Recommendation systems (RecSys) enhance information retrieval efficiency across different domains by delivering personalized content that aligns with user preferences. These systems address user-item relationships using methods such as matrix factorization and graph attention networks (GATs). Despite advances in accuracy, existing approaches often focus narrowly on predictive performance while neglecting the broader utility of confidence estimation. This estimation is crucial for quantifying the confidence behind recommendations, particularly when balancing risk and reward is needed to foster trust. RecSys can mitigate uncertainties stemming from data noise and model limitations by utilizing confidence. Existing approaches to confidence integration face critical limitations. Parametric methods risk non-convergence, inflexible uncertainty design, and often degrade accuracy. Non-parametric techniques, such as neural network-based probabilistic calibration, remain confined to classification tasks and fail to address regression scenarios, including rating prediction and listwise learn-to-rank. Many confidence models operate independently of core recommendation processes, limiting their adaptability and their impact on calibration. Notably, the literature neglects integrating confidence estimation into GAT-based models. Additionally, the literature lacks experimental evaluation of different distribution-based methods. Therefore, this study proposes an experimental evaluation of prior distribution-based methods and explores integrating suitable confidence into GAT-based models. For instance, we evaluate four prior solutions based on rating prediction accuracy, ranking accuracy, and correlation with confidence and error. These solutions and our proposal are evaluated in public datasets from varying contexts and characteristics. Results reveal that distribution-based confidence integration often harms model accuracy and leaves room for improvement in the correlation between confidence and error. Although these findings also hold for our method, it still achieves superior performance compared to all prior solutions and shows a promising negative correlation between confidence and error. Furthermore, as a second part of this study, we propose and evaluate integrating confidence into embedding-based learn-to-rank methods. This proposal and its baselines are also evaluated across various public datasets using different ranking metrics, and the correlation with confidence and error is assessed. The results reveal that the proposed method consistently achieves competitive ranks and even outperforms the baselines on some datasets. Therefore, the proposed learn-to-rank method does not significantly degrade the model’s performance. Additionally, we observed a cubic polynomial relationship between confidence and error.

Educational Robotics stands out as a pedagogical approach capable of enhancing learning in computer programming and fostering student engagement in Vocational and Technological Education (EPT, in Portuguese). This dissertation aims to investigate the impact of Educational Robotics in the context of integrated technical secondary education, using a participatory research carried out through a Robotics Club in a Territorial Center for Vocational Education. The adopted methodology included an initial assessment of the students' levels of engagement and prior knowledge, followed by weekly practical activities with the application of programming concepts mediated by robotics. In the final stage, the evolution of the participants' engagement and technical knowledge was quantitatively and qualitatively evaluated. The results demonstrate a significant increase in the students' emotional, cognitive and behavioral engagement, as well as a consistent improvement in the mastery of technical skills in programming. It is concluded that the use of Educational Robotics is effective in reducing cognitive barriers and promoting a more interactive and accessible learning environment for students in the technical modality integrated with High School.

In the context of government procurement in Brazil, efficiency and continuous monitoring of spending represent significant challenges to public management. Brazil’s government, in 2023, issued the amount of 1.761.910 invoices to different types of purchases, resulting in an amount of R$ 76,62 billions in negotiations with private entities (Transparˆencia 2024). The acquisition of these supplies is distributed in various locations throughout the country, generating a growing and diverse volume of information, found in contracts and invoices of products and services. However, these government purchases are frequently a fertile soil to the occurrence of colluding and fraud (OECD 2007), as overbilling on the price of products, suppliers monopoly, bribes to public officials, etc. The goal of this paper is to compare the performance of Natural Language Processing (NLP) models in the task of detection - based on the extract of government purchasing - of companies that were already punished by government agencies, such as the Office of the Comptroller General (CGU). The used data is public and periodically updated through Federal Government Open Data’s page. The results of this paper show that it is possible to use the natural language models as an early step to investigate suspicious purchases, providing a classification of potentially problematic purchases and afterwards it can be.

In this work, two optimization problems related to routing and scheduling are studied. The first study addresses the PDPTW-SE.
The challenge lies in determining routes for a heterogeneous fleet of vehicles to transport requests with specific pickup and delivery locations, considering that some traversals require synchronized operations with machines. Since the number of machines is limited, their usage must be properly scheduled. The objective is to minimize the total completion time while respecting capacity, time window, and precedence constraints. To this end, a MIP formulation and a multistart heuristic with a linear programming-based improvement procedure are developed. A benchmark set of instances is also proposed, consisting of two families of instances representing different applications of the problem. In the computational experiments, the MIP formulation solves instances with up to twelve requests and finds feasible solutions for 93.40% of the cases. The heuristic, in turn, obtains feasible solutions for all instances, with quality equivalent to or better than that of the MIP formulation. The second study addresses the SMCTSP.
This is a job scheduling problem in which each job consists of two coupled, non-preemptive tasks separated by an exact delay.
The objective is to minimize the completion time of the last scheduled task. The problem was modeled using CP, and a BRKGA was developed, incorporating a warm-start solution generator, periodic restarts with varying intensities, and a local search algorithm. Computational experiments showed that the proposed BRKGA provides high-quality solutions with reduced computational times compared to the CP model. On the other hand, the CP model significantly outperformed the BRKGA solutions when run for one hour with multiple threads. Finally, the proposed approaches combined obtained new best solutions for 93.33% of the instances that had not yet been solved to optimality by previous approaches in the literature.

The advancement of natural language models has been marked by the transition from
rule-based approaches and statistical methods to deep neural architectures, such as the

Transformer, which enable the modeling of contextual dependencies in texts in a distri-
buted manner. This study investigates the capacity of a monolingual BERT-based model,

trained on Brazilian Portuguese data, to represent syntactic governor→dependent relati-
ons, as described within the Universal Dependencies (UD) framework.

To conduct the analysis, we used the annotated UD Portuguese-Bosque corpus, from
which we extracted sentences containing diverse grammatical patterns, including verbal

transitivity, passive voice, reflexive pronouns, subject predicatives, and subordinate clau-
ses. The sentences were processed by the model, and attention values were extracted per

layer and head, aiming to identify alignments between attention weights and the syntactic
dependencies recorded in the corpus. The model’s tokenizer was used alongside lexical
tracing mechanisms that allow associating subtokens with their respective positions in
the original texts, enabling an interpretable analysis of syntactic pairs.
The evaluation relied on metrics such as grammatical pattern-wise accuracy, attention

distribution entropy, and the Undirected Unlabeled Attachment Score (UUAS). A compo-
site metric combining selectivity and structural adherence was also applied. The results

reveal that certain attention heads exhibit systematic activation patterns concerning spe-
cific dependencies. Notably, head 3 in layer 2 consistently aligned with relations between

verbal nuclei and their arguments, representing an example of emerging functional speci-
alization. These findings contribute to the understanding of the internal attention mecha-
nisms of Transformer-based models applied to Brazilian Portuguese and provide insights

for future approaches in supervised compression and automated linguistic analysis.

Interpersonal trust among students is essential for the success of collaborative activities in Virtual Learning Environments (VLEs), but establishing it is challenging due to the absence of physical presence and the limitations of computer-mediated communication. Lack of trust can compromise knowledge sharing, motivation, engagement, and student retention in courses. In order to support the development of interpersonal trust among students in VLEs, this research aimed to identify software requirements for features that promote trust. To this end, a Systematic Mapping Study (SMS) was conducted to understand how trust has been conceptualized in the literature, as well as the intrinsic aspects of this phenomenon. Next, attributes and characteristics that influence trust among students were analyzed across four evolutionary phases: acquire, maintain, loose, and restore trust. As a result, 37 attributes were identified, most of which originated from other domains but can influence the dynamics of trust among students in VLEs. However, the evidence analyzed focused mainly on the trust acquire phase, while the loose phase was little explored. The SMS revealed significant gaps in the field of trust among students, highlighting, in particular, the absence of characteristics associated with maintain and restore trust. In this sense, it became clear that there was a need to explore new characteristics based on the students' own perceptions of the phases of acquire and lose trust, as well as to understand, from their perspective, the factors that influence the maintain and restore of trust. Based on the gaps identified, an exploratory study was conducted through an online survey with 170 students, which aimed primarily to understand which personal and behavioral characteristics can influence trust in the four evolutionary phases. A total of 248 characteristics were identified that can influence trust, with 96 in the acquire phase, 78 in the loose phase, and 37 characteristics in each of the maintain and restore phases. Based on these results, 26 software requirements were identified for the following features: student trust profile, collaborative work groups, trusted peer recommendation systems, peer evaluation, and friend network. In addition, a platform-independent conceptual model was developed that illustrates how the features can be integrated during the development process of VLEs that aim to promote trust. Finally, the software requirements were validated with stakeholders using a high-fidelity mockup that illustrated the software requirements in the intended features. The main artifacts generated from this research were: a conceptual map based on definitions from the literature, a conceptual model with attributes by evolutionary phase, cataloging of characteristics identified in the literature and by students, software requirements, and a conceptual model with features and requirements aimed at promoting trust.

The software industry faces constant challenges in adapting to rapid market changes,

meeting the growing demands of customers, and maintaining the productivity of development

teams. In this context, there is increasing interest in hybrid development approaches

that combine agile methods with plan-driven approaches. Due to its iterative,

incremental, and adaptable nature, Scrum stands out as a frequent foundation for building

hybrid processes. However, difficulties persist in formulating and effectively adopting

these approaches, which can compromise productivity, product quality, and stakeholder

satisfaction.

This dissertation aimed to investigate, characterize, and analyze the use of hybrid

processes in the software industry, particularly those involving Scrum, from the perspective

of professionals working in development. An empirical study was conducted with 158

Information Technology professionals worldwide, including 143 survey participants and

15 respondents in semi-structured interviews. Data analysis enabled the identification of

practices, methodological combinations, motivations for adoption, challenges faced, and

strategies used to overcome them.

The results indicate that hybridization is largely driven by the pursuit of flexibility,

shorter delivery times, higher quality, risk mitigation, and greater project control. The

most recurrent combinations were ScrumBan and Water-Scrum-Fall. Reported challenges

included defining which practices to combine, adapting teams, resistance to change,

and the lack of consolidated guidelines to guide adoption. Professionals’ perceptions reveal

that the practical application of hybridization does not always align with academic

definitions, highlighting a gap between theory and practice.

Based on these findings, 34 recommendations were formulated and organized into

human, technical, and organizational categories, aiming to support the adjustment and

improvement of software development processes. These recommendations are expected

to contribute to building more effective processes tailored to the needs of each project,

helping to overcome common obstacles to the adoption of hybrid methods.

By sharing these findings, this dissertation seeks to expand knowledge about software

process hybridization, promote a clearer understanding of its benefits and challenges,

and encourage more efficient practices in the industry, with positive impacts on project

productivity and quality.

The practice of listening to music is widely adopted by software engineering professionals to improve concentration and attenuate environmental noise. However, the neurophysiological and behavioral impact of this practice on software testers, a group with
distinct cognitive demands, remains largely unexplored. This dissertation investigates
the influence of music on the performance (accuracy) and neurophysiological activity of
professional software testers during the execution of testing tasks. We conducted a controlled experiment with 14 professionals, divided into an experimental group (exposed to
Lo-Fi music) and a control group (in silence). The participants performed four blocks of
distinct tasks: test code comprehension, syntax error identification, logic error identification, and test case creation. Brain activity was analyzed using Electroencephalography
(EEG), focusing on metrics such as Power Spectral Density (PSD) and Event-Related
Desynchronization/Synchronization (ERD/ERS). The results indicate that the musicexposed group achieved higher accuracy in the analytical tasks of comprehension, syntax
detection, and logic detection. However, the experimental group exhibited slightly lower
performance on the generative task of test case creation. The music group also exhibited
more stable EEG activity, notably in the Beta and Delta bands, associated with attention
and internal cognitive processing. Furthermore, they exhibited a more pronounced Alpha
desynchronization (ERD) during the analytical tasks, suggesting greater concentration.
Conversely, in the generative task, the music group demonstrated Alpha synchronization
(ERS), indicating a more relaxed state and lower focus. This study, pioneering the application of EEG to professional software testers, provides objective evidence that the
impact of music on performance is task-dependent. Instrumental music seems to act as
a cognitive enhancer for analytical tasks but may not be beneficial for generative tasks.
The contributions include a new methodological foundation for the field and a public
dataset for replication and future investigations.

Identity is essential for recognizing entities (individuals, things, and organizations) and their various relationships within their contextual environment. Identity Management (IdM) involves processes such as authentication, authorization, accountability, and auditing, traditionally carried out by centralized or federated systems. However, these models limit privacy, interoperability, and user control over personal data. In this context, Decentralized Digital Identities (DDIs) emerge, in which individuals themselves hold and manage their own identity—now unique, portable, and securely shareable. This work presents ChainID, a platform for decentralized identity management based on blockchain, developed within the RNP (Brazilian National Education and Research Network) Working Group GT-ChainID. The platform adopts a service-oriented approach to enable the creation, issuance, verification, and revocation of decentralized identifiers (DIDs) and verifiable credentials (VCs), abstracting the complexity of the standards, protocols, and distributed infrastructure involved in this new digital identity paradigm. ChainID was initially built on the Hyperledger Indy blockchain, which is oriented toward privacy and credential management. As the solution’s requirements and architecture evolved, the platform migrated to the Hyperledger Besu blockchain, compatible with the Ethereum Virtual Machine (EVM). This transition allowed compatibility with the RNP Testbed experimentation environment, improved operational capabilities, and opened possibilities for future integration with the Brazil Blockchain Network (RBB). The solution provides a RESTful API, support for asynchronous events, integration with protocols such as SAML2 and CAS authentication, as well as dedicated interfaces, including the ChainID Console for configuration and administrative management, and the ChainID Wallet for individual control of user DIDs and VCs. As proofs of concept, two use cases were implemented: (1) the ChainID federated authentication component, demonstrating its applicability in educational scenarios using Moodle and the feasibility of replacing centralized infrastructures with DDI-based solutions without compromising security, privacy, or interoperability; and (2) the CarbonID project, which explores the issuance, validation, and traceability of sustainability certificates using DIDs and smart contracts, showcasing the platform’s potential for adoption in other applications and domains, including environmental, social, and corporate contexts.

Ridesharing systems have emerged as a potential solution to urban mobility challenges, promoting collaborative vehicle usage and route optimization. These systems require efficient routing algorithms to balance conflicting objective functions such as travel cost, travel time, and driver bonuses. Previous studies have modeled such systems using the Traveling Salesman Problem with Profits (TSPP), where the driver shares the vehicle with passengers and minimizes travel cost. The Bi-objective Traveling Salesman Problem (BiTSP) has also been investigated in prior work, but it ignores passenger and bonus collection. Consequently, the literature lacks a multi-objective formulation that captures the real-world trade-offs among such objective functions. This work introduces the Multi-objective Traveling Salesman Problem with Profits and Passengers (MoTSPPP), an NP-hard optimization problem that minimizes travel cost and time while maximizing bonus collection. A mathematical formulation and a proof of NP-hardness are provided. Eight algorithms are investigated: an exact solver, three naïve heuristics, and four evolutionary metaheuristics (NSGA-II, MOEA/D, IBEA, and SPEA2). A comprehensive experimental study is conducted on 252 benchmark instances, comprising symmetric and asymmetric graphs with varying edge-weight correlations. Supported by statistical tests, performance evaluation concerns processing time, solution quality, and solution diversity. Results demonstrate that the MoTSPPP is computationally more challenging than the TSPP and BiTSP, and that metaheuristic approaches yield significantly better results than naïve heuristics.

Real-time systems (RTS) are composed of a set of tasks (code segments) that are recurrently launched to be executed and must meet deadlines. Designing such a system in a provably correct manner requires information about the worst-case execution time (WCET) for each of its tasks. However, estimating the WCET is becoming increasingly challenging due to the high complexity of hardware and software in modern platforms. This has motivated the use of techniques to derive the probabilistic worst-case execution time (pWCET). Most existing approaches rely on measuring the execution time of system tasks on the target platform. As measurements are taken during the design time, collected samples may lead to unreliable estimates (due to possible measurement bias) or non-representative ones (due to difficulties in reproducing operational conditions). The need to make samples compatible with the assumptions of statistical modeling is an additional source of difficulty. Given the complexities presented, two studies with distinct objectives were developed. In the first study, a representation of execution time is performed based on hardware events, considering different computational intelligence tools. Specifically, for a program under analysis, it is shown that the execution time T(n) per number n of executed instructions can be correlated with occurrences of hardware-related events. In the second study, a new approach for Measurement-Based Probabilistic Timing Analysis (MBPTA) is presented. Unlike the usual MBPTA, which considers only T(n) as the variable of interest, this new approach incorporates a variable of interest that considers both T(n) and n. Using tuples (n, T(n)) for different values of n allows for exploring multiple execution paths. Additionally, this new approach allows the set of measurements to be evaluated and improved. For this purpose, deep neural networks (DNN) were employed. Since the measurements are considered representative, it is possible to estimate probabilistic bounds on the execution time. Experimental results indicate a difference of up to 30% between the estimates obtained using samples refined by the proposed approach and those obtained using non-refined samples. The approaches are evaluated considering different hardware and program models, and the results obtained demonstrate effectiveness in the proposed studies.

Collaborative Recommender Systems generate personalized recommendations by analyzing users' past interactions. However, traditional approaches often prioritize relevance, leading to issues such as super-specialization, popularity bias, and class imbalance. These limitations can result in recommendation lists that fail to represent the full spectrum of a user’s interests fairly. In this sense, Calibrated Recommendations address this problem by balancing relevance with fairness (calibration), ensuring that the distribution of recommended items aligns more closely with the user’s preference distribution. For example, when the user's profile comprises 80% Adventure and 20% Sci-fi, the calibrated recommendation seeks to generate a list following this distribution. Relevance and calibration are two distinct goals that the system should achieve. This multi-objective is reached through a trade-off balance approach. This thesis addresses calibrated recommendations as a multi-objective recommendation system, aiming to measure and improve the calibration of the recommendation list using the user's preferences as a target. Thus, we divide the goals of this thesis into studies, and inside each study, research questions were raised and answered. In the first study, we systematically benchmark 57 fairness measures, introducing novel methods for extracting user preference distributions and refining relevance estimation. As a result, four measures achieve the same four best performances. In the second study, we explored the broader impact of calibration on key recommendation objectives, including novelty, coverage, personalization, unexpectedness, and serendipity. Our findings indicate that calibration enhances item coverage and personalization while maintaining high recommendation utility. In the third study, we investigate the structural properties of the distributions used in calibrated recommendations. Unlike traditional recommender systems that operate in a one-dimensional space, calibrated recommendations involve high-dimensional user preference distributions. Our analysis shows that calibrated recommendation lists naturally form distinct user clusters, a phenomenon best understood through outlier detection models. In the fourth study, we propose two novel approaches for modeling user preferences to enhance the accuracy and adaptability of calibration techniques. The first method incorporates time-sensitive weighting to discount outdated preference information. The second method introduces an entropy-based approach to better capture user preferences in domains where item features are set-valued, such as movies with multiple genres. Experimental evaluations confirm that these approaches effectively reduce miscalibration while maintaining recommendation accuracy. Overall, this thesis advances the field of calibrated recommendations by providing a comprehensive evaluation of fairness measures, proposing novel calibration techniques, and analyzing the structural properties of user preference distributions. 

Context. Empathy is the ability to understand and share the emotions of others, a critical skill for software practitioners as it contributes to improved software quality, communication, collaboration, and work environments. Despite its importance, empathy remains an underexplored topic in Software Engineering (SE) research.

Aims. To address this knowledge gap, this dissertation aims to deepen our understanding of empathy in SE by investigating how it is defined, practiced, and experienced by software practitioners, as well as identifying barriers and effects related to empathetic behavior in development contexts.

Method. We adopted a mixed-methods approach. First, we conducted a qualitative analysis of grey literature from practitioner platforms (DEV and Medium). Then, we surveyed software practitioners within a large software organization to quantitatively and qualitatively examine empathy-related perceptions and behaviors. We also evaluated our proposed framework with empathy experts.

Results. The study revealed different meanings and a high perceived value of empathy, as well as barriers that hinder its application in software teams. We identified a set of empathetic practices and categorized them into empathy dimensions based on exploratory factor analysis. We also found that empathy practices are closely linked to positive outcomes, ranging from productivity and technical quality to collaboration, well-being, and professional growth, reinforcing the practical relevance of empathy across multiple dimensions of software practice. These insights informed the development of a conceptual framework of empathy in SE, which was validated through experts' feedback.

Conclusion. This research advances the understanding of empathy in software engineering by offering a theoretically and empirically grounded framework, a curated dataset, and practical implications for teams and organizations. It lays the foundation for future work on empathy-driven practices, tools, and interventions in software development.

This dissertation presents the development of an innovative technological artifact that integrates Internet of Things (IoT) devices, blockchain technology, and a service-oriented middleware to enable the traceability and automated control of the fermentation and drying processes of fine cocoa. The work begins with a contextualization of the subject, followed by a conceptual and technical review of the fundamentals of IoT, consensus pro- tocols used in blockchains, and architectural models that support the efficient integration of IoT and blockchain systems. Within this context, both functional and non-functional requirements are analyzed, particularly those enabling the adoption of blockchain solutions in computationally constrained environments, which are typical in IoT scenarios. Subsequently, the development of a service-oriented middleware is described. This mid- dleware supports semantic interoperability through the use of ontologies and is desig- ned to ensure compatibility among heterogeneous devices. It offers APIs based on the REST architectural style, which were adapted to emulate the behavior of SNMP (Simple Network Management Protocol) messages. This approach led to a significant reduction in message exchange and computational resource consumption. Additionally, the dis- sertation describes the distributed applications developed to configure and manage the control of fermentation and drying actions, supported by custom-designed IoT hardware specifically built for this purpose. The data captured by the sensors are pre-processed and forwarded to the middleware, which then records them in smart contracts deployed on the Ethereum blockchain, thus ensuring the process benefits from the inherent im- mutability, auditability, and reliability provided by blockchain technology. The thesis also presents a contextual overview of the cocoa-producing region of Southern Bahia, dis- cussing the challenges and opportunities associated with the adoption of Agriculture 4.0 technologies. As part of the evaluation of the proposed solution, a proof of concept was developed, demonstrating key results such as computational resource savings, messaging efficiency, and reliability in data traceability. Finally, it is worth noting that beyond its applicability in the cocoa sector, the developed middleware has the potential to be exten- ded to other domains, such as asset inventory control and water resource management, thus demonstrating the versatility and robustness of the proposed architecture.

With the advancement of digital technologies, Virtual Learning Environments (VLEs) have evolved from simple content repositories to interactive spaces that support active methodologies and promote meaningful teaching and learning experiences. However, many of these environments still need to align with the principles of Instructional Design (ID), which makes it difficult for teachers to develop coherent pedagogical practices. Given this scenario, this research aimed to propose, structure, and evaluate a conceptual framework to support education professionals in the use of pedagogical strategies in VLEs, based on instructional design. The work was conducted according to the Design Science Research (DSR) approach and followed a methodological path composed of: a systematic mapping study, application of a survey with 276 education professionals, conducting interviews with 26 teachers, triangulating the data, evaluating the conceptual structure in 03 cycles of focus groups, and a case study with 10 computer science teachers. Based on the analysis of data collected from the survey and interviews, three active methodologies that are recurrent and most significant in the practice of educators were identified: the flipped classroom, Problem-Based Learning, and Project-Based Learning. These methodologies were adopted as structuring axes of the framework, due to their relevance in promoting collaboration and problem-solving in educational contexts mediated by Digital Technological Resources (DTRs). The resulting conceptual framework aims to integrate elements of instructional design with the functionalities of VLEs, providing a theoretical and practical guide that supports informed decision-making regarding the selection of methodologies, strategies, and digital resources. Its evaluation demonstrated applicability and relevance, especially as a planning and reflection tool for teachers in hybrid and remote contexts. Thus, this thesis represents the conclusion of an investigative cycle and contributes to the field of educational technologies by offering a structured model that strengthens the pedagogical use of VLEs.

The accelerated growth of applications and services on the Internet, with heterogeneous requirements, demands more efficient service differentiation mechanisms than those offered by traditional IP networks. In this context, Named Data Networking (NDN) emerges as a promising alternative to meet the needs of the current communication model. NDN introduces a paradigm shift from the traditional IP address–oriented architecture by prioritizing content access by name rather than by location, and by offering features such as in-network caching, stateful forwarding, and packet-level security—capabilities that enable more sophisticated resource management. Despite its benefits, NDN by default does not implement service differentiation mechanisms, treating all traffic uniformly, which limits support for applications with distinct Quality of Service (QoS) requirements. Furthermore, the efficiency of on-path caching and conventional forwarding strategies can be compromised in dynamic environments with high competition for resources. This thesis proposes a set of adaptive mechanisms for QoS provisioning in NDN, based on the integration of naming semantics, cache management, and forwarding state maintenance, focusing on service differentiation. The developed works include: a cooperative and adaptive forwarding mechanism for traffic control based on forwarding state to mitigate Interest flooding in wireless NDN; an opportunistic and cooperative caching mechanism to support producer mobility; a dynamic, content-centric load-balancing mechanism that promotes efficient traffic distribution and differentiated content delivery; and an integrated approach that combines naming semantics and cache management to enhance QoS provisioning and the performance of heterogeneous applications. The proposed mechanisms are designed, implemented, and evaluated through simulations using ndnSIM, considering different network topologies and scenarios, including both mobile and wired networks. These mechanisms explore several components of the NDN stack, such as naming, caching, and forwarding. In this way, the thesis presents new perspectives for service differentiation in NDN, demonstrating that the integration among different components has the potential to provide QoS more efficiently and to support heterogeneous applications. This work establishes conceptual foundations and practical mechanisms that can guide future research in this field.

Technical Debt (TD) can provide short-term benefits for software projects, but its presence can lead to decreased product quality. Although recent literature has proposed several approaches to identify and manage Agile Software Development Technical Debt (TD-ASD), most of them look at the point of view of software developers. However, there is a gap in knowledge about how project management professionals perceive and deal with TD-ASD items. The objective of this study is to investigate how project management practitioners discuss and face TD items in their professional activities. To achieve this objective, an analysis of 108 discussions related to TD-ASD was carried out on the Stack Exchange Project Management (SEPM) question and answer website. These discussions totaled 547 posts and 882 text comments, providing a good overview of the topic. The analysis was conducted using quantitative and qualitative methods, employing open coding to identify types of TD and indicators. The study identified 74 indicators used to recognize TD-ASD items, 126 TD management practices and 9 types of TD. The most frequent discussions in SEPM about technical debt focus on two types of process and people, differing from previous studies with developers, which tended to focus on code and design debt items. Furthermore, the Product Owner and Development Team are the roles most involved with TD-ASD. Sprint Backlog and Sprint Planning are the agile elements most affected by TD-ASD. Project management professionals who use SEMP adopt a different perspective than developers when analyzing TD. TD management indicators and practices were structured in a Sankey diagram, providing a valuable tool to guide ID management and guide future research on the subject.

Face detection and recognition algorithms have been widely adopted in a diversity of applications, such as social networks that automatically detect and recognize every person present in published images. However, with the growing adoption of Artificial Intelligence (AI) algorithms in general, questions related to the existence of biases began to arise. In many situations, it was found that there were biases affecting historically oppressed minorities. As an example, there was racial bias in many facial recognition systems used by American police, which led to the suspension of the use of this technology in some states, the discontinuation of development in some companies, such as IBM, and researchers to ask their colleagues to stop working on these systems because of the impact on people of different races and ethnicities. The aforementioned problem motivates the study and evaluation of the existence of bias in a fraud detection, AI-based, system used by the public transportation in Salvador (Brazil). Taking into account the fact that Salvador is the Brazilian city with the highest percentage of black people, any error can affect a significant number of users, leading to a high number of false positives. In previous studies developed by the research group, statistical tests were performed to verify if there is a correlation between the error rate and race and gender. The results indicated the existence of this correlation, that is, there is a higher error rate in face detection of black or brown users and women. From these results, a main question motivates the development of this project: Is there, in fact, a causal relationship between race and the error rate in detection? To evaluate this question, this work’s main objective is to create a causal model that allows the evaluation of the influence of skin color in the face detection system used by public transportation in Salvador.

In addition to smart traffic lights, systems are being advanced for the efficient management of intersections, including controlling the flow of pedestrians, cyclists and autonomous vehicles. On the other hand, inefficient management of intersections can lead to congestion, delays and increased risk of accidents. Furthermore, failures in data communication, resulting from problems such as physical obstacles, interference, network failures and faulty sensors, can generate gaps in information transmission, adversely impacting management solutions. In this context, a proposal for an intersection management system is presented, aiming to improve the safety and efficiency of urban traffic for autonomous vehicles. The system's decision-making is based on continuous data communication between vehicles and infrastructure. Based on this data, the system performs a conflict analysis that identifies possible collisions between vehicles and performs dynamic speed adjustment of these vehicles. In order to mitigate the negative effects of missing information, we incorporated missing data imputation methods by using cubic polynomial segmented interpolation (PCHIP), a process we call GIVAC. The results indicate that GIVAC was stable in all scenarios with data loss, keeping the average total travel time in simulations 68.4\% lower than that of the baseline algorithm, Intelligent Intersection (II).

Real-time systems are present in several applications, from critical applications, like
automotive and aviation, to non-critical such as multimedia and online games. For critical
real-time systems to function correctly, it’s necessary to ensure that all tasks execute
within pre-defined deadlines. To this end, the system’s task scheduler plays a fundamental
role, as it determines at each moment which task must be executed so that they can all
meet their deadlines. In systems with one processor, the Earliest Deadline First (EDF)
scheduler is one of the best options, as it guarantees that all tasks meet their deadlines
whenever possible. On systems with more than one processor, the same is not true. This
work proposes the Partitioning and Server Shadowing Algorithm (PSSA) scheduler for
multiprocessor systems composed of periodic and independent tasks. The idea of the
algorithm is to partition tasks between processors and then use EDF scheduling on each
processor. When such partitioning is not found, some tasks must execute on more than
one processor, while some processors are left with idle capacities. In order to manage
the execution of tasks that run on more than one processor, PSSA uses a new approach
called server shadowing, according to which the idle parts of real processors are used to
create logical processors. When a task runs on a logical processor, it actually runs on
one of the real processors that have provided their idle capacity to the logical processor
where the task is allocated. The concepts of server shadowing and logical processors gave
PSSA the ability to obtain results similar to the best results published to date in terms of
the number of migrations and preemptions generated for periodic tasks, as demonstrated
through extensive simulations.

The growing technological advancement has driven the development of smart homes, requiring adaptive and personalized systems to meet the needs of residents. Additionally, with the increased use and integration of devices, smart homes have the potential to adapt to users' needs, offering personalized services such as energy management, home security, and comfort. At the same time, this evolution has introduced more challenges and complexity to these scenarios, leading to many possible states for smart devices, such as different colors, temperatures, and intensities for a light bulb. One alternative to address some of these challenges is the application of recommendation systems in smart home approaches, where the recommended items are states or actions for devices within the environment. The recommendation system proposed in this work aims to optimize the interaction between the smart devices in the home and its inhabitants, promoting more efficient decisions within the household environment. This study explores new approaches to enhance the efficiency and adaptability of recommendation systems in intelligent residential environments by applying reinforcement learning algorithms, implicit feedback, and cooperation among multiple agents for controlling actuating devices. In this study, we conducted experiments with two different reinforcement learning algorithms on two different datasets based on three distinct routines. The results demonstrate the system's promising ability to anticipate residents' needs and adapt to routine changes, contributing to a more efficient, intuitive, and adaptable residential experience.

In this work, a framework was developed that integrates an Internet of Things (IoT) network with Blockchain and Smart Contract (SC) technologies, aiming to make the IoT a reliable data source for Machine Learning (ML) model training systems. Reliability in this research consists of the provision of automation, availability, privacy, traceability, and immutability of data by the architecture. The architecture of the proposed framework was conceived from a Digital Twin (DT), capable of virtually representing the topology of a Fog of Things (FoT), with the aid of the MININET project. Thus, the ability to ensure reliability was evaluated through experiments that simulated the behavior of the proposed solution compared to the conventional architecture. The experiments focused on the interaction of the architectures as a data source for an ML system, which uses the Federated Learning (FL) algorithm to train a Neural Network (NN) capable of predicting the thermal comfort of an environment, guided by the calculation of the Thermal Discomfort Index (IDT). The results revealed that conventional IoT does not have effective measures to prevent or mitigate data manipulation attacks. In contrast, the proposed framework, even if an attacker manages to alter information, restores the original data, ensuring that the consumer receives the data exactly as it was collected, without compromising availability and without the need for human intervention. To determine the viability of the architecture, analyses of computational cost, market resources, and temporal impact of technology integration were carried out, demonstrating that the solution, in addition to being efficient, offers a low cost compared to centralized solutions.

Test smells are considered bad practices for developing the test code. Their presence can reduce the test code quality, thus harming software testing and maintenance activities. Software refactoring has been a key practice to handle smells and improve software quality without changing its behavior. However, existing refactoring tools target production code with very different characteristics than test code. Despite the research invested in test smell refactoring, little is known about whether current refactorings improve the test code quality. This thesis proposal presents our research to help developers decide when and how to refactor test smells through a machine learning-based approach. First, we aim to mine refactorings performed by developers to derive a catalog of test-specific refactorings and their impact on the test code. Our findings show that developers prefer specific features of the testing frameworks, which may lead to test smells such as Inappropriate Assertion and Exception Handling. While the refactorings proposed in the literature aligned with the evolution of testing frameworks to help refactor test smells, the Inappropriate Assertion remains unexplored in the literature. Second, we aim to understand whether developers target low-quality test codes to perform refactorings and the effects of refactorings on test code quality improvement. Our findings show that low-quality test code, especially regarding structural metrics, is more likely to undergo refactorings. Common refactorings between test and production code contribute more to improving test code quality in terms of cohesion, size, and complexity. Test-specific refactorings enhance quality concerning the resolution of test smells. Third, we aim to learn whether developers would perform refactorings and which refactorings they would apply to improve the test code quality. Results indicate that the accuracy of Support Vector Machines models varies between 30% and 100% in different projects for detecting when a developer would perform a refactoring. However, accuracy decreases for detecting specific refactorings due to the low data on test refactorings found in analyzed projects. Overall, this research demonstrates the feasibility of using structural metrics and test smells for detecting test refactorings. In addition, it highlights the need for improvements through the analysis of synthetic data and project development context. The proposed approach represents a promising step towards supporting the detection and refactoring of test smells aligned with development practices currently adopted by developers.

Dental panoramic radiographs are not only a highly valuable exam but also a versatile one. They can be used to diagnose periodontal bone loss, lesions, cysts, and tumors, as well as estimate the age and biological sex of the patient. The works that use deep learning to determine such conditions in panoramic radiographs are based on supervised approaches that require manual annotation of each attribute and condition considered. However, manual annotation of radiographs is demanding, as it requires qualified labor and is, consequently, expensive. This work seeks to overcome this difficulty by exploring the Human-in-the-Loop concept, a semi-supervised learning technique that expedites the labeling process through an interaction between human experts and machine learning models. To support this approach, special focus was given to teeth, as they are the main objects of attention and reference points for radiologists when reading panoramic radiographs. As a result, a dataset for tooth instance segmentation of panoramic radiographs was produced: the O$^2$PR dataset, containing 4,000 images. The remaining data of work consists of 4,795 radiographs in the Raw Panoramic Radiographs (RPR) dataset, with images in their crude format, and the Textual Report Panoramic Radiographs (TRPR) dataset, containing 8,029 pairs of radiograph images and textual reports. These groups of data comprise the most extensive dataset in the literature. Starting from these datasets, we classify thirteen dental conditions in the tooth or its surroundings. To classify all the considered conditions, a holistic approach was necessary. First, using the labeled radiographs of the O$^2$PR dataset, we trained an instance segmentation neural network to pseudolabel the teeth in the unlabeled radiographs. Subsequently, all tooth images were cropped to facilitate the classification of dental conditions. The O$^2$PR and RPR datasets do not include textual reports, making it impossible to generate labels for training or evaluating these images for dental conditions. Instead, the tooth crops from these datasets were used to pre-train Vision Transformers (which were later employed as classification networks for dental conditions) through a self-supervised learning technique called Masked Autoencoders. This approach proved effective as it allowed the use of unlabeled data to improve performance. The label extraction procedure follows a different branch. We explored the API of a Large Language Model, GPT-4, to avoid the pure manual labeling of the dental conditions. The goal of using it was to identify the noun phrases from the textual reports to find the dental conditions. Later, a heuristic associated each tooth present in the report sentences with all the dental conditions of the same sentence. We leverage the pretrained Vision Transformer to train several dental condition classification models. Encouragingly, the results consistently met or surpassed the baseline metrics for the Matthews correlation coefficient. A comparison of the proposed solution with human practitioners, supported by statistical analysis, highlighted its effectiveness and performance limitations; based on the degree of agreement among specialists, the solution demonstrated an accuracy level comparable to that of a junior specialist.

The increasing access to digital technologies leads to a large-scale data production and consumption worldwide. Consequently, there is a high demand for computing analysis techniques that help users to obtain a better understanding of data. An inherent challenge in this context is to analyze large volume of complex and heterogeneous datasets, such as conceived by software repositories. Areas such as Software Visualization and Visual Software Analytics have become increasingly used to support developers in software comprehension by providing a visual panorama of the entire software development process. Those graphical representations enable useful information extraction, especially for system maintenance when analyzing technical debts. Although repositories are mostly composed of multidimensional datasets, there is a lack of works that apply multidimensional visualizations to identify and track distinct groups of technical debts. In this sense, the proposal presents the Visual Debts Analytics tool, an approach based on Visual Analytics consisting of multiple coordinated multidimensional visualizations for the analysis of different groups of technical debts in software repositories. In particular, the proposed technique aims to identify and track the correlation, structure, evolution and similarities of technical debts in open-source software repositories. To evaluate our approach we presented case studies in two different repositories that demonstrate the a better comprehension in the project life cycle, enabling insights into the project quality.

Blockchain is a disruptive technology that offers a secure decentralized network, and allows for direct transactions between distributed entities, without the need for a trusted third party. Smart contracts (SmC) are executable codes hosted on the blockchain and enable the implementation of several decentralized applications, in different domains. However, due to the complexity of SmC construction, it is necessary to seek ways to facilitate and make this innovative development process robust. Many projects in the literature proposed solutions based on automatic code generation from high-level modeling, such as diagrams. However, this approach brings with it a constant need for a person to model the contracts. This paper presents Gifflar, a framework for generating SmC code on the fly, in such a way that it allows a system to write, compile and implement blockchain SmCs with an application still running, which reduces the frequent precision of the SmC developer and allows to further automate the SmC development process. One of the main tools is a component-structured library that implements design patterns to better define the responsibility of each of these parts. The Gifflar library offers an API that allows the developer to model, generate code and manage SmCs through methods that abstract to a certain level the SmC code using the JSON as a model to the code generation. In addition, project evaluations were carried out to validate the project: (i) a usability evaluation of the Gifflar library and (ii) a conceptual evaluation of Gifflar's application in other projects. As far as we know, this framework is one of the first to allow dynamically generating SmCs on the fly, thus contributing to the state of the art by approaching a new paradigm, where systems can act as smart contract developers.

Self-adaptive Systems (SAS) can monitor themselves and their context. They can detect changes and react to unexpected conditions with minimal human supervision during their execution. One of the challenges behind developing SAS is dealing with the decision-making process while analyzing the tradeoff points among the multiple quality attributes (QA). In Software Engineering, a widely accepted method of evaluating QA goals in software projects is the Architecture Tradeoff Analysis Method (ATAM). However, despite its importance and wide acceptance, there are few reports of empirical studies on analyzing QA tradeoffs in SAS. In this sense, the present investigation proposes an adapted version of ATAM called ATAM-4SAS to deal with the particularities of SAS. To achieve the research goal, we employed the UPPAAL SMC (statistical verification model) to analyze a set of QA. To evaluate the feasibility of the proposed method, we performed an empirical study on the execution of the ATAM-4SAS in a SAS developed according to the MAPE-K model. This model encompasses the Monitoring, Analysis, Planning, and Execution phases. Such steps share a knowledge base (K), which is fundamental in supporting decision-making. We complemented the empirical evaluation by conducting a focus group, which sought to assess the perceived ease of use and the perceived usefulness of the ATAM-4SAS to support the strategic choice of QA in a SAS. As a result, we observed that most participants agreed that ATAM-4SAS provides adequate support for the strategic choice of QA in SAS.

Real-time systems (RTS) are those whose actions are subject to time constraints, defined in terms of execution time deadlines. A hard RTS is usually designed to never miss a deadline, as in the case of avionics and space systems. To guarantee this it is essential to know the maximum time that each task takes to execute, a parameter known as Worst Case Execution Time (WCET). Estimating the WCET of a task is not simple and, in the case of modern architectures, it is subject to unpredictable influences caused by different hardware and software elements, which prevent the derivation of an absolute  WCET value. One of the most used techniques in these cases is the Measurement-Based Probabilistic Timing Analysis (MBPTA) method based on the Extreme Values Theory (EVT) statistical area, for estimating worst-case probability distributions, called Probabilistic WCET (pWCET) . The present work aims to carry out an experimental analysis of the use of EVT via MBPTA for pWCET inference, presenting the process of applying the technique in a real RPi environment and exposing the challenges and flaws encountered during this process; a point rarely demonstrated in the literature. The partial results indicate that EVT is robust, but that has weaknesses because it does not always produce adequate models and
coherent pWCET results. This work is part of the Kepler project, a cooperation between UFBA and INRIA-Paris.

The objective of this work is to proposes a new hybrid SVR-ANN model for water demand forecasting. Where an adaptation of the methodology proposed by Zhang (2003) is used to decompose the time series of 10 reservoirs that supply the Metropolitan Region of Salvador (RMS). The data used are from the historical consumption from January/2017 to February/2022, obtained from the local supply company, Empresa Baiana de Águas e Saneamento, and meteorological data obtained from the National Institute of Meteorology of Brazil. The results demonstrated the feasibility of using the proposed model, compared to other traditional models such as the Multilayer Perceptron (MLP), Support Vector Regression (SVR), Short Long Term Memory (LSTM) and Autoregressive and Integrated Moving Average (ARIMA).

Ordering food online is present in people's lives every day, however, this procedure requires sending customer data to different restaurant systems, and often these restaurants share sensitive information with their partners without authorization, creating a security risk when sharing such information. data. Also, in many cases, when a person travels to a country other than their origin and wants to order food, an exchange transaction for local currency is required, which is expensive. Blockchains have revolutionized the way of making payments online and intrinsically secure, eliminating the centralizing and authorizing entity, such as Bitcoin, the first blockchain used on a large scale. However, the scientific literature records the difficulty in view of the high frequency of transactions in Bitcoin, the high transaction time, and also its high cost per transaction, in addition to the limitation of the Script language for application development. In this step, the Ethereum blockchain emerges as an alternative to Bitcoin, in an environment of more complex applications with a higher frequency of applications and number of transactions per second, but still far from the performance of solutions such as credit cards, which disadvantages this “chain of blocks”. ” compared to traditional payment solutions such as credit cards, which also had high transaction cost rates for a long time. From the study of this context, we propose a payment model for retail transactions through the blockchain based on a Proof of Concept (POC) aimed at restaurants using the Software Development Kit (SDK) of the Lisk blockchain. POC Lisk Restaurante represents an integration of Lisk blockchain in the food industry allowing customers to buy food quickly and securely through sidechains - exclusive blockchains for special restaurant-type transactions, without intermediation and with lower fees when using a single crypto asset in the world : the LSK. We performed an empirical study of the proposed solution in which the performance of block transactions, transaction fees, auditability, solution scale, privacy and security in sending messages in transactions is evaluated. The results were used in comparison with data from the literature of Bitcoin, Ethereum, and Multichain blockchains. The results indicate that the special transactions used at Lisk Restaurante allow guaranteeing data privacy for its customers, since only the recipient of a transaction can access sensitive sender data. Thus, through research, literature review and performance evaluation experiments, it was observed that the sidechain solution for restaurants demonstrated the efficiency of blockchain technology in providing global services at a lower cost than common means of payment; also, when compared to their transaction fees with other blockchains like Bitcoin or Ethereum, their values are lower and the privacy of their customers is maintained. Also, our solution has shown an increase in the capacity of Food transactions in a single block and better scalability through sidechains, which accept transactions specific to the business involved, which differs from single blockchains such as Bitcoin or Ethereum that execute transactions for different purposes, also surpassing the first version of the Proof of Concept (PoC) Lisk Restaurante, exceeding twice the capacity of number of transactions or frequency, in a block. Finally, experiments were carried out with the Lightning network, an off-chain technology that works on top of Bitcoin, being a scalability solution for Bitcoin. The evaluation of experimental results of the Lisk sidechain solution and the Lightning network demonstrated the differences of each technology and the advantages and disadvantages of each one were praised. Thus, it was observed that the Lightning network technology can also be used in the retail industry, but without the possibility of customizing transaction costs.

Neighborhood detection in ad hoc vehicular networks involves determining neighboring vehicles for collaborative organization. This problem presents challenges to communication between vehicles in vehicular networks due to the dynamic nature of the network, which leads to unpredictable variations in the quality of the communication channel due to factors such as environmental conditions, interferences, and congestion. In this dissertation, we address the neighborhood detection between platoons in dynamic networks, facing challenges such as varied vehicle movements and building interferences. The platooning organization represents a relevant scenario for studying vehicular networks, as it brings benefits to traffic quality, such as fuel consumption reduction, traffic flow improvement, and increased safety. We have developed a basic communication protocol designed for communication between platoon leaders, which ensures mutual knowledge between them, enabling the exchange of essential information for neighborhood detection and collaborative operation of the platoons. We have verified the protocol using the PRISM model checker. Using this protocol, we conducted experiments to determine the required times that occur between the sending and receiving of messages exchanged between platoon leaders for neighborhood determination. We performed experiments on both reliable and unreliable channels and determined progress time bound through simulations using the VEINS framework, which integrates SUMO and OMNET++. The results of the experiments provided information about the progress time bound in various scenarios that can serve as a reference for the development of applications related to communication in VANETs

Multidimensional visualizations are graphical representations that assist in presenting multidimensional datasets. However, many of these techniques exhibit limited visual scalability as the number of dimensions increases, requiring ever-expanding visual space to accommodate the entire dataset. One way to address this problem is by utilizing Multidimensional Projection techniques, which perform dimensionality reduction while aiming to preserve the patterns of the original data in the projected space. However, it is highly likely that errors and distortions may be present in the projection layouts. A challenge in researching this topic is to measure and visualize the errors and distortions embedded in the projection mappings. In this regard, this article introduces MultiVisD3, a Visual Analytics approach to analyzing the quality of Multidimensional Projections. This approach comprises an interactive visual dashboard containing multiple coordinated views to facilitate understanding and comparison of error metrics between two projections. Other visualization techniques such as Treemap and tabular views support the extraction of information regarding projection quality. Finally, the proposed approach underwent user evaluation using well-known datasets to analyze functionality and usability aspects. The results demonstrated that users positively evaluated the proposed approach. More than 90% of the participants correctly completed eight out of nine tasks assigned to them, and over 90% of them approved of usability aspects related to interactions and the arrangement of visualizations.

In today's dynamic digital realm, the plethora of listing services, spanning from music platforms to product recommenders and social media content suggestions, often leaves users searching for items that truly align with their tastes. Addressing this intricacy, the rise of Recommender Systems has proven invaluable. These systems efficiently filter vast data to align items with individual preferences, enhancing user choices. This work centers on the creation of an advanced Recommender Systems API. Distinctively crafted, this API boasts universal accessibility and an uncomplicated deployment procedure. As the foundation for various Web Services, the API draws strength from the stalwart REST architecture. It is designed with a commitment to modularity, championing adaptability and flexibility. The API processes user data and queries to provide tailor-made recommendations quickly. Performance evaluations illuminated the API's commendable accuracy. It particularly shone with smaller datasets, displaying impressive data processing and algorithm execution times. The API manifested exceptional efficiency and resilience under specific test conditions, including cloud environments, especially notable in extensive 16,000-item dataset scenarios. The API is more than a tool; it paves the way for personalized digital experiences, showcasing its prowess in CRUD operations and tailored recommendations. The user evaluation phase encompassed a varied demographic: novice to experienced developers. Over half had considerable software development experience, and a significant percentage had prior engagements with coding recommender systems. With diverse knowledge of recommender libraries, most feedback praised the API's effectiveness. 81% valued the recommendations, and many felt confident in its filtering techniques. The highlight of this work is the Recommender System API's versatility. Despite positive feedback, users suggested improvements in documentation, data security, and features. These insights will shape future API refinements and user experience. Participants' enthusiastic engagement and feedback underscore the API's potential to enhance applications requiring a recommendation system, especially for developers who are perhaps less versed in the theoretical nuances. The solid research foundation and participant dedication highlight the API's potential for broader adoption by developers.

Refactorings are operations performed on source code aimed at improving the maintainability of a software system. Although literature contains a high number of studies on refactorings, there are only few works that investigate the motivations reported by developers to perform refactorings. Therefore, this research aims to investigate the motivations behind extract method refactorings in real systems. To achieve this, we conducted two empirical studies: a preliminary study involving a single software system, and another study considering multiple real systems. The studies were based on mining software repository commits in which extract method refactorings occurred. Essentially, the goal was to analyze commit messages in order to capture developers' motivations for performing such refactorings. The results identified 16 different motivations for applying extract method. Additionally, it was observed that only 16% of the refactorings had an explicit intention of improving code quality. It was also identified that refactorings with the intention of improving code quality occurred more frequently in methods with size higher than 61 lines of code and less frequently in methods smaller than 20 lines of code.

Commit bursts are sequences of changes made by developers that occur in code within a short period of time. In projects that adopt the practice of Continuous Integration (CI), every time a modification is completed, a new version of the code is created, generating a new build. In this new version, the changes made are verified automatically, running unit tests and reporting the result of failure or success of build to the developers. In this sense, the objective of this research is to carry out an empirical study to verify the association between bursts of commits and failures in build. The first step was to carry out an empirical study based on the mining of deposits, through which the relationship between bursts of commits and construction failure was identified. After this study, a survey was followed to evaluate the results obtained, considering the opinions of developers who work with CI. Among the discoveries made, the results of the first study show that in some projects the success rate of builds after bursts of changes tends to decrease. However, we cannot generalize the results to all projects since, in most of the studied projects, the difference was not statistically significant. In the second study, most survey participants agree that the proximity of a project's delivery data is a factor responsible for bursts of commits and build failure. However, we cannot generalize the results, since some developers had small disagreements about who was responsible for the construction failure. Thus, the results of this work intend to contribute to the community of developers who use CI, helping them to reduce failures in build, facilitating future work on good development practices.

With the growth in the number of vehicles in the world in recent years, it has become necessary to adopt technologies to deal with the consequences that this vehicle volume can generate for large cities, such as increased congestion on highways. Vehicular Ad-Hoc Networks (VANETs) present themselves as a promising technology in this scenario, helping to form vehicular networks capable of interconnecting vehicles and infrastructure to understand and deal with vehicle congestion. Considering this, the amount of data generated by this environment increases as the number of vehicles on the roads increases. Consequently, sending data from the vehicular environment to the structure that identifies congestion can be increasingly costly from the point of view of network use, potentially generating overloads and increased latency, making quick decision-making difficult. Therefore, in this work, we propose the construction of a Framework that aims to identify vehicle congestion, with an approach to reduce the data generated by a VANET in the fog layer and then send only the most relevant data to the cloud. for decision making. In addition to congestion detection, with historical data in time series format we perform congestion prediction using ARIMA. To work with data reduction, \textit{Framework} uses simple sampling algorithms and clustering techniques (DBSCAN and XMEANS). The results showed that the use of clustering algorithms in this Framework can achieve a significant level of accuracy in detecting traffic congestion together with a marked reduction in network usage.

Context: Technical debt (TD) describes the effects of immature artifacts on software development that can bring benefits in the short term but may have to be paid with interest in the long term. TD management balances short-term and long-term goals, supporting development teams to decide on the need and the best time to eliminate the debt. TD management activities include prevention, monitoring, and payment. Through prevention, it is possible to prevent teams from incurring TD, while monitoring helps them follow the evolution of TD items concerning the cost-benefit of eliminating them or not, that is, paying the debt items. Knowing the practices used to prevent, monitor, and pay TD items can help development teams to choose the best practice to be used in their projects. Identifying the practice avoidance reasons (PARs) that lead to nonprevention, non-monitoring, and non-payment of TD can help teams understand which aspects need to be improved to enable TD management. Although the technical literature has investigated the prevention, monitoring, and payment of TD, current results only reflect the viewpoint of a small number of professionals and organizations. To achieve the benefits of TD management, it is necessary to investigate more deeply the practices and PARs associated with these TD activities.

Aims: This Ph.D. dissertation aims to investigate, through the continuous and independent replication of a family of surveys conducted globally, the state of practice on the prevention, monitoring, and payment of TD items in software projects.

Method: Initially, we conducted a literature review on the current state of research on TD and its prevention, monitoring, and payment. Then, we analyzed data collected by six replication teams from the InsighTD project, which is a family of globally distributed surveys on the causes, effects, and management of TD. From the body of knowledge resulting from the analysis of InsighTD data, we defined three artifacts: an updated version of the conceptual model for TD, a set of conceptual maps, and IDEA (Impediments, Decision factors, Enabling practices, and Actions) diagrams. Finally, we assessed these artifacts through case studies in academic and industrial settings.

Results: This Ph.D. dissertation presents the leading practices used to prevent, monitor, and pay off TD items and the PARs that justify the non-application of these practices. Regarding the prevention of TD, well-defined requirements, adopting good programming practices, and better project management are among the five most cited practices related to prevention, while short deadlines, ineffective management, and lack of predictability in the software development are among the five most cited PARs to justify the non-prevention of debt. About TD monitoring, TD item backlog, use of tools, and team meetings are among the five most cited practices related to monitoring, while lack of interest, focus on short-term goals, and lack of time are among the five PARs used to explain the non-monitoring of TD items. Regarding TD payment, code refactoring, investing effort in TD payment activities, and design refactoring are among the top five payment-related practices, while focusing on short-term goals, lack of organizational interest, and lack of time are among the five most cited PARs to explain the non-payment of TD. We update the conceptual model for TD by including the knowledge we learn from the state of practice and organize all practices and PARs along with their types, natures, and categories into maps and IDEA diagrams. From the conceptual model and TD payment map assessment, we found that they are well organized and provide valuable information to define strategies for TD management. The IDEA diagrams assessment provided positive evidence that the diagrams are easy to read and follow and can influence decisions on how to manage TD items.

Conclusion: Using the InsighTD data, this Ph.D. dissertation explores the state of practice on TD prevention, monitoring, and payment, revealing the primary practices used to perform these activities and the PARs that avoid their execution. All body of knowledge was organized into three artifacts that can drive new investigations on TD and support software practitioners in increasing their capabilities and reducing their issues in managing debt items.

The demand for memory has increased due to various applications such as big data, machine learning, social networks, and streaming analytics. DRAM technology is facing issues with scalability, energy, and costs, despite its low latency advantage. Non-Volatile Memory (NVM) is an alternative to DRAM scalability issues. Although NVM has high density, low cost, and low energy, its high latency prevents it from replacing DRAM. Heterogeneous memories like DRAM+NVM are likely to become common, and efficient data placement is an important research question. This PhD dissertation focuses on managing data placement for graph analytics applications. It studies machine learning models to predict application performance in a context with multiple applications and explores data mapping choices at the object or chunk level for a single application. The dissertation contributes to offline identification of performance-influencing object attributes, implementation of an offline scheduler at the chunk level, and implementation of an online scheduler at the chunk level.

Conditional compilation is often used to implement variability in configurable systems. This technique relies on #ifdefs to delimit feature code. Previous studies have shown that #ifdefs may hinder code comprehensibility. However, those studies did not explicitly take feature dependencies into account. Feature dependency occurs when different features refer to the same program element, such as a variable. Comprehensibility may be even more affected in the presence of feature dependency, as the developer must reason about different scenarios affecting the same variable. Our goal is to understand how feature dependency affects the comprehensibility of configurable system source code. We conducted four complementary empirical studies. In Study 1, forty-six developers responded an online survey. They executed tasks in which they had to analyze programs containing #ifdefs with and without feature dependency. However, feature dependencies may be of different types depending on the scope of the shared variable. In Study 1, we were not concerned with different types of dependency. Thus, in Study 2, we carried out an experiment in which 30 developers debugged programs with different types of feature dependency. Each program included a different type of feature dependency: global, intraprocedural or interprocedural. We used an eye-tracking device to record developers' gaze movements while they debugged programs. However, feature dependencies do not differ from each other only in terms of types. They can also present differences in terms of number of dependent variables and degree of variability, among others. To address those characteristics, we complemented Study 1 and 2 by means of Studies 3 and 4. In Study 3, we executed a controlled experiment with 12 participants who analyzed programs with different numbers of dependent variables and number of uses of dependent variables. In Study 4, we carried out an experiment in which 12 developers analyzed programs with different degrees of variability. Our results show that: (i) analyzing programs containing #ifdefs and feature dependency required more time than programs containing #ifdefs but without feature dependency, (ii) debugging programs with #ifdefs and global or interprocedural dependency required more time and higher visual effort than programs with intraprocedural dependency, (iii) the higher the number of dependent variables the more difficult it was to understand programs with feature dependency, and (iv) the degree of variability did not affect the comprehensibility of programs with feature dependency. In summary, our studies showed that #ifdefs affected comprehensibility of configurable systems in different degrees depending on the presence or not of feature dependency, on the type of feature dependency, and on the number of dependent variables.

Mapas Conceituais são ferramentas pedagógicas amplamente utilizadas para facilitar o
processo de aprendizagem. De forma geral, podem ser definidos como diagramas com
organização hierárquica que apresentam relações entre conceitos. São um importante
instrumento para o ensino e avaliação do conhecimento por representarem, entre outras
características, a estrutura cognitiva do aprendiz. Os mapas conceituais estão fundamentados
na Teoria da Aprendizagem Significativa de Ausubel, assim, é possível avaliar os mapas com
base em aspectos dessa Teoria, como as categorias-chave da diferenciação progressiva e
reconciliação integrativa. Estas representam informações sobre o processo de construção do
conhecimento sendo, portanto, fundamentais para uma avaliação de aprendizagem
significativa. As categorias-chave podem ser identificadas no processo de elaboração dos
mapas, no entanto, este processo não vem sendo muito explorado na literatura. A maioria dos
trabalhos focam na análise comparativa entre o mapa do aprendiz e um mapa de referência
para verificar se os conceitos e relações estão representados corretamente. A análise
comparativa entre mapas é importante, mas não suficiente para uma avaliação da
aprendizagem significativa, visto que as categorias-chave bem como outras informações da
estrutura cognitiva do aprendiz, ou seja, como ele constrói e reconstrói seu conhecimento no
decurso do seu processo de aprendizagem, são essenciais para identificar indícios de
aprendizagem. Na literatura, há trabalhos que analisam a estrutura do mapa conceitual para
classificar a qualidade da aprendizagem em aprendizagem mecânica, significativa ou não-
aprendizagem, mas não analisa o processo de construção do mapa e nenhuma discussão é
feita sobre a automatização da avaliação. Pesquisas no campo da Informática na Educação têm
investigado o problema de se avaliar a aprendizagem, comparando mapas conceituais de
aprendizes com um mapa de referência, por meio de ferramentas computacionais. Há diversas
soluções que envolvem análise estrutural utilizando padrões de comparação e análise
semântica por meio de ontologias de domínio. Esta tese de doutorado apresenta um modelo
e um método para avaliação de mapas conceituais que considera avaliação sintática e
semântica bem como o processo de construção de mapas, propondo critérios para identificar
indícios de aprendizagem mecânica, significativa e não-aprendizagem. São consideradas
avaliação qualitativa e quantitativa. Um framework conceitual é especificado fornecendo um
arcabouço sobre a organização e estruturação de avaliação de mapas conceituais para apoiar a
construção de sistemas computacionais para a sua automatização. Foram desenvolvidos
estudos de caso, que serviram como uma primeira avaliação do método proposto.

A 1-factorization is a partition of the edge set of a graph into perfect matchings. The concept of 1-factorization is of great interest due to its applications in modeling sports tournaments. Two 1-factorizations are said to be isomorphic (belong to the same isomorphism class) if there exists a bijection between their sets of vertices that transforms one into the other. The non-isomorphic 1-factorization search space is a graph in which each isomorphism class is represented by a vertex and each edge that connects the vertices $\mathcal{F}_a$ and $\mathcal{F}_b$ corresponds to a move in a neighborhood structure, which from a 1-factorization isomorphic to $\mathcal{F}_a$ generates a 1-factorization isomorphic to $\mathcal{F}_b$. An invariant of a 1-factorization is a property that depends only on its structure such that isomorphic 1-factorizations are guaranteed to have equal invariant values. An invariant is complete when any two non-isomorphic 1-factorizations have distinct invariant values. This thesis reviews seven invariants used to distinguish non-isomorphic 1-factorizations of $K_{2n}$ (complete graph with an even number of vertices). Additionally, considering that the invariants available in the literature are not complete, we propose two new ones, denoted lantern profiles and even-size bichromatic chains. The invariants are compared regarding their sizes and calculation time complexity. Furthermore, we conduct computational experiments to assess their ability to distinguish non-isomorphic 1-factorizations. To accomplish that we use the sets of non-isomorphic 1-factorizations of $K_{10}$ and $K_{12}$, as well as the sets of non-isomorphic perfect 1-factorizations of $K_{14}$ and $K_{16}$.
We also consider algorithmic and computational aspects for exploring the generalized partial team swap (GPTS) neighborhood, a neighborhood structure for round-robin sports scheduling problems recently proposed in the literature. In this regard, we present a framework to explore the GPTS neighborhood. Furthermore, a discussion is presented on how this neighborhood structure increases the connectivity of the search space defined by non-isomorphic 1-factorizations of $K_{2n}$ (for $8 \le 2n \le 12$) when compared to other neighborhood structures.

Early-stage startups, often founded by small teams with innovative yet unproven ideas, face the challenge of validating their concepts in the market. These ideas take shape through a fluid business model, adapting until a repeatable and scalable approach is established. Startups must grapple with the pressure to deliver a minimum viable product or service swiftly. For software startups, this often centers around software or technology-mediated offerings, operating within a competitive landscape alongside companies of various sizes. Recent research has explored software startups' context, objectives, challenges, and practices. While extensive research has delved into how software startups engage in software development, there remains a gap in understanding how they select and implement specific development practices. Uncovering their software development processes and methodologies is equally essential, as it directly impacts their ability to overcome software development challenges. Key questions include: What characterizes software development in startup environments? How do startups prioritize product quality attributes in their development processes? What software engineering practices support their development endeavors, and which tools aid their progress? This thesis seeks to gain insights into how early-stage software startups navigate the software development journey, shedding light on their priorities, processes, and the tools they leverage. Our research adopts a qualitative approach in the form of a multiple case study involving 14 organizations. We analyze the data employing Grounded Theory techniques, including open, axial, and selective coding. The findings underscore the critical relevance of human factors, software development processes, software engineering practices, and external influences in the software development journey. Understanding these factors has led to the formulation of practical recommendations to bolster early-stage software development within startups. This work identifies pivotal factors influencing software development within startups and offers practical recommendations for their early-stage software development endeavors.

We propose a method that can perform one-class classification given only a small number of examples from the target class and none from the others. We formulate the learning of meaningful features for one-class classification as a meta-learning problem in which the meta-training stage repeatedly simulates one-class classification, using the classification loss of the chosen algorithm to learn a feature representation. To learn these representations, we require only multiclass data from similar tasks. We show how the SVDD method can be used with our method, and also propose a simpler variant based on Prototypical Networks that obtains comparable performance, indicating that learning feature representations directly from data may be more important than which one-class algorithm we choose. We validate our approach by adapting few-shot classification datasets to the few-shot one-class classification scenario, obtaining similar results to the state-of-the-art of traditional one-class classification, and that improves upon that of one-class classification baselines employed in the few-shot setting. Moreover, as a practical application, we employ our method to the biometric task of on-device face verification. In this scenario, it compares unfavorably to the state-of-the-art metric learning technique.

Cloud computing provides a centralized, service-oriented infrastructure for users around the world. A cloud environment allows running critical applications sensitive to user interaction (such as an online web search), and is also used to run best-effort oriented batch applications (such as audio/video compression, indexing of web pages). Data centers are designed on a large scale to support the various cloud services, distributed across hundreds of thousands of servers, with significant operational and capital expenditures. Thus, optimizing the use of server resources results in significant savings on these systems. As a research opportunity, there is the possibility of exploring idle moments of servers that run critical services, especially during periods of low demand. This research aims to co-locate different types of cloud application tasks (latency-sensitive and best-effort) on the same server in order to increase its utilization, thus improving its efficiency. The work demonstrates that performing co-allocation of latency-sensitive and best-effort tasks provides greater efficiency for cloud servers, as long as the performance of critical application tasks is guaranteed. The work explored the use of specific schedulers for each type of task, latency-sensitive tasks were scheduled and parameterized via real-time scheduling, and batch application tasks used fair best-effort scheduling. The gains in relation to the standard scheduler of the system reached up to 150% when compared to the one proposed in this work, here called EDFCoaloc.

The design and implementation of effective Self-Adaptive Systems (SAS) is a very challenging task, usually due to the existence of complex dynamics between the various elements that make up the system and between these elements and the environment in which they operate. A crucial point, in the SAS project, is the execution of the planning activity (planning): definition of which actions (adaptation plans) must be carried out so that the system adapts and returns to meet the quality levels expected service. Recent research has investigated the use of formal verification techniques to enable the generation of reliable adaptation plans. Through the model checking technique (model checking) and the MAPE-K reference model, we developed the design, implementation and validation of an architecture that defines self-adaptive structures and behaviors using formal models, for the specification planners that generate adaptation plans autonomously, considering self-adaptive properties such as self-organizing and self-healing. The architecture evaluation was carried out through a case study related to the traffic light control problem, with the creation of a planner, in which the adaptation plans were generated, in an off-line way, with the verifier of UPPAAL models and executed in run-time through an implementation of this work. Some application scenarios were simulated, using the SUMO tool, to evaluate the effectiveness of the solutions found. The created planner generated efficient adaptation plans, autonomously, in scenarios where the number of possible adaptation plans was around 2¹². The results indicate that the solutions adopted with our planner were more effective compared to solutions without any self-adaptation approach, when the vehicle flow approached collapse.

Due to the impacts caused by the advent of the new Coronavirus (SARS-CoV-2), the mobilization and popular participation in the face of public health policies are fundamental for facing Covid-19 and mitigating its damages, which is widely affected knowledge, attitudes and practices in relation to the disease. However, in view of the context of people in a state of social vulnerability and with low access to education and information, the absence of participation and engagement in socioeconomically disadvantaged communities presents itself as an obstacle that stands in the way of the objectives proposed by health policies. . Thus, the present work proposes the implementation of the GAFCC theory-oriented gamification design model on the +Lugar platform (Covid-19 version), to support participatory crowdsourcing approaches and participatory mapping in favor of initiatives aimed at coping with Covid-19. -19, mainly the mapping and signaling of disease transmission predictors in real time, through a risk perception approach. In order to assess the acceptance and use of the proposed solution, UTAUT was used to identify the determining factors that have a significant impact on the intention to use the platform. To this end, a validation study was carried out with 20 young people and adolescents living in the Marechal Rondon neighborhood, where a form was applied, consisting of questions related to moderating and determining factors, and behavioral intention. The result of the study pointed out that the respondents have high levels of performance and effort expectation, social influence, facilitating conditions, and behavioral intention to use the platform. However, through the analysis of Spearman's correlation coefficient, it was possible to identify that only the factors “effort expectancy” and “social influence” significantly impact behavioral intention.

For a long time, electrocardiograms (ECGs) have been used to diagnose heart problems. However, obtaining a satisfactory classification through the technology of digital systems applied to e-health medicine is a difficult task. To study this problem, it was considered that the heart is a reactive system with some possible modes or states of functioning, so it is intuitive to use computational tools such as finite state machines (FEMs) to describe the health status of this human organ. In this paradigm, we propose the Cyber-ECG, a computational model for automatic classification of ECG signals with detection of electrocardiograph or sensor failures. To accomplish this, the Simscape and Stateflow libraries from Simulink were used in the Matlab ecosystem. In addition, a prototype for cardiac pulse detection is presented, using the Medical Internet of Things (IoTM) technology. This prototype consists of an ESP32 board, to collect, aggregate and visualize data on a person's heart pulse in the cloud. The proposed Cyber-ECG tool obtained an accuracy of 84% and 80% when classifying the ARR and NSR classes, respectively. For these same classes, the F1-Score values are approximately 82% and 83%. The precision was 82.5% and the sensitivity was 80%, respectively. Based on these results, the tool 

Recommender Systems provide suggestions for items that are most likely of interest to users. Providing personalized recommendations is a challenge that can be addressed by filtering algorithms among which Collaborative Filtering (CF) has demonstrated much progress in the last few years. By using Matrix Factorization (MF) techniques, CF methods reduce prediction error by using optimization algorithms. However, they usually face problems such as data sparsity and prediction error. Studies point to the use of data available in Semantic Web as a path to improve recommender systems and address the challenges related to CF techniques. Motivated by these premises, the present work developed a data pipeline along with an algorithm that processes the Ratings Matrix combining semantic similarities of Linked Open Data (LOD) and estimates missing rat- ings. The experiments take subsets of three different datasets (Movielens, LastFM and LibraryThing), two semantic similarity metrics, Linked Data Similarity Distance (LDSD) and Resource Similarity (RESIM), and three MF-based algorithms (SVD, SVD++ and NMF). Our experiments reduced sparsity by more than 75% in Movielens subset and 28% in LastFM. Prediction error is reduced in all subsets with statistical confidence using parametric test one-way ANOVA followed by Tukey’s multiple comparison test.

Technical Debt (TD) contextualizes a set of tasks that are not performed and are accumulated to be performed later during software development. In software testing, insufficient test coverage, lack of testing, and lack of test case planning are situations in which test debt can be found. In contrast, other types of debt can affect the quality of testing activities. Outdated requirements documentation is an example of a requirements debt situation that can compromise test planning activities. In this scenario, the term test-related debt (TRTD) refers to the problems faced by test professionals due to the presence of TD items. Recently, a conceptual map has been proposed in the technical literature that relates test activities to TD items. This map presents test-related TD causes, indicators and payment practices. However, this map was based on an ad hoc literature review and experimental evidence on its feasibility for use in industry is limited. Despite several works investigating indicators, causes, effects and prevention of TD, there is still a long way to go on how to use this information to support the management of debt items in the context of testing activities. The objective of this work is (i) to present an updated version of a conceptual framework that organizes a set of indicators, causes, effects and preventive practices of TRTD, based on data from the InsighTD project, which is a global questionnaire, which seeks to generate experimental data on TD, its causes, effects and management from software engineering professionals and (ii) to evaluate it for its ease of use, usefulness and possible future use. This evaluation methodology, based on the technology acceptance model (TAM), ran a feasibility study of the maps with 95 participants. As a result of the research, it was observed that the conceptual framework with TD elements can be useful to support the management of TRTD items, as pointed out by 89% of the participants. Most participants also indicated that they would gain productivity, performance, agility and effectiveness using the proposed framework. The framework was also considered easy to learn and use by the study participants.

In recent years the availability of data on the Web has been driven by the increasing use of social networks and smartphone applications. In addition to textual content, geo-location information is also shared, favoring the emergence of numerous location-based services. The location information of a Point of Interest (POI) can be used to understand the profile of users, their interests, and movements. In this way, it is possible to identify, for example, places of interest for a particular user, and even classify them into categories, such as cafes, universities, bars, malls, etc. This type of data is widely used for Points of Interest Recommender Systems, which aim to assist users in the search for places of interest, whether on a daily basis or during a trip. These systems are traditionally recommended for individuals, however, there are scenarios where individuals gather in groups, thus increasing the complexity of the problem. In addition to the need to find individual preferences, the recommendation must consider the preferences of the group as a whole, which requires the application of a consensus technique. Another obstacle is that non-diversified recommendations tend to always be in the same category, decreasing the group's interest in recommendations from already known Points of Interest. This master's thesis proposes a Points of Interest Recommendation System for Groups using diversity. To evaluate the proposed model, an exhaustive experiment was carried out with 19 groups, some with 3 and others with 5 members. To evaluate the diversified recommendations, precision metrics were used in positions 3, 5, and 10. According to the results, the recommendations for positions 5 and 10 obtained more satisfactory results when diversity was applied. After the experiment with real users, an offline analysis was also performed with variations of the proposed model and aggregation techniques. According to the results obtained, it was possible to verify that the recommendation models with diversity obtained better results than the non-diversified approach in most of the tested configurations.

Recommender Systems traditionally recommend items to individual users. However, there are scenarios where individuals gather in groups, and with that arises the need to recommend to groups. Most of these groups form naturally, for example, to watch a movie, have lunch at a restaurant, or even plan a trip. In all these hypotheses, it is possible to use Recommender Systems to offer personalized information to the group as a whole. To do so, it is necessary to consider the individual preferences of the group members, to satisfy them fully, and, in this sense, to use techniques for aggregating this information. Although there are consensus techniques for aggregating information, the recommendations can be repetitive among themselves, as they will always serve the same group profile. This inconvenience sets a precedent for adopting diversity techniques for recommendations to the group. In this work, we investigate how to apply such diversification techniques in group recommendations, based on the members' preferences, to avoid over-specialization of the system and thus keep the group members satisfied in general. To do so, it is necessary to develop a group formation model and then model how the recommendation will be carried out to be able to apply diversification techniques to the items to be recommended. Two experiments were carried out in this research, considering related work as a baseline. In the offline experiment, the result of the proposed model was slightly higher than the baseline, having been 3.8% more accurate and 1.8% more diverse. In the second experiment, the online experiment, the proposed model completely dominated the baseline in terms of accuracy and was slightly inferior in terms of diversity.

Over the years, Cloud Computing has provided infrastructure for processing applications and services. However, the distance between Cloud computing resources and Internet of Things (IoT) devices can cause delays in response times for some services and applications that have sensitivity and temporal constraints. In this context, Fog Computing emerges as a computational paradigm that seeks to bring Cloud Computing resources closer to IoT devices, geographically distributing micro-data centers (Cloudlets) and reducing response time. Studies using Fog Computing present solutions for the application allocation, addressing that it is possible to meet different types of requirements. These studies evaluate certain requirements for decision making, such as response time and communication cost. However, it is observed in these studies that as a requirement is prioritized, another is neglected. Thus, this work proposes an approach for allocating modular applications in a Fog Computing architecture with multiple hierarchical levels, which seeks to reduce the response time of latency-sensitive applications and reduce data traffic on the network. To achieve this goal, the proposed approach, Least Impact - X (LI-X), seeks to reduce the idleness of resources at the lowest levels of the fog and takes into account the communication between application modules when deciding on the allocation. LI-X was compared to predecessor studies in a simulated iFogSim environment. The results show that the LI-X was able to overcome these studies in most of the proposed scenarios, reducing the response time and the cost of data communication on the network.

In recent years, Information-centric networking (ICN) has gained attention from the research and industry communities as an efficient and reliable content distribution network paradigm, especially to address content-centric and bandwidth-needed applications together with the heterogeneous requirements of emergent networks, such as the Internet of Things, Vehicular Ad-hoc NETwork, and Mobile Edge Computing. In-network caching is an essential part of ICN architecture design, and the performance of the overall network relies on caching policy efficiency. Therefore, a large number of cache replacement strategies have been proposed to suit the needs of different networks. The literature extensively presents studies on the performance of the replacement schemes in different contexts. The evaluations may present different variations of context characteristics leading to different impacts on the performance of the policies or different results of most suitable policies. Conversely, there is a lack of research efforts to understand how the context characteristics influence policy performance. There is also a lack of initiatives to assist the process of choosing a suitable policy given a specific scenario. In this direction, this thesis address those research gaps by (i) pointing out what is context from the perspective of cache replacement policies and the context characteristics that influence cache behavior, and (ii) proposing a caching meta-policy strategy to assist the choosing process of suitable policies according to the current context. For the context delimitation study, we have conducted an extensive survey of the ICN literature to map reported evidence of different aspects of context regarding the cache replacement schemes. Beyond the contribution of understanding what is context for caching policies, the survey provided a helpful classification of policies based on the context dimensions used to determine the relevance of contents. Moreover, as an investigation of holistic aspects to represent context, and motivated by the emergent area of human-centric networking, we have performed an exploratory case study on a human behavior influence over the policies performance. To accomplish such goal, we carry out a simulation-based study that evaluated the performance of cache replacement policies through clusters formed by users according to their music listening habits. The results fostered the evidence that distinct context aspects have an effect on caching policy performances. Following the context studies, we present a meta-policy strategy capable of learning the most appropriate policy for cache online and dynamically adapting to context variations that leads to changes in which policy is best. The meta-policy benefits from the diversity of policies and its context aspects, decouples the eviction strategy from managing the context information used by the policy, and models the choice of suitable policies as online learning with bandit feedback problem. The meta-policy can support the deployment of a diverse set of self-contained caching policies in different networks. It enables cache routers to work as adaptive systems agnostic to the underlying contexts, such as content request patterns or popularity variations. Experimental results in single and network of caches have shown the meta-policy effectiveness and adaptability to different contexts in synthetic and trace-driven simulations.

Recommendation Systems focus on recommending the relevant items to end-users. Usually, the most important items are the most popular ones. However, the emergence of new forms of content distribution in the most diverse markets implied the emergence of the long tail phenomenon, i.e., the creation of new niche products market. Considering the growth of groups related to niche markets, the study of recommendations in the long tail has become increasingly demanding. Nevertheless, long-tail items naturally tend to generate fewer inputs to information systems such as purchase logs, user feedback, and ratings. As a result, it becomes more difficult to recommend items from the long tail. Motivated by these premisses, the primary purpose of this thesis is to develop and exploit recommendation models capable of leading users to niche items located in the long tail, but at the same time highly relevant ones. For this, two major techniques of clustering and representation of matrices through graphs are explored. The first technique adopts Markov chains to calculate similarities of the nodes of a user-item graph. The second technique applies clustering to the set of items in a dataset. Such a combination aims to give more visibility to the long tail items. Experiments were carried out to evaluate the approach and measure the effectiveness of the recommendations considering the long tail context. Metrics such as recall, diversity, and popularity of the generated recommendations were calculated and compared with techniques whose objectives do not directly cover long-tail recommendations. In addition, a questionnaire applied to experts in the business domain also complemented the evaluation of the proposed approaches. By comparing our proposals against three state-of-the-art baselines, the results show that it is possible to improve the accuracy of the recommendations even by focusing on less popular items, in this case, niche products that form the long tail. The recall in some cases improved by about 27.9%, while the popularity of recommended items has declined. In addition, the recommendations show to contain more diversified items indicating better exploitation of the long tail.

One of the great challenges in MAS is the creation of cooperative plans to deal with the different scenarios that present themselves in a dynamic, real-time environment composed of teams of mobile robots. In this scenario, an agent of the MAS controls each robot, which needs to make complex decisions in a short time in a coordinated manner with the other robots on its team. Despite the many solutions developed based on multi-agent planning and reinforcement learning, a human expert in the problem domain usually sees opportunities for better cooperative plans in many scenarios where robots underperform. The research presented in this thesis consists of capturing the human expert's knowledge to demonstrate how robot teams can better cooperate in solving the problem they must solve. The human expert can indicate the situations in which a cooperative plan can better solve a given problem by watching the performance of a team of robots in action.
Consequently, a dataset for training the agents that control the robots can gather the various human observations. For the development of this research, this work used the environment 3DSSIM and the collection of human demonstrations was carried out through a set of tools developed from the adaptation of existing solutions in the RoboCup community using a strategy of crowdsourcing. In addition, fuzzy clustering was used to gather expert demonstrations (setplays) with the same semantic meaning, even with small differences. With the data organized, this thesis used a reinforcement learning mechanism to learn a classification policy that allows agents to decide which group of setplays is best suited to each situation that presents itself in the environment. The results show the ability of the robot team to evolve, from the learning of the suggested setplays and its use in an appropriate way to the abilities of each robot.

Crowdsourcing is a new concept associated with problem-solving where the crowd is its main element. Knowing what drives this audience to engage in crowdsourcing projects is a challenge, as is building projects that attract this audience. Our studies focused on two aspects. First, in the investigation of the planning and development of crowdsourcing projects, and second in the identification of the factors that lead people to collaborate with these projects. Crowdsourcing has a dynamic structure where knowledge, problem- solving, and sometimes decision-making can be attributed to unknown members of the crowd. These unique characteristics of crowdsourcing must be considered in its building process. In this thesis, our goal is to investigate crowdsourcing systems from two perspec- tives. From the developer’s perspective, we investigated the construction of crowdsourc- ing. And from the user’s perspective, we investigated the factors that motivate users to collaborate on these projects. To achieve this goal, we carried out several studies, using an exploratory qualitative research method. The investigation of the construction process and the investigation of the users’ motivation followed the same procedures: (i) studies selected by literature review, and (ii) semi-structured interviews with developers and users. As the older adults’ population is constantly growing and is little studied in tech- nological field, our motivation study focused on this public. We conducted interviews with 46 older adults over 60 years of age. The interviews’ analysis suggest that older adults are motivated by dynamic tasks, which stimulate their curiosity, increase their knowledge and skills, tasks that help them to care for their physical and mental health and, mainly, by tasks related to altruistic causes. With the results of the studies, we proposed a process model for the construction of crowdsourcing and a motivation model aimed at older adults, both to be used as a guide by managers and software developers.

Contextualization: Natural disasters lead to damage and loss of life around the world. Early Warning Systems (EWS) are systems that send alerts, in advance, to warn the population about these events. Prediction institutions send information about events that can cause disasters in a standard file format known as CAP. In general, an EWS receives a CAP file, extracts its data, and sends warnings to the population. Most EWS does not consider the specific needs of vulnerable groups, and send the same alert to everyone, whether they are, e.g., visually or hearing impaired people, or are in an area with a higher level of risk for the event. Objective: Given the context, the aim of this study is to purpose an EWS to warn people (vulnerable groups or not) who are in risk areas. Method: This research was carried out through a combination approach of the following methods: Systematic Mapping Study, exploratory studies, and empirical studies. Exploratory studies, through semi-structured interviews, served to achieve more familiarity with the research problem. Based on these studies, a context model, behavioral rules, and an automated process for creating an accessible resources library were proposed, in order to enable the development of an EWS that sends alerts to different groups of users. Finally, the evaluation was carried out from two perspectives: (i) evaluating whether the alerts are delivered to the expected recipients and with the expected personalization; (ii) evaluating the context model and behavioral rules. For evaluation (i) an empirical approach was used through quantitative techniques. For evaluation (ii), a survey was conducted with Civil Defense experts from the states of Brazil. Results: This study presents an architecture and a context model with contextual rules for an EWS that considers people from vulnerable groups (for example deaf or blind people). Finally, it presents an implemented and functional prototype, the UAware Alert, with a management interface for sending alerts and a mobile app for receiving the warning in an accessible way. Conclusion: EWS uses CAP messages and the person’s location to send alerts. But to alert people who have different needs, it is also necessary to consider the profile of the person who will receive the alert and their location. Thus, the context model, behavioral rules, and media generation process of this study enables the development of an EWS, which sends instructions focused on the region where the person is located, either in text or in formats accessible to the needs of each user.

Context: Teaching Software Engineering (SE) in higher education is a challenging task. A solid theoretical base should be provided to students, including basic SE concepts (requirements, design, tests, among others), and topics related to project management, human-computer interaction, configuration and release management. Practical experience, communication skills and professionalism are dimensions that should also be considered, including teamwork and ethics, to prepare students for a career in the software industry.

Free/open source software (FLOSS) has been used in Software Engineering Education (SEE) as a viable strategy to address the need for training students' technical and non-technical skills and prepare them to deal with the challenges posed by the evolving software industry.
Some studies report that the pedagogical use of FLOSS projects has brought benefits but also challenges that may hinder their effective adoption in SEE.

A common challenge is project selection, due to the diversity of available FLOSS projects and the complexity of the task. In general, the teacher must search for, manually or with the support of different tools (not necessarily designed for use in Education), one or more projects for pedagogical use in the course. Several criteria can be considered for the selection of a FLOSS project for pedagogical use in SEE, including project size, number of tests, programming language, community size and their best practices.

Problem: The selection of FLOSS projects for pedagogical use in SEE requires effort and experience, and there are few tools that support such complex task while providing clear and well-defined selection criteria to be used by the teacher. Thus, the burden of selecting FLOSS projects can hinder their use and adoption in the context of SEE.

Objective: This research aimed to investigate the usefulness of an automated approach to support the teacher in the selection of open source projects guided by criteria that take into account socio-technical aspects of the FLOSS project, for pedagogical use in SEE higher education.

Reasearch Methods: A literature review was carried out to identify criteria already in use for the selection of FLOSS projects. Then, a set of selection criteria was identified, documented, operationalized and implemented in an open source tool called FlossSearch.Edu. The first version of the tool was evaluated by students and professors from different Brazilian higher education institutions. Initially, students used FlossSearch.Edu, in the context of a SE discipline, to select FLOSS projects based on criteria defined by the teacher, and then evaluated the use of the tool through a survey with questions based on the Technology Acceptance Model (TAM). In a qualitative and quantitative study, teachers already familiar with the pedagogical use of FLOSS projects, used FlossSearch.Edu in an individual session, recorded and guided by predefined scenarios, while reporting their impressions out loud (``Think Aloud''protocol). At the end of the sessions, teachers evaluated the use of the tool using a TAM questionnaire.

Findings: In the context of selecting FLOSS projects in which FlossSearch.Edu was used, the tool was well evaluated. The majority of students who used FlossSearch.Edu in the classroom and teachers who participated in the second assessment said that the tool is useful, easy to use and that they intend to use it in the future. Several suggestions for improvement were received and should guide the evolution of the tool and further studies.

Startups are small companies that seek to explore new businesses by embodying new technologies to different markets through innovation. Startups' Ecosystems exist to provide a supportive environment for these companies, being a valuable source of networking and knowledge to them. White label software projects are developed by Startups and known in Ecosystems as highly adaptable products, capable of generating new products faster than stand-alone applications, ensuring the best cost x benefit. Startups are divided into stages, and they face different challenges depending on the current stage. In the early stages, they barely plan their development activities but rather assess the market's needs and find users for their initial product, which leaves technical debt. White Label software projects suffer more from these issues since Startups do not address advanced code reuse techniques widely known by the academy. In the late stages, startups need to solve the technical debt managed in the early stages and need even more scalable processes to grow their business that involves White Label Software projects. This work seeks to unveil the concept of White Label software projects and their feasibility for software startups; investigate if and how advanced code reuse techniques, such as highly-configurable systems, could be used as opportunity-lever for software startups; build an adaptable framework for helping White Label software projects manage technical debt and face its challenges. In order to accomplish the aforementioned goals, we applied three empirical studies under different public. A semi-structured interview with startups using White Label software projects, an online questionnaire applied for stakeholders from ecosystems of innovation, and a Multivocal Literature Review with gray literature articles over the internet about White Label software projects. The yielded results presented a compelling portrait of how software startups have dealt with software reuse in their daily practices, with particular attention on how White Label software development has been explored in their projects. The data gathered allowed the creation of the framework above for mobile White Label software projects. In addition, this work gives a path of improvement that can be applied for startups working with White Label software projects under different stages, besides giving researchers a set of open challenges to be studied in future work.

Recommendation systems are tools used to suggest items that might be of interest to users. These systems are based on the user's preference history to generate a list of suggestions that are more similar to the items in the user's history, aiming at better accuracy and less error. It is expected that when recommending an item, the user gives a feedback to the system, indicating if he liked or how much he liked the recommended item. User interaction with the system enables a better understanding of users' tastes, which over time, add more and more items to their preference profile. The recommendation based on the similarity of the item with the preferences, seeking the best precision can cause side effects in the list, such as: overspecialization of the recommendations in a certain core of items, little diversity of categories and imbalance of category or gender. Thus, this dissertation aims to explore calibration, which is a means to produce recommendations that are relevant to users and at the same time consider all areas of their preferences, seeking to avoid disproportion in the recommendation list. For this, ways of weighing the balance between the relevance of the recommendations and the calibration based on measures of divergence are discussed. The hypothesis is that calibration can positively contribute to fairer recommendations according to user preference. The research is carried out through a broad approach that contemplates nine recommendation algorithms applied in the film and music domains, analyzing three of divergence measures, two custom balance weights and two balances between relevance-calibration. The assessment is analyzed with widely used metrics as well as proposed metrics. The results indicate that calibration has positive effects both for recommendation accuracy and fairness with user preferences, creating recommendation lists that respect all areas. The results also indicate which is the best combination to obtain the best performance when applying the calibration proposals.

ABSTRACT

Model Driven Development (MDD) and the Technical Debt (TD) metaphor are software

engineering approaches that look for promoting quality of systems under development.

While MDD does it through the use of models as primary artifacts in the software development

process, TD management does it by correctly dealing with problems that, out

of control, can impair system maintenance and evolution. Most research on TD focuses

on application code as primary TD sources. In an MDD project, however, dealing with

technical debt only on the source code may not be an adequate strategy because code

generation is often done at a later stage than creating models, and then dealing with TD

only in source code can lead to unnecessary interest payments due to unmanaged debts.

Recent works concluded that MDD project codes are not technical debt free, making it

necessary to investigate the possibility and bene ts of applying TD identi cation techniques

in earlier stages of the development process, such as in modeling phases. This

work intends to analyze whether it is possible to use source code technical debt detection

strategies to identify TD on code-generating models in the context of model-driven development

projects. For this, we investigated source code TD detection techniques were

evaluated with the purpose of adapting them to be used in the model abstraction level.

A catalog of nine di erent model technical debt items for platform-independent codegenerating

models was speci ed, with detection strategies that can automate the TD

identi cation process in the models. Then, the proposed detection strategies were implemented

in a tool that allows both speci cation and automatic identi cation of so-called

model smells in EMF models. A preliminary study was performed to evaluate and re ne

the proposed detection strategies by using 3 opensource projects versioned on Github.

Then, a deeper experimental study was performed in order to evaluate the detection strategies

real e ectiveness in anticipating the technical debt identi cation, detecting them

in the code-generating models. A total of 9 di erent projects, with 36 EMF models and

more than 400 thousand lines of code were used in this evaluation, all opensource, git

versioned and developed with EMF technology. Results showed that the evaluated detection

strategies are able to anticipate a great amount of source code technical debt in the

code-generating models. However, di erent catalog items presented di erent anticipation

e ectiveness levels and various levels of occurrences in the studied projects. A discussion

was performed for each evaluated item in order to explain the obtained results. Finally,

aspects were obtained that may guide future works in improvements and also extension

of the catalog.

System of systems (SoS) is composed of a set of systems, which interact together for

a common goal. Thereby, it tends to be larger and more complex than traditional systems.

To address the question of the inherent complexity of this kind of complex systems,

they are frequently modeled using the Business Process Modeling and Notation (BPMN),

which is a standard modeling notation for business processes. The choreography diagram,

introduced in version BPMN 2.0, provides suitable concepts for representing the interactions

among constituents of a SoS. However, models created using this notation may

contain errors, some of which can be detected at design-time and others only at runtime.

Syntax errors are easily detected with the assistance of modeling tools. Nevertheless,

the absence of a formal semantics for BPMN makes it harder to identify runtime errors,

as for example, deadlocks, livelocks and other safety properties in BPMN choreography

diagrams. These errors are challenging to detect and may lead to an improper operation

or even a system lockup. In this context, a method for identifying runtime errors consists

of translating the BPMN diagram into a formal model that can be analyzed in a model

checker. Thereby we present an approach to build a formal model for the BPMN choreography

diagrams in terms of the formal language ⇡-ADL, which is properly designed for

the specification of dynamic architectures, an intrinsic characteristic of SoS. Therefore,

we define the mapping of the elements of the choreography diagram to ⇡-ADL, in order

to obtain its formal description in ⇡-ADL. Such ⇡-ADL models allow its formal verification

using a specific model checker, enabling the prior detection of runtime errors in the

modeled system.

The mobile device market has grown exponentially in recent years, as the need for apps that strive for quality and that offer features that increase user retention. In recent years, a growing number of studies have provided solutions to problems inherent to the demand mentioned above. The software testing activity plays an essential role in the software quality assurance process. In particular, regression tests are presented as a viable strategy to deal with the complexity and constant evolution of applications since its main objective is to ensure that changes made between versions of a product do not change the system's behavior. Although the literature has devoted efforts to developing new regression testing techniques for Android - the most popular operating system for mobile devices - existing studies do not address how professionals use testing techniques. This investigation aims to carry out a thematic synthesis on the adoption of regression testing techniques in software projects for the Android platform. We organized the research in four steps: (i) carrying out a structured review of the technical literature on regression testing of Android applications, (ii) applying a survey, (iii) carrying out an identifier with professionals from the industry, and (iv) the construction of a thematic synthesis. The survey got 100 responses and provided preliminary insight into how the testing process is performed during and after application maintenance, how automated regression testing of Android applications is in practice, and why the professionals do not perform regression testing after updating applications. In the interview study, we inquired 16 industry professionals. This study contributed to the research as it could unveil the level of knowledge of professionals and its relationship with test automation. In addition, we could identify the most used languages in the development of Android applications and the tools used for testing automation. Furthermore, the study provided us with a preliminary understanding of how professionals perform regression testing in practice. Finally, the thematic synthesis presents a model resulting from the comparison of evidence gathered from different data sources - literature, survey and declares about the use of regression test techniques. These studies bring an insight into the subject and provide possible future investigations.

The registration and digital publication of texts in natural language is part of different domains of human activity. Among the information recorded, events are a significant part of the content, since, regardless of the domain, new events continue to happen, and those considered relevant are recorded. With the spread of information technologies, the ease of access to these records has increased the supply of registered events and also the demand for knowledge to aid decision making, derived from new applications of a large volume of information increasingly globalized and instantaneous. The task of event extraction aims to automate the extraction of events from text by identifying the event type and its participants and attributes. This is an information extraction task, which is part of the area of natural language processing. There is no known solution with a performance equal or superior to that of a human for this task in any natural language. In the specific case of the Portuguese language, there is not only a lack of a solution, but there is a need  for methods and linguistic resources for the development and evaluation of those methods (via learning from data, i.e. machine learning). To the best of our knowledge, the few works that were published did not perform event extraction in a closed domain configuration, when one knows the structure of the events that one aims to extract. In this research project proposal, we intend to develop a method for automatic event extraction in a closed domain configuration in Portuguese sentences and to build or adapt a corpus in Portuguese that supports the development of this method. This work presents a method for event extraction in a closed domain configuration in Portuguese language sentences based on deep learning using contextualized word embeddings. Furthermore, to support the development and evaluation of the proposed method, we construct a  Portuguese corpus with annotations of events in closed domain.

Residues Defined Network (RDN) have the ability to use labels for end-to-end routing in autonomous systems. RDN is characterized by dispensing with the use of routing tables at the core of the network. In this way, packets are forwarded based on the remainder of the division between the label found in the packet and the local key found in the core switch. Initially, RDN was intended to improve latency in SDNs, since the switches at the core of the network do not communicate with the controller. On the other hand, a range of solutions adopts the Quality of Service (QoS) based on the use of MPLS, which allows providing differentiated services to different traffic classes and patterns and generating considerable benefits for network communications. However, such solutions have high OPEX and CAPEX costs.Motivated by the capacity of RDNs, this dissertation brings an alternative method for using the labels in the core of the network, in addition to trying to solve the problem of scalability and flexibility found in the proposed KeyFlow solution and offer new functionalities for this technology.Through the P4 language, a new flexible, scalable and reprogrammable protocol for an RDN was implemented, in order to implement the model for provisioning QoS by programmable labels, ProgLab {Programmable Labels}.ProgLab implements a new method for reading labels by core switches, which uses the residual number system to define the output ports and queue priority, thus providing QoS to the RDN. To validate the proposal, emulation tests were performed with QoS and Quality of Experience (QoE) metrics to evaluate and demonstrate the feasibility of the model.The results demonstrated the viability of this proposal for the provision of QoS in core networks that use the RNS as packet forwarding, showing that in the future an RDN can become a low-cost alternative for packet forwarding with QoS services.

Online reviews are valuable sources of information to support the decision-making process, both for individuals and companies. Nevertheless, the large volume of reviews that have a low quality hinders the process of gathering helpful information from those reviews. Several retailers’ websites provide a voting system to allow customers to evaluate other product reviews as helpful or not. However, those votes are often biased and most of the reviews do not receive any votes at all. Besides that, several websites do not even have this voting mechanism or any other component for organizing the reviews in terms of their helpfulness. Therefore, classifying reviews according to their helpfulness has paramount importance in facilitating access to truly informative content. In this context, previous studies have unveiled several features and architectures that are beneficial for the perceived helpfulness prediction. In the present work, we argue that text specificity, defined as the level of details expressed in a text, can influence the perception of review helpfulness and, consequently, can also be a novel useful linguistic aspect for modeling the helpfulness prediction. We proposed two approaches to incorporate the specificity aspect into helpfulness classification models: i) using hand-crafted features based on text specificity and ii) using the review specificity prediction as an auxiliary task in a Multitask Learning (MTL) setting. First, we conducted an unsupervised domain adaptation approach [Ko, Durrett and Li 2019] to label text specificity scores on sentences from online reviews automatically. To evaluate the different trained models using this approach, we proposed a measure named Specificity Prediction Evaluation (SPE), which is based on the assumption that, on average, reliable specific sentences tend to be longer than reliable general sentences [Li and Nenkova 2015]. For the 18 collections of online reviews used in all of our experiments, we could achieve more reliable specificity predictions, according to SPE, by varying only the training set size and the number of training epochs. Finally, we performed experiments to assess the performance of the helpfulness classification models using two neural architectures: Convolutional Neural Network (CNN) [Kim 2014] and Bidirectional Encoder Representations from Transformers (BERT) [Devlin et al. 2019]. In summary, using balanced datasets, the perceived helpfulness classification models, embodied with text specificity- either as features or MTL - showed significantly higher precision results in comparison to a popular SVM baseline when using CNN. With BERT, the experiments showed that MTL outperformed the single-task models for most of the 18 datasets and both accuracy and precision were improved compared to the SVM baseline.

Context: The term Technical Debt (TD) refers to activities deferred during the development of the software project. TD can be compared to a financial debt whose debtor has the option of paying the principal amount of the debt with interest. Several initiatives have sought to identify the causes and effects associated with TD. Among them, the InsightTD project has carried out a survey in several countries to collect evidence on the state of TD practice. One of the results obtained by the project was the list of causes and effects of TD reported by software professionals. TD can be incurred at any stage of development, such as software testing. This activity aims to ensure the quality of the software being developed. However, both the project and the technical literature have not addressed the TD items that occur in software testing activities. Objective: To investigate, from the point of view of software testing professionals, the types of TD associated with test-related TD, the causes that lead to it and the effects of its existence. Method: A replication of the survey of the InsighTD project was carried out with software testing professionals in Brazil. In addition, responses from software professionals collected in the first execution of the survey were used. Thus, 54 responses from participants in the Brazilian software industry were collected and analyzed, quantitatively and qualitatively. Result: This study highlights the point of view of testing professionals on TD. It was noticed that test-related debt items are related to different types of TD, such as documentation, code, and the test debt itself. Sixty-one causes of TD were identified and organized into eight categories. The main causes identified were deadlines, outdated/incomplete documentation and inadequate planning. Forty-five TD effects were identified and grouped into six categories. The main effects are delivery delays, rework and poor external quality. Conclusion: The list of causes and effects along with their categories was organized into a set of cause and effect diagrams that can help software professionals in managing TD. 

Nowadays information is an essential resource in various sectors of the economy. The popularization of social networks, smartphone applications, and online services has increased the volume of data available online. Among this extensive amount of information, there is a specific data type called spatial data. It represents a physical object using its spatial coordinates (e.g. latitude and longitude). Spatial data is critical in a large number of application domains (e.g. land use, transportation plan). For instance, the user can find points of interest (POIs) or be warned of critical situations by spatial data applications like web search engines or emergency response applications. It’s been asserted that 80% of all data business has some locational reference. Spatial queries are widely employed to manipulate spatial data more efficiently. However, the user has a crucial role in the spatial information retrieval process when querying the needed information. For decades, researchers have proposed several techniques to aid users in expressing their information needs, such as Boolean models, pattern matching operators, and query expansion. Despite the existence of relevant alternatives in the field, there is still a lack of solutions applied to keyword preference queries. The Spatial Keyword Preference Query (SKPQ) arises as a potential solution to assist users in finding POIs. SKPQ selects POI based on the description of features in their neighborhood. In essence, the user defines a spatial (i.e. radius) and textual (i.e. query keywords) constraint to be satisfied. In this context, this thesis aims at proposing strategies to improve SKPQ results. The contribution is threefold. First, two Linked Open Data (LOD) repositories (i.e. DBpedia and LinkedGeoData) are exploited to improve the features description. The feature description in LOD contains more information than traditional spatial databases, leading to a more detailed description. Second, the query results are personalized to present the best POIs for the underlying user. By exploiting reviews on POIs, the system identifies the object that best satisfies the user and re-order the rank with respect to the user preference. Third, we model the user preference in visiting locations near to each other using a probabilistic function. This function is incorporated into the ranking function to retrieve POIs considering this user preference. We evaluate each technique employed in this proposal separately. The first technique achieves a relative NDCG improvement of  20% when using random query keywords. Also, it finds POIs where SKPQ is unable to find. The second technique further improves the relative NDCG by 92%. Finally, the third technique improves the rank consistency achieving a Tau performance of 52%.

The increasing number of modern applications, that rely on the processor capacity of multicore CPUs, has been pushing the area of real-time systems towards multiprocessor architecture. In fact, such applications are present in a wide variety of areas, such as automotive, avionics and personal electronics industries, most of them being composed of both time critical (hard) and non-critical (soft) services.
Indeed, for multiprocessor real-time systems that deal with both hard and soft tasks, one is interested in minimizing the response time of the soft tasks, while not jeopardizing timeliness of hard tasks. Several scheduling approaches have been proposed to deal with hard and soft tasks on a single processor, while for a multiprocessor environment such approaches are less common and are typically extensions of the mechanisms described for uniprocessor systems.

This work describes a flexible scheduling approach designed for multiprocessor real-time systems composed of both hard and soft tasks. According to the proposed approach, hard tasks are assigned to processors in an off-line manner and do not migrate during their execution. As for soft tasks, two main configurations are possible. They may be placed either in each processor's local queue, for which migration is not allowed, or in a global queue, being accessed by all processors. The scheduling approach described in this thesis is strictly linked to the concept of servers, which are scheduling proxies responsible for scheduling the tasks (or other servers) assigned to them. Two special servers are described, whose definition, rules and correctness proof are also presented in this thesis. Together, these servers are able to (a) encapsulate soft tasks so that their execution does not put at risk the timing requirements of hard tasks and to (b) reclaim system unused slack to favor the execution of soft tasks. By doing so, the average response time of soft tasks are improved, without jeopardizing the timeliness of hard tasks.

State-of-the-art techniques and Automated Static Analysis Tools (ASATs) for identifying code smells rely on metric-based assessment. However, most of these techniques have low accuracy. One possible reason is that source code elements, such as methods implemented according to different design decisions, are assessed through the same generic threshold for each metric. Other possible reason is that these metric thresholds are usually derived from classes driven by different design decisions. Using generic metric thresholds that do not consider the design context of each evaluated class can generate many false positives and false negatives for software developers. Our goal is to propose design-sensitive techniques to derive contextual metric thresholds. Our primary hypothesis is that using the design role played by each system class to define this context may point out more relevant code smells to software developers. We conducted some empirical studies to define the proposed techniques. Firstly, we performed a large-scale survey that showed that practitioners recognize difficulties in fitting ASATs into the software development process. They also claim that there is no routine for application. One possible reason practitioners recognize that most of these tools use a single metric threshold, which might not be adequate to evaluate all system classes. Secondly, we conducted an empirical study to investigate whether fine-grained design decisions also influence the distribution of software metrics and, therefore, should be considered to derive metric thresholds. Our findings show that the distribution of metrics is sensitive to the following design decisions: (i) design role of the class (ii) used libraries, (iii) coding style, (iv) exception handling, and (v) logging and debugging code mechanisms. We used these findings to propose two new techniques to derive design-sensitive metric thresholds using the class design role as context. Then, we carried out two large-scale empirical studies to evaluate them. The first study showed that our proposed techniques improved precision according to developers' perceptions. Since it is impossible and tiring to perform a complete source code quality assessment with developers, we conducted a second study mining the evolution of software projects from popular architectural domains. We found that our techniques improved recall to point out methods effectively refactored during software evolution.

Distance Learning, along with the evolution of Information and Communication Technology (ICT), has higher enrollment rates than traditional education. With this growth, educators started to manage classes with a more significant number of geographically distributed students. Challenges arise for teachers regarding how to monitor academic progress based on the fulfillment of planned educational objectives and assessing students' academic performance, supervising the acquisition of knowledge. Tracking students' learning experiences in Learning Management Systems (LMS) to assess academic performance and student progress is one of the most intense and exhausting challenges for educators due to the amount of educational data produced and the various learning indicators available on the LMS. Thus, this research aims to investigate how an approach to monitoring and evaluating the student's academic trajectory in the LMS, called SapeS, can positively impact the teaching and learning process in Distance Learning. The proposed solution aims to process educational data from the LMS, extract useful knowledge, and make information available to educators and students with techniques of Learning Analytics and Computational Ontologies, guided by Taxonomy of Educational Objectives. The solution was developed based on the Design Science Research (DSR) methodology and applied in a real context for evaluation with the population (n=24). Through the analysis carried out it was verified that there was a strong adherence to the use of the proposed tool and a significant difference between the data collected at different moments of the investigation. There was a greater engagement, motivation, and participation in the academic activities of the investigated course and greater attention from the teachers regarding the evaluation of the student's trajectory, in addition to checking and adapting the pedagogical planning in relation to the student's situation, and the planned educational objectives.

Image segmentation consists of assigning a label to each pixel in the image in such a way that pixels belonging to the same objects in the image must have the same labels. The segmented area of an object must span all pixels up to the limits (boundaries) with the other objects. The boundary region can provide helpful information for the segmentation process, as it marks a discontinuity that can define a segment limit. However, segmentation methods commonly suffer to explore boundary information and consequently to segment this region. This is so mainly due to the proximity between image regions containing different labels. In view of that, we propose to investigate how to take into account the boundary information when semantically segmenting an image object. Our first contribution is a graph-based image segmenter, called interactive dynamic programming (IDP)-expansion. This is a weakly-supervised method that requires a seed into each object targeted to be segmented in the image, subsequently minimizing an energy function to obtain the image labels. IDP-expansion explores dynamic programming to initialize an alpha expansion algorithm over superpixels to improve boundary information in a segmentation process. Over the Berkeley segmentation data set, our experiments showed that IDP-expansion is 51.2% faster than a traditional alpha-expansion based segmentation. Although IDP-expansion has shown to be faster, it suffers from two matters: A mandatory seed initialization and the lack of semantic information. This further led us to develop a supervised convolutional neural network architecture to semantically explore boundary information. Our novel method, called DS-FNet, uses two streams integrated in an  end-to-end convolutional network to combine segmentation and boundary information based on an attention-aware mechanism. To evaluate DS-FNet, we initially conducted experiments on general-purpose (Pascal Context) and traffic (Cityscapes, CamVid, and Mapillary Vistas) image data sets, having the mean intersection over union (mIoU) as the reference metric. DS-FNet outperformed ten segmentation networks in the Pascal Context, Cityscapes, and CamVid data sets. In the Mapillary Vistas data set, DS-FNet achieved second place when compared to five other methods. A second round of experiments was performed to evaluate the generalization of our proposed method on challenging medical imaging data sets, containing several kidney biopsy whole slide images (WSIs). The data sets used to evaluate the second version of our network were HubMAP, WSI Fiocruz, and a subset of Neptune data set, all considering glomerulus segmentation. After training DS-FNet only over HubMAP data set, containing periodic acid-Schiff (PAS)-stained WSIs with only non-injured glomeruli, we found that our network was capable to segment glomeruli on WSIs stained by other methods (hematoxylin-eosin (HE), periodic acid-methenamine silver (PAMS), trichrome (TRI), and silver (SIL)). The results of these latter experiments show that our model is more robust than other models based on U-Net architecture. All the experiments and analyses presented in this work demonstrated that the explicit and adequate consideration of boundary information improves the results over non-boundary segmentation methods.

The generation of planetary terrains, in the context of electronic games, presents several challenges. Among them, the main one is the management of multi-scale data. A player may be orbiting the planet in an instant and then descend to battle on its surface. To deal with this challenge, we present a method of procedural generation of planets with support for multiple scales, making use of Brownian fractal noise functions with analytical partial derivatives. Through several iterations of this fractal noise we can simulate different types of terrain. The multiscale terrain model is managed by an indexed quadtree with hash tables. This same structure supports the view-frustum culling process, which optimizes the rendering process. In addition, our approach is able to manage levels of detail based on the observer's point of view and the introduction of high resolution details through the use of the GPU tessellation process. The results show that our method is capable of generating realistic planetary bodies, with temporal coherence and diversity of geological compositions in real time, and, therefore, being applicable to the domain of electronic games.

Institutional Repositories (IR) are academic repositories that enable the storage and dissemination of scientific productions from universities and research centers. These repositories allow self-archiving, where the researcher himself can deposit his publication. The information about each deposited item is stored in its metadata, but as the researcher himself usually does this manually, the terms chosen do not always help in this description, which leads to the intervention of librarians in the process. Suggesting keywords during metadata validation would certainly help librarians as it would identify representative terms of each publication and semantically enrich these metadata, favoring the retrieval of items in an IR. Binary textual classification machine learning methods may suggest that a publication is also associated with another collection if it is identified that it is multidisciplinary work. You can also help classify publications not yet filed with IR that are not organized by collection but by other criteria, such as collegiate date or date of defense. Based on the representative terms of each community and sub-community, it is possible to train a multi-hierarchical classifier to identify which community and sub-community each job should belong to. On the other hand, semantic annotation of metadata referring to collections and keywords enables the enrichment of RI item descriptions and facilitates the retrieval process. In this context, the objective of this paper is to semantically annotate the items of an academic repository in Dspace semi-automatically using the Dublin Core RDF standard from the results obtained in the textual classification and validation of keyword suggestions. Additionally, a case study was conducted at the UFBA IR, where domain specialists validated the extracted keywords, in this case, the SIBI (UFBA Library System) librarians. Thus, the semantic annotation of the results obtained in the textual classification experiment and the case study were performed. The main contributions of this project relate to the exploratory study of publication deposit validation and classification methods, as well as the semantic description and enrichment method of an IR item.

The interest of the Information and Communication Technology community on sustainability has grown considerably in recent years. Although still at an early stage, the theme has become of great relevance since it forces us to think about what we have done to ensure the planet and future generations' continuity. In the Software Engineering context, when we think about sustainable software development, we face gaps, since this subject is still nebulous for software engineers and developers, as well as for the research community in Software Engineering. For software to be produced in a sustainable way, software engineers need to understand how sustainability concepts are incorporated into software development, so that they can have a clear, common and shared understanding of that knowledge. However, the development of a recent study on the state-of-the-art about software approaches that support sustainable Software Engineering showed that there is still a gap about what is sustainable software development, in fact. The lack of such an understanding can prevent the industry from building software with sustainable awareness. Given the question, this research aims to characterize sustainable software engineering by highlighting the sustainable concerns present in the software development life-cycle. To achieve this goal, this dissertation adopted a multi-method approach and produced a series of qualitative studies. The multi-method approach is a methodological research strategy that combines two or more qualitative research methods or two or more quantitative data collection and analysis methods. We elaborated: (1) a systematic mapping study with the intention of knowing the Sustainable Software Engineering domain; (2) a survey to obtain the software industry perception on the adoption of sustainable practices; and (3) a grounded theory, with the aim of generating a theory to provide an understanding of sustainable software development to readers. The theory was organized around the sustainability dimensions. This dissertation concluded that sustainable software development is explained through the following propositions: (a) technical, environmental and social concerns are present in all phases of sustainable software development. This means that researchers and software engineers are concerned with considering the longevity of the software produced, as well as environmental resources, in addition to social welfare; (b) the sustainable requirements identification must occur in the project initial phase (c) with the support of experts engaged in sustainability, who must be part of the group of stakeholders; and (d) the use of sustainable concerns can generate trade-offs in the project. Therefore, the results contribute to a greater understanding of sustainable software development, from the literature and software practitioners' perspectives; and, consequently, with the evolution of the state-of-the-art in Sustainable Software Engineering.

Most of the well-known and widely used conventional clustering algorithms, as k-Means and Fuzzy c-Means (FCM), were designed by assuming that, in most cases, the number of objects in a dataset will be greater than its number of dimensions (features). However, this assumption fails when a dataset consists of text documents or DNA microarrays, in which the number of dimensions is much bigger than the number of objects. Most studies have revealed that FCM and the fuzzy cluster validity indices (CVIs) perform poorly when they are used with high-dimensional data even when a similarity or dissimilarity measure suitable to this type of data is used. The problems faced by high dimensionality are known as the \textit{curse of dimensionality} and some approaches such as feature transformation, feature selection, feature weighting, and subspace clustering were defined to deal with thousands of dimensions. To be convinced that the number of dimensions should be maintained to learn as much as possible from an object and to know that just one subset of features might not be enough to all clusters, the soft subspace clustering technique was used in the proposed work. Besides FCM, three soft subspace algorithms, Simultaneous Clustering and Attribute Discrimination (SCAD), Maximum-entropy-regularized Weighted Fuzzy c-Means (EWFCM) and Enhanced Soft Subspace Clustering (ESSC) were performed to cluster three types of high-dimensional data (Gaussian mixture, text, microarray) and they were evaluated employing fuzzy CVIs instead of using external measures like Clustering Accuracy, Rand Index, Normalized Mutual Information, that use information from class labels, as usually done in most research studies. From the experimental results, in a general evaluation, all the clustering algorithms had similar performances highlighting that ESSC presented the best result and FCM was better than the remaining soft subspace algorithms. Besides the use of the soft subspace technique, in the search for the cause of the poor performance of the conventional techniques for high-dimensional data, it was investigated which distance measure or value of weighting fuzzy exponent ($m$) produced the best clustering result. Furthermore, the performance of nineteen fuzzy CVIs was evaluated by verifying if some tendencies and problems related to previous research studies are maintained when validating soft subspace clustering results. From the analysis made in this work, it was clear that the type of data was determinant to the performance of the clustering algorithms and fuzzy CVIs.

Unit testing is a specific type of test that deals with the smallest units in the system. It represents the first test after the implementation of a component. When the implementation of unit tests does not follow good practices, anti-standards are likely to be introduced in the code. Anti-patterns in tests, also known as test smells, are poor decisions for designing and implementing test code. Test smells impair the quality of the test code and reduce the ability of developers to interact with the test code, which makes it difficult to understand, read and, consequently, maintainability and evolution of the system. One strategy for removing test smells is to refactor the test code. Few studies in the literature offer automated support for the detection and refactoring of test smells. Thus, the present study provides a semi-automatic approach to identify test smells and suggest refactorings for the test code. Our approach is supported by the RAIDE tool (RefActorIng test Design Errors). We developed the RAIDE integrated with a development IDE to help test engineers refactor test code. RAIDE supports two test smells, Assertion Roulette and Duplicate Assert. Although there are other tools capable of identifying test smells, as far as we know, there is still no evidence of tools that provide automated support for refactoring test smells. In addition, no tool provides a user-friendly interface integrated with an IDE for the identification of test smells. To assess how our approach can help improve the quality of the test code, we also conducted an experimental study. We compared our tool with the state of the art and found that participants who used RAIDE were able to identify test smells more easily and quickly. In addition, RAIDE also proved to be efficient in refactoring test methods. Since RAIDE is a tool integrated with the Eclipse IDE, the identification and refactoring processes are faster and more intuitive compared to the state of the art and manual refactoring.

Data modeling in the temporal domain has been applied to different systems such as meteorology, engineering, medicine and economics. In this sense, the Fuzzy Time Series area has stood out due to its ability to approximate mathematical functions and linguistic variables to create rules, which are easier to be interpreted by specialists. As a consequence of this capacity, more accurate models can be obtained to understand the behavior of systems and, for example, to make predictions of future observations. Since the appearance of fuzzy time series models, researchers have been proposing improvements aiming, for example, to reduce errors in the forecasting task. In general, such tasks are composed of three stages: i) fuzzification; ii) fuzzy logical relationship; and iii) defuzzification. According to the literature, one of the points most studied by researchers is the fuzzification stage, with a focus on partitioning the universe of discourse for modeling fuzzy sets. However, few studies in the literature consider the separation of influences from the stochastic and deterministic components present in order to assist in the modeling process of fuzzy sets. In order to overcome this limitation, the present work presents two approaches that make improvements in the modeling of fuzzy sets in the fuzzification stage, considering influences of stochastic and deterministic components present in time series. In the first approach, the series was initially decomposed, separating its components into different monocomponents. Then, single components with high frequency were removed, resulting in a new smoothed series whose combination with the original series allows to generate a new scatter plot. The second approach was developed to model series with chaotic behavior.

In this sense, instead of modeling the series as a scatter plot, tools from the field of Dynamic Systems and Chaos Theory were used to reconstruct the series in phase space. Although using different methods, the two approaches remove the temporal dependence between the observations to assist in the modeling process of the fuzzy modeling sets, which is conducted considering the Fuzzy C-Means clustering method and cluster validation indexes, to estimate the quantity and function of fuzzy sets. In the experiments, time series with influences of stochastic and deterministic components were used to validate and test the generalization of the approach. The results achieved were promising and superior compared to the state of the art.

Currently, there is an expressive number of social networks used for different purposes, such as connecting people with a common interest in research, job offers, musical preferences, and general content.  These networks have gained significant popularity in recent years.  To demonstrate this phenomenon,  research shows that 71% of young American adults use a social network at least once a day.  With this frequent access and the freedom given by the networks, users started to publish numerous information, from personal photos to texts with opinions on different topics such as politics, entertainment, and health. In this sense, a new volume of information started to be produced, since, before social networks, only specialized professionals with access to conventional media were able to publish their opinions.  From a scientific point of view, several techniques have been proposed in the literature aiming at analyzing the content produced in such social networks. Specifically related to the users’ behavior, it is common to observe their modeling through graphs or time series,  however,  these methods tend to ignore aspects of this behavior, for example, the temporal relationship or the dependence between terms used in publications.  Considering these limitations, this research project was developed based on the hypothesis that the adoption of temporal graphs, together with tools from the areas of text Mining and Time Series, allows the detection of changes in the behavior of users of social networks.  To validate this hypothesis, a new approach was developed to identify points of change in users’ behavior and to associate them with real events that influenced public opinion.   This procedure uses  Text  Mining techniques to find terms,  which will be used later in the creation of temporal graphs, maintaining their relationships in the original texts and their temporal dependencies.  A new measure has also been developed, to quantify how users’ opinions evolve with time.  Finally, a method for automatic detection of behavior change is presented, which aims to identify points when changes occur. This approach was evaluated considering a historic event in Brazil:  the 2018 presidential elections.  This period was chosen due to the volume of publications that effectively established social networks as the main mechanism for political activism.   The results obtained emphasize the importance of the proposed approach and open new possibilities, for example, for the identification of bots that propagate fake news.

O escalonamento de frequências é uma técnica essencial para maximizar a eficiência dos recursos computacionais existentes, especialmente em sistemas com arquiteturas capazes de executar tais tarefas em núcleos de frequência variável. Um dos fatores mais críticos capazes de afetar a experiência do usuário em serviços como busca ou redes sociais é a latência de cauda, definida como sendo a latência no 95 ou 99-percentil. Esta latência pode ser fortemente influenciada pela resposta mais lenta de um núcleo de baixo desempenho de um sistema, com seu impacto largamente amplificado quão mais núcleos de baixo desempenho estejam hospedados neste mesmo sistema. Do lado corporativo, reduzir a latência de cauda para a meta desejada é tão necessário quanto a diminuição do gasto energético, haja vista que este é diretamente proporcional aos custos financeiros para a operação do serviço e consequentemente ao lucro. Trabalhos anteriores em escalonamento de frequências para núcleos com capacidade de variação de frequência são de granulometria grossa, no sentido que se observa o estado de toda a aplicação para a tomada de decisões (sem distinção entre aplicação, threads ou funções). Tais trabalhos também muitas vezes dependem de um processo externo rodando em paralelo para coletar o comportamento dinâmico de determinada tarefa.

Hydrographic Printing, also called Hydrographics, is a viable method for coloring objects created with 3D printers. However, executing the hydrographic technique leads to a complex interaction between a thin film and a 3D printed object. First, an image is printed on a film, which is put in the water. Then, by chemical reaction with an activator, the film becomes an adhesive sheet that carries the image pigment. When the 3D object dip, the film adheres to the object, projecting the 2D image to the object 3D surface. The hydrographic printing resultant projection is hard to predict precisely, because of the variable stretch in the film and the color whitening in the regions with high stretch. To address the difficulty of predicting the final result of hydrographic printing, we propose a computational solution that enables the generation of a flatted image of an arbitrary texture. Printing this image in a hydrographic film, it is possible to color a 3D object. The proposal allows executing a physical texture mapping in an object, with computational and physical steps. We propose a 3D computational simulation that uses Position-Based Dynamics, a popular technique for simulating deformable bodies and widely used in physics engines. We take advantage of this technique running in parallel a GPU-based simulation with suitable performance. We simulate the film behavior and its interaction with the 3D printed object, as an interaction between a soft-body colliding with a rigid one. To evaluate the achieved performance consistency, we made tests varying the number of vertices and voxels in the bodies involved and observed that the simulation kept running in real-time. We also execute the hydrographic technique in different printed models and compare these results with the simulated models.

The gravity fed water distribution network design (WDND) optimization problem consists in determining the pipe diameters of a water network such that hydraulic constraints are satisfied and the total cost is minimized. Traditionally, such design decisions are made on the basis of expert experience. When networks increase in size, however, rules of thumb will rarely lead to near optimal decisions.
Over the past thirty years, a large number of techniques have been developed to tackle the problem of optimally designing a water distribution network. This work tackles the NP-hard water distribution network design (WDND) optimization problem in a multi-period setting where time varying demand patterns occur. A new enhanced simulation-based iterated local search (ILS) metaheuristic is proposed which further explores the structure of the problem in an attempt to obtain high quality solutions. More specifically, four novelties are proposed: (a) a local search strategy to smartly dimension pipes in the shortest paths between the reservoirs and the nodes with highest demands; (b) a technique to speed up convergence based on an aggressive pipe diameter reduction scheme; (c) a novel concentrated perturbation mechanism to allow escaping from very restrained local optima solutions; and (d) a pool of solutions to achieve a good compromise between intensification and diversification. Computational experiments show that the proposed approach is able to improve over a state-of-the-art metaheuristic for most of the performed tests. Furthermore, it converges much faster to low cost solutions and demonstrates a more robust performance in that it obtains smaller deviations from the best known solutions.

Two-Dimensional bin packing problems (2BP) are classic combinatorial optimization problems belonging to the NP-Hard class and have applications in several sectors such as the textile, metallurgical and glass industries. 2BP consists in allocating a set of rectangular items in larger rectangular plates with a standardized size in order to minimize the waste of raw material. In this dissertation, we approach a restricted version of the 2BP with guillotine cuts presented in the ROADEF/EURO Challenge 2018, in which there is a possibility to rotate items in 90° and soma itens have a order of precedence to be produced, and the rectangular plates may present defects in certain points. We propose two randomized greedy heuristics and a method for improving solutions based on a constraint programming model combined with a random greedy heuristic. The techniques are combined in Multistart and Greedy Randomized Adaptative Search Procedure (GRASP) metaheuristics in order to obtain better solutions. Computational experiments show that the use of the solution improvement method is advantageous, and that some combinations of the proposed techniques are more effective for certain types of instances. A preliminary version of this work was qualified for the final phase of the ROADEF/EURO Challenge 2018.

With the growth in the demand for traffic on computer networks, Elastic Optical Networks emerged as a promising technological evolution of optical networks for the development of solutions focused on the efficient use of network resources. Unlike traditional WDM networks, which work with allocation of fixed grid resources, in EONs the bandwidth of an optical path is scalable and can vary according to the need for each connection, adapting dynamically to each traffic demand, ensuring the most efficient use of spectral resources. In this context, the use of resources is becoming more and more balanced, enabling greater cost reduction in several technological factors. However, with the popularization of the internet and the search to serve a greater range of technological possibilities with efficiency, researchers have been focusing their efforts towards an even more strategic use of these network resources. Translucent Elastic Optical Networks, through efficient strategies, strongly use the concept of elasticity and adaptability with the use of signal regeneration equipment. The Virtualized Elastic Regenerators are devices capable of adapting dynamically according to the traffic demand, where as they are no longer needed, VER is released to use new demands. In this way it is possible to serve a greater number of optical connections and reach greater distances with energy efficiency and cost reduction. These new transmission technologies have complementary characteristics, which using them combined with other concepts, brings great benefits to companies and users. The establishment of term-oriented connections can be postponed until such time as resources are available to meet the required term. The possibility of scheduling a requisition can bring benefits to both users and service providers. If there are no resources available, the establishment of a new connection may be postponed, so it is possible to handle a greater number of requests, avoiding an increase in the blocking rate. In conjunction with the scheduled provisioning of requisitions, the creation of requisition batches contributes to flexibility when establishing new connections. A batch is a set of requests that arrive at a certain time and that are not immediately provisioned, that is, each request will receive a certain deadline where your establishment must be attended to within this period. If there are no resources, at the end of the deadline the connection must be blocked. Scheduling in batches allows for greater combinations, making the most of the available network resources, thus avoiding increased energy consumption and reducing the number of blocking requests. Therefore, this work presents a new methodology to face the problem of exponential increase in traffic on the communication network, considering cost reduction naturally a priority. In order to achieve these expectations, the proposed methodology consists of the development of a dynamic pooling and resource sharing model for regeneration equipment, in addition to the combination of technologies in the creation of requisition batches with scheduled provisioning. The results obtained through simulations by software implemented by the author himself, indicate that the model proved to be effective, and was able to considerably reduce the blocking rate, the amount of resources with VER used and consequently generated greater energy efficiency, when compared to them methods without using this grouping methodology, creating requisition batches and scheduled provisioning.

Source code metrics quantify dierent software attributes and have the potential to support
the identication of design problems that may aect software comprehensibility and
maintainability. Identifying design problems can reveal parts of the source code that
need to be monitored more closely. However, one of the major challenges in using metrics
in source code quality monitoring activities is the denition of threshold values that are
capable of identifying design problems that are actually considered problems according
to developers' perceptions. Although there are a number of techniques for extracting
threshold values, the threshold values obtained by means of them generate many false
positives. That is, there are many of code elements (e.g., classes or methods) whose
metric values exceed threshold values that, when evaluated by developers, are not considered
to be problematic. Therefore, the participation of developers is important to
analyze threshold value eectiveness. Few studies evaluate the eectiveness of dierent
threshold extraction techniques based on the perception of developers. Therefore, the
goal of this work was to conduct an experimental study to evaluate the perception of
developers about design problems detected with threshold values obtained by means of
ve dierent techniques that extract threshold values from system benchmarks. In this
scenario, we conducted two studies, a preliminary one in which we analyzed two systems
and two developers' perception, and a second study in which we analyzed four systems
and eight developers. The results indicate that developers use design roles to assess
whether or not design problems exist, and techniques that take into account some contextual
information, such as reference systems or design papers to derive thresholds may
reduce the number of false-positives.

Contexto: O conceito de dívida técnica (DT) contextualiza o problema das tarefas de desenvolvimento pendentes como um tipo de dívida que traz um benefício a curto prazo para o projeto, mas que poderão ter de ser pagas com juros mais tarde no processo de desenvolvimento. É comum que um projeto de software incorra em dívida durante o processo de desenvolvimento, no entanto, sua presença traz riscos para o projeto e dificulta sua gestão. Diferentes estratégias de gerenciamento da DT têm sido propostas, contudo, considerar ações que possam evitar sua inserção e monitorar seus efeitos ainda não é uma prática comum. Este é um ponto que merece ser melhor investigado por diferentes motivos: (i) conhecer as causas da DT pode auxiliar as equipes de desenvolvimento na definição de ações que possam ser tomadas para evitar sua ocorrência e (ii) conhecer os possíveis os efeitos da DT pode apoiar na realização de análises de impacto mais precisas e também na definição de ações corretivas para minimizar possíveis consequências negativas da dívida inserida. Dessa forma, enquanto o gerenciamento da DT é um tópico de pesquisa importante, também é necessário entender as causas que podem levar equipes de desenvolvimento a incorrer em diferentes tipos de dívida, bem como os efeitos de sua presença em projetos de software.
Objetivo: O objetivo desta tese é investigar, através da replicação contínua e independente de surveys distribuídos globalmente, o estado da prática e tendências da indústria sobre DT incluindo causas que levam à sua ocorrência, efeitos de sua existência e como esses problemas se manifestam no processo de desenvolvimento de software. A partir da organização das causas e efeitos identificados, estruturar diagramas de causa e efeito probabilísticos que possam ser utilizados no apoio a atividades de gestão da DT.
Método: As atividades realizadas nesta tese são fundamentadas no paradigma da engenharia de software experimental. Inicialmente, foi realizado um estudo terciário com o objetivo de investigar o estado atual da pesquisa sobre DT identificando quais tópicos de pesquisa têm sido considerados, organizando direcionamentos de pesquisa e conhecimentos práticos que já foram definidos, identificando os tipos conhecidos de DT, e mapeando quais atividades, estratégias e ferramentas têm sido propostas para apoiar o gerenciamento da DT. Em seguida, foi planejado InsighTD, uma família de surveys globalmente distribuída. InsighTD foi planejado de forma cooperativa com pesquisadores da área de DT de diferentes instituições ao redor do mundo. Trata-se do primeiro estudo em larga escala realizado na área. Ele busca organizar um conjunto aberto e generalizável de dados experimentais sobre causas e efeitos da DT em projetos de software. Por fim, os diagramas probabilísticos de causa e efeito da DT propostos foram avaliados através de estudo de caso executado na academia.
Resultados: Esta tese apresenta os resultados do estudo terciário executado e as análises realizadas sobre a primeira execução de InsighTD no Brasil e sua primeira replicação nos Estados Unidos. Ao total, 107 profissionais da indústria brasileira e 99 da indústria estadunidense de software responderam ao questionário. Os resultados indicam que há uma ampla familiaridade com o conceito de DT. Prazos, planejamento inadequado, falta de conhecimento e falta de um processo bem definido estão entre as 10 causas mais citadas e mais prováveis de levar à ocorrência da DT em projetos de software. Por outro lado, baixa qualidade, atraso na entrega, baixa manutenibilidade, retrabalho e perdas financeiras estão entre os 10 efeitos mais citados e de maior impacto em um projeto com a presença da DT. Diagramas probabilísticos de causa e efeito da DT para os diferentes tipos de DT foram elaborados. Os resultados também indicaram que o tipo de modelo de processo (ágil, híbrido ou tradicional) impacta nos efeitos da DT sentidos por equipes de desenvolvimento.
Conclusão: Com InsighTD, pretende-se reduzir o problema de investigações isoladas em DT que ainda não são representativas e, assim, construir um conjunto aberto e generalizável de dados experimentais para a compreensão de problemas práticos e desafios da área. Parte do conhecimento organizado será estruturado em diagramas probabilísticos de causa e efeito que permitem apoiar atividades de gestão da DT.

The Dynamic Software Product Lines (DSPL) engineering processes aim to design Dynamically Adaptable Software (DAS) by increasing the flexibility for the generation of a huge number of configurations. It results in a software configuration space explosion making the analysis more difficult and complicating the developer’s work. In this sense, software engineers need to find a combination of systems features that can simultaneously satisfy constraints specified in feature and context models, Non-functional Requirements (NFRs), and stakeholder’s preferences. It means that they have to measure many configurations until finding the optimal ones, characterizing the product configuration process in a complex optimization problem. Most of the existing studies do not focus on the interactions between contextual information and NFRs when dealing with feature selection to meet the desired quality objectives in DAS. In addition, such studies do not use any planning strategy to support the configuration selection process. Based on these research gaps, we propose an approach that (i) manages the system’s features and contexts; (ii) facilitates the understanding of how DSPL applications can behave from a certain context change, and (iii) enables to conduct trade-off analysis in order to find valid and optimal configurations, which meet the constraints and the interactions between contextual information and NFRs. Aiming to support the context variability modeling of DAS, we proposed the Extended Context-aware Feature Modeling (eCFM) technique to deal with constraints among contexts. Next, we defined the DAS Trade-off Analysis (ToffA-DAS) approach to deal with the configuration selection process embracing interactions between contextual information and NFRs. We also proposed a strategy to analyze context changes in order to define adaptation models for each prioritization of the system’s features, contexts, and NFRs. Finally, we evolved our approach and named it as DAS Trade-off Analysis PLUS (ToffA-DAS+). ToffA-DAS is based on the integer linear programming technique, whereas ToffA-DAS+ uses a genetic algorithm. We performed a set of empirical studies in order to evaluate the proposal for this thesis. First, we conducted a survey to evaluate eCFM from the viewpoint of expressiveness to model the context constraints and easiness of use. Indeed, the analysis was focused on the comprehensibility of contextual variability modeling. Next, we performed a study based on simulations to gather initial evidence about the feasibility of using ToffA-DAS. It is based on how to conduct trade-off analysis and define adaptation models from optimal configurations found in the analysis. We also conducted an exploratory study to evaluate how the configurations obtained by the execution of ToffA-DAS affect the overall satisfaction level of NFRs. Finally, we evaluated the evolution of our approach in comparison with the previous release. As a result of the first study, the eCFM was considered a technique with a great expressiveness to represent adaptation rules among contexts and system features, besides the easiness of use and organization with the grouping of contexts. Therefore, we argue that the software engineers may take into account the use of eCFM technique to model DAS. In the second study, ToffA-DAS presented consistent results in accordance with the real-world scenarios and satisfied the estimated utility values and model constraints. The third study showed that the set of configurations generated by ToffA-DAS execution provide high satisfaction levels of NFRs. In the last study, we collected evidence that ToffA-DAS+ suggests more solutions from then possible valid configurations of the model. Based on the aforementioned studies, we evidenced that our approach can be handy when software engineers need assistance in the understanding of how to design a variety of configurable options for DSPL applications. It is based on the principle that each configuration option must be optimal to meet certain contextual changes without losing service quality. With the usage of our approach, software engineers can exhaustively analyze and simulate a solution before implementing it.

Interoperability is the ability of heterogeneous systems to communicate transparently and

can be achieved through levels, such as: (i) syntactic, (ii) semantic, (iii) pragmatic, and

(iv) organizational. However, there are several challenges for systems to provide interoperability

such as what requirements are necessary to achieve the desire interoperability.

Several proposals have been made to discuss aspects of a specific level of interoperability

or for a given domain. However, the interoperability of a system is an aspect that can

evolve throughout its life cycle, so support is needed to aid this process. Maturity models

can help in this case as they have been used in several domains to assess the maturity of

the system according to specific aspects (e.g, interoperability). This work presents AMortisse

(mAturity Model fOR inTeroperability In Software SystEms), a maturity model to

check interoperability in software systems towards the specification of a methodology for

maturity model definition. This methodology aims to aid the Maturity Model (MM) developers

by presenting a MM development life-cycle applicable for different domains. We aim

to systematize tasks involved in MM development, such as MM domain requirements,

organization of related concepts into levels, dimensions, and the path to maturity, therefore

leveraging the produced MM quality. Despite their popularity, maturity models

have been criticized due to lack of empirical validation and effective methods to aid in

their definition. A validation was performed by applying it to an organization's system to

attest Amortisse's ability to indicate the system's maturity. As a result, Amortisse was

able to measure the level of interoperability of the system and show which requirements

must be met to evolve between levels of maturity. The results of this investigation show

that the Amortisse and the methodology are feasible. We hope the presented methodology

provides clarity while obtaining this model and may help the development of new

maturity models in different domains. We hope that Amortisse will be able to show

systems interoperability level and indicate missing requirements.

The Internet of Things (IoT) has been playing an important role in the technology scenario due to its high potential and impact on different society segments. Estimates suggest a trend for an increase in the number of IoT devices connected to the Internet for the next few years. As a consequence, the volume of data produced by IoT devices will follow this growth perspective, and there will be a demand for systems that are able to process, store and promote access to large amounts of data. In the typical IoT system, the data collected from sensors is stored and processed in cloud servers; however, some IoT solutions use edge devices to perform actions, such as processing, storage and access, using only local infrastructure for low latency requirements. Fog Computing has been used to improve IoT solutions with the aim to transfer some of the complexity from the cloud to the edge of the network, i.e., closer to devices, applications and/or users, working as a kind of ``local and private cloud''. The cooperation of devices and applications between edge and cloud creates a need for an interplay to allow data flow among the layers of IoT systems deployed in the edge and the cloud. Thus, it is necessary to support the data life cycle since its collection, analysis and use. Performance efficiency is a quality factor of systems and software engineering, which measures ``performance relative to the amount of resources used under stated conditions''. In special, in IoT systems which involve a large volume of data, the performance efficiency of data interplay is a relevant requirement. This thesis proposes a data interplay model of Internet of Things to provide the definition and deployment of the IoT data life cycle, in collection, analytics and data use stages. This data interplay proposal aims to improve performance efficiency in the operations regarding the IoT data life cycle: collection, analytics and use among devices and applications in edge and cloud infrastructures.

Hours of video are uploaded to streaming platforms every minute, with recommender systems suggesting popular and relevant videos that can help users save time in the searching process. Video summarizers have been developed to detect the video's most relevant parts, automatically condensing them into a shorter video. Currently, the evaluation of this type of method is challenging since the metrics do not assess user annotations' subjective criteria, such as conciseness. To address the conciseness criterion, we propose a novel metric to evaluate video summarizers at multiple compression rates. Our metric, called Compression Level of USer Annotation (CLUSA), assesses the video summarizers' performance by matching the predicted relevance scores directly. To do so, CLUSA generates video summaries by gradually discarding video segments from the relevance scores annotated by users. After grouping the generated video summaries by the compression rates, CLUSA matches them to the predicted relevance scores. To preserve relevant information in concise video summaries, CLUSA weighs the video summarizers' performance in each compression range to compute an overall performance score. As CLUSA weighs all compression ranges even that user annotations do not span some compression rates, the baseline changes with each video summarization data set. Hence, the interpretation of the video summarizers' performance score is not as straightforward as other metrics.

In 2006 the Brazilian Computer Society (Sociedade Brasileira de Computação - SBC) proposed five major challenges in computer research in Brazil. The fourth challenge, "Participatory and universal access of the Brazilian citizen to knowledge", involved Human-Computer Interaction (HCI). The objective of this challenge was to overcome the barriers that impeded access and interaction by designing systems, tools, models, methods, procedures and theories capable of giving Brazilian citizens access to knowledge. In addition, interactive technologies are penetrating into every aspect of daily life. In this context, interaction design raises questions about how to design systems that are suitable for all users. This research aimed to understand how to adapt an interaction codesign process for people who are deaf or hard of hearing (D/HH) participate. The starting point was the analysis of a participatory codesign process, SPIDe, with the objective of identifying, adapting and evaluating design solutions for the participation of D/HH in the entire process. A case study was conducted at a state college for deaf people with a group of six year old students to test the adapted process. In this case study, it was created an educational game by codesign with the children. As result, the children produced two video games, created by e for D/HH childrens, whose goals were to help deaf children’s alphabetization, teaching numbers systems and the alphabet. In addition, the children recognized themselves as producers and the games are imbued with elements of their culture. From this results, guidelines and indications were defined on how to conduct a codesign process with D/HH children and also about possible extensions of this process.

On the one hand, the simplicity of the HTML document standard was an initial design requirement, with the goal of rapidly expanding the Web by creating various HTML-authoring tools. On the other hand, this same simplicity also generates limitations for the use of the Web by machines due to the lack of structuring and semantics of HTML documents. In this context, the Semantic Web aims at extending the current Web, or the Web of Documents, such that the Web has become susceptible to interpretation and use by machines (e.g. software agents). To that end, the Web of Data attempts to describe the current Web’s data through structured documents that contain metadata and explicit semantics. In addition, rules for authoring and publication of structured and explicitly semantic data called Linked Data have been proposed for the Web of Data. However, the proposed document patterns in the Linked Data seems to undermine the initial goal of the HTML standard, i.e. simplicity. Thus, the aforementioned authoring process must now take into account that such simplicity is no longer a valid prerequisite for Web documents. Therefore, to generate and publish data for the Wed of Data just as it was for the Web of Documents, some studies have proposed methods and tools for the following: automatic generation of Linked Data from databases; semiautomatic extraction of Linked Data from HTML pages; and authored Linked Data for ordinary users. However, in relation to authoring and publication by ordinary users, these methods and tools are still being used in an ad hoc manner. That is, the previously mentioned studies present solutions that are domain specific such as social networks, content managers, wikis, blogs, etc. In this context, this work proposes a framework for data authoring and publication in the Web of Data according to the rules of the Linked Data standard while still focusing on ordinary users.

Attribute selection is a widely used dimensionality reduction technique to deal with the difficulties associated with the "dimensionality curse" in text classification tasks. The most common attribute selection approach for textual databases is to weigh the relevance of each attribute to the learning process and to select the best-valued N's, where N is generally an empirically defined number. Although this strategy is widely applied, it can lead to the partial or complete exclusion of attributes essential for learning In this sense, the research presented in this paper aims to foster the generation of more reliable text classifiers through the improved Xangô, a framework for selecting attributes in tasks of classifying texts, whose selection process seeks to construct a reduced dimensional space where all classes are represented in a balanced way by the its most discriminatory terms. Experimental results indicate that Xangô is a framework that is adaptable to different state-of-art attribute selection methods, surpassing its individual performances in learning tasks performed under varied conditions, including multi-class problems, imbalance between classes, different classification algorithms and drastic dimensionality reductions.

In the last decades, the need to store and process large volumes of data, generated daily under the most varied sources and commonly referred to as big data, has grown significantly. The importance of these data increases as they are integrated to enable the execution of studies and analyzes, to support decision-making processes and to generate comprehensive information about a particular domain. The integration and processing of large volumes of data impose considerable computational complexity. The analysis of these data, for example, requires machine learning and visualization techniques to allow the correct identification of the information present in these data, not always easily perceived. In public health, specifically in the context of malaria, there is a worldwide effort to eradicate the disease, since it still has high incidence rates. This research is part of a project that deals with integration techniques and visual data mining applied in a malaria surveillance ecosystem in the construction of an integrated database comprising notifications, deaths, vector control and climatic data. This database is accessed through Malaria-VisAnalytics, a mining platform for descriptive, predictive and visual analysis, assisting public policy decision-making by government and health stakeholders. Our experimental and validation results favor interaction and visual exploration allowing effective surveillance. Finally, our study can be easily extended with new features and data sources to accommodate more complex scenarios.

Ontologies play an important role across several domains as they represent and define categories, properties and relationships among concepts, data and entities existing in those domains. Specially in Robotics, ontologies can be used as a standard way to represent and share knowledge and reasoning among autonomous agents. CORA (core ontology for robotics and automation) is a standard ontology developed by the IEEE ORA working group mapping main concepts and axioms from the robotics and automation (R&A) domain. After CORA has been approved as an official IEEE standard, ORA was split in different sub-groups addressing, among other topics, task representation (RTR) and autonomous robots architectures (AuR). This work presents a simulation study on a multi-robot reconnaissance application intended to serve as proof-of-concept on the feasibility of the CORA ontology and its related architecture. We show how robots and tasks are represented and dynamically mapped into the ontology, as well discuss some technical aspects of our simulation when dealing with the proposed architecture. Our results show the proposed ontology (CORA) and its associated reference architecture (ROA) are able to be correctly mapped during application runtime.

In Mobile Crowdsourcing, tasks are distributed to workers according to sensor data, aiming to solve problems using collective intelligence. Requesters create tasks and distribute to the crowd workers in order to solve collectively a certain issue. Matching a task to the appropriate worker is an open issue in the area. This task distribution influences the quality of the workers’ response or the task acceptance rate. In this research we investigate the usage of context on mobile crowdsourcing, and propose ConTask, a context-aware approach to support task distribution. ConTask comprises a context model and a mobile app, which extends the task model proposed by Mrazovic and Matskin (2015) implementing the five context dimensions proposed by Zimmermann et al (2007): Individuality, Time, Relation, Activity, and Location. Additionally, ConTask presents a task distribution architecture, instantiated in the ConTask app prototype, which aims to support task matching to crowd workers according to the task's contextual requirements as defined by the requesters. The ConTask App was implemented according to the proposed model and achitecture, using popular sensors for each context dimension. Experiments were performed with different users in a University, and provided some insights regarding sensors usage; the accuracy of the implemented task distribution was 63%, with 73% precision and 63% revocation. To evaluate the ConTask model, we instantiated an performed an investigation with users in the University Administration domain. This experiment aimed to get insights about how the context dimensions influence tasks' acceptance from the crowd workers point of view. The results illustrated some tendencies in the influence of context usage in the worker willingness to perform the task; such as the higher the distance between the worker and the task location the lower the willingness of the worker on performing the task; the activity the worker is performing when receive the task request influences its acceptance, such as the activities of "waiting" or "talking" in contrast to "on cellphone" and "working"; the time of the day the worker received the task, such as morning versus night; the worker expertise regarding the task requirements; and the social relation between the worker and the task requester.

In software development that uses some agile method, the source code, rather than documentation, is the main and most up-to-date artifact. In contrast to this reality, the Model Driven Development (DDM) approach has the model as the main artifact of the development process. Intending to get th advantages of both approaches, (SALES, 2017) developed the ScrumDDM meta-process (agile MDD) that added modeling and model transformation practices coming from model driven development to the Scrum framework. Unlike the work of (SALES, 2017), this paper will investigate whether the ScrumDDM metaprocess preserves the characteristic agility of SCRUM, supports software evolution through user stories and documentation provided by DDM using for that aexisting software and developed by ScrumDDM, as well as whether this metaprocess is effective in creating new processes that integrate SCRUM and DDM. To evaluate the metaprocess regarding its ability to evolve user stories and the agility in software development, a controlled experiment was developed. To evaluate the generalization of metaprocess, in turn, a new software process from the academic literature was instantiated from this metaprocess. Through the controlled experiment developed in this work, it was possible to demonstrate that the evaluated metaprocess supported the software evolution through DDM documentation. It has also been shown that development agility has been enhanced through the transformations and models of the DDM approach. In addition to the results cited, at the end of the development process, it was observed that, both the project code and documentation, were up to date.

In this thesis, we propose new matheuristics for two NP-hard graph optimization problems. The first studied problem is denoted the minimum weighted feedback vertex set problem (MWFVS), for which given a weighted graph $G=(V,E)$ it consists in obtaining a minimum weight subset $F\subseteq V$ of the vertex set whose removal makes the graph acyclic. Differently from other approaches in the literature, we tackle this problem via the maximum weighted induced forest problem (MWIF). First, we propose two new compact mixed integer programming (MIP) formulations, using a polynomial number of variables and constraints. Next, we develop a matheuristic that hybridizes a multi-start iterated local search metaheuristic with a MIP-based local search procedure. This thesis also studies the $b$-coloring problem, for which given a graph $G=(V,E)$ it consists in attributing a color to every vertex in $V$ such that adjacent vertices receive different colors, every color has a $b$-vertex, and the number of colors is maximized. A $b$-vertex is a vertex adjacent to vertices colored with all used colors but his own. The optimal solution of the $b$-coloring problem determines the $b$-chromatic number of $G$, denoted $\rchi_b$. We present an integer programming formulation and a very effective multi-greedy randomized heuristic which can be used in a multi-start fashion. Furthermore, a matheuristic is proposed combining the heuristic with the integer programming formulation. Extensive computational experiments show that the proposed techniques for both problems outperform the state-of-art metaheuristics for the majority of tested instances.

Software metrics have been widely adopted during the process of developing source code, allowing to manage projects with higher quality by quantifying different characteristics that may impact, for example, the project costs and the planned activities. In general, such metrics can be divided into direct and indirect. The direct ones are commonly used in software development and maintenance. Those measures are directly extracted from source codes such as, for instance, the number of written lines and the total number of errors collected in a given time instant. The indirect measures, on the other hand, are obtained after performing some interpretation of the project. They may also include direct metrics and were developed to be generally applied to different software projects. However, the interpretation of these metrics depends on the experience of the project manager (domain specialist) and must consider specific characteristics related to the team of developers. In this project, we present a new approach, referred to as Fuzzy Software Analyzer -- FSA, designed to automatically extract characteristics and patterns in software repositories, aiming at: i) assisting the expert during the task of interpreting metrics, especially by working on large volumes of source codes; and ii) monitoring the evolution of the software as new releases are submitted to the repositories. The proposed approach was assessed by analyzing the Linux Test Project repository, demonstrating its usefulness and applicability.

The growth of the Internet of Things has allowed many devices to be daily connected. Each device has its own action rules and its peculiarities that can be monitored by a domain expert. Many of these devices are also configured to work in conjunction with other devices in controlled environments. With the advancement of the Internet of Things, it becomes more difficult to manage when devices are connected and/or disconnected from the environment. Moreover, with the presence of a new device, such as restricted usage rules for a controlled environment, may be violated and cause unexpected behavior in the environment. These cases are related to feature interaction. Feature Interaction occur when two or more devices/services generate instability in the environment due to input or output of a device. In dynamic environments such as IoT, device inclusion and exclusion makes it impossible for an expert to maintain these rules. According to the systematic mapping review performed, most works use the rules applied by the expert, requiring a method that allows to detect these feature interactions automatically. In this sense, the main objective of the present work is to detect features interacionista based on data analysis. Feature interaction data sets were annotated to obtain a model for automatically identifying rules. Experiments were performed and the results showed evidence of automatic detection of resource interactions in similar or complementary domains.

Large amounts of heterogeneous data produced by social networks, Internet-connected devices, and web applications has been stored and managed in different formats at distinct service levels Data-as-a-Service (DaaS) and Database-as-a-Service (DBaaS). This solutions provides on demand services via Application Programming Interface (API). To access these distributed cloud services, users in most cases face data ambiguity issues. To solve these problems and to avoid performing complex and laborious tasks to access this data, an automated solution that serves as an intermediate layer for establishing communication between SaaS and DaaS / DBaaS service levels is required. This solution allows SaaS service level consumers to access their data stored at different DaaS service levels through a single query. The middleware layer called Middleware for Interoperability Between SaaS and DaaS (MIDAS) provides this solution to users in a transparent way. Over time, DaaS service level parameter updates can semantically affect query terms (that are automatically or not automatically) configured for applications at SaaS service levels. To ensure that cloud consumers continue to access DaaS data, an approach that ensures semantic similarity between DaaS parameters to maintain the reliability of the original request is recommended. Our work proposes a method (SM) with two similarity ways: (i) edge counting (Cosine and Jaccard) to measure the similarity between two attributes and the Information Content(IC) to measure similarity based on knowledge through a WordNet corpus. The IC method is used in the model to fill the limitations of the edge counting method. To choose the edge count methods (Cosine and Jaccard), an environment with eight distance measurement algorithms, twenty-two authentic parameters from eleven DaaS providers, and five possible parameter change situations was simulated. As a proof of concept, our model is implemented in MIDAS to evaluate three criteria: overload, performance and correctness. The results of our experiments showed that we are in the first direction to provide semantic interoperability between SaaS and DaaS in MIDAS.

Data mining techniques allow extracting relationships and features present in data that are not easily noticeable, especially in scenarios involving large volumes of data. These techniques allow grouping and classify data, identify patterns and exceptions, as well as establish association relationships between different data. THE Predictive analytics is a sub-area of data mining that uses machine learning techniques to implement of predictive analytical models. Such models are based on prior knowledge of the data (descriptive analysis) to design possible future events about the context being analyzed. This work is an integral part of a research project which aims to integrate malaria reporting data from different sources of information and develop models of predictive analysis on these data. Specifically, this work involves the study of different prediction techniques based on in models and data and, based on this study, the development of predictive models for the epidemiological scenario of malaria in Brazil. Four case studies involving socioeconomic and epidemiological data are discussed as evidence of concept of the proposed models. The results show that the models developed have an accuracy variable between 60% and 85% for the tested scenarios, with prediction windows of up to 3 weeks and considering the large variation observed in the malaria cases in the last four years

Software Product Line Engineering (SPLE) has been widely adopted for applying systematic reuse in families of systems. Given the high upfront investment required for SPLE adoption, organizations commonly start with more opportunistic reuse approaches (e.g., a single system that they clone and modify). However, maintenance problems appear when managing a large number of similar systems where each of them implements and evolves particular characteristics. One viable solution to solve this issue is to migrate to SPLs using an extractive approach. This initiative, in its early phases, includes the definition of a Product Line Architecture (PLA) to support the derivation of product variants and also to allow customization according to customers' needs. In this way, the systematic use of Software Architecture Recovery (SAR) techniques enables PLA recovery and keeps the PLA aligned with development. Our objective is to provide an automatic approach to recover PLAs and guidelines to support the PLA recovery. We gathered knowledge by means of literature reviews and exploratory studies to characterize the state-of-the-art and identify research gaps on SAR techniques and tools that support the recovery of architectural variability information from source code for a family of products. The use of SAR techniques and tools to recover a PLA that documents variability information at the architecure level may address issues related to SPL adoption, design and evolution. Unfortunately, few studies investigate PLA recovery and also provide empirical evaluation. One of the main issues in the extractive approach is the explosion of the variability in the PLA representation. Our approach is based on identifying variability on architectural level by extracting information from variants' source code. To evaluate our approach, we performed a set of empirical studies.

Data clustering refers to an unsupervised machine learning method, often used when no labeled data is available. This class of machine learning models is used to retrieve information by splitting the data into coherent and contrasted groups. The most popular partitional and hierarchical clustering algorithms are K-means and AGNES, respectively. Nevertheless, to make either K-means and AGNES suitable to categorical data, some extensions are required in terms of similarity measures to be used and central position measures to update centroids. Categorical variables bring together several scales with a finite and moderate set of values, such as nominal variables, ordinal variables, discrete interval variables, and continuous variables grouped into few categories. As an extension of K-means, K-modes aims to cluster categorical data by using the mode to update clusters’ centroids, as well as an overlap measure to define the distance between objects and centers. The extension of AGNES can follow the same logic. This work proposes the Frequency Factor, a new probability-based measure, specifically forged to update the centroids on clustering algorithms. The use of Frequency Factor to edit both K-means and AGNES yielded two proposed alternatives, K-fact and MARI, respectively. In our validation scheme, we used the Rand index from external validity criteria to measure the clustering performance and compare both algorithms. Those algorithms extended by Frequency Factor outperform or tie with K-modes, AGNES, and ROCK and can be used as a competitive alternative to cluster nominal-categorical variables.

The increase in the amount of text generated daily on the web has made it more difficult the analysis and extra information of these data. In addition, the number of forms with that people can express themselves using natural language increases the complexity the development of computational tools to extract useful information from texts. Extra Information (EI) has the role of automating the transformation of corpus textual information in structured information. Extra informations of texts are structured in different ways. The relational triple of the type (subject-predicate-object) and structure most used in IE. The EI task can be classified as traditional and open. THE The traditional one makes extra cations from a limited set of predicates. Haha In addition to Open Information (EIA), the relationships are not limited by a set of therefore, one of the objectives of this task and the independence of the domain. The EIA can be divided into two stages: (i) extra and (ii) classi cation. M etodos of EIA present in the literature use spe- cific language functions of the target language in their stages of classi - cation. Language resources for English are more the other languages, thereby increasing the OIE's languages. Most of these methods are based on English. Therefore, the work, and in the classi cation stage, use a set of multi-language features to classi cation of relational triples, contributing to language independence in EIA. Experiments were performed to compare the multi-language features with the proposals by ReVerb. The results obtained showed that, at the 5% level of significance, the multi-language features presented superior performance for English (p <0.01). After the to translate the part of the ReVerb features, both sets achieved similar in Spanish (p = 0.76) and Portuguese (p = 0.07)..

Many approaches have emerged in the last few years to address the challenges of dynamic
hand gesture recognition systems. These systems use mostly computer vision techniques
based on RGB images, subject to limitations such as variation of light and color. The
emergence of RGB-D sensors allowed to incorporate geometric information into the image,
making recognition systems more robust. However, there are still dierent challenges
related to the problem of the classication of gestures as invariance to physical aspects
and location, choice of characteristics and robust classiers that allow classication in
real time. This work proposes an analysis of classiers for dynamic 3D hand gestures
applied to geometric descriptors that represent the 3D curve generated by the hand of
the user over time. Two approaches were analyzed, Dynamic Time Warping (DTW) and
Decision Tree (DT). Two datasets were used to evaluate the results: MSR Action3D and
GRUFBA 3D, which was built by the author. The classiers were evaluated to identify
the importance of the curve tting steps, and the extraction and selection of the curve
characteristics, analyzing their impact on the gesture recognition rate. The obtained
results show that this approach is competitive with the other techniques of the state of
the art for similar approaches, being invariant to physical aspects and lighting with an
average rate of 97:2% and 97:6% with the textit DTW and textit DT, respectively, in
one of the analyzed datasets.

Organizations are increasingly aware of the importance of mapping their business processes. Business process modeling helps to identify and understand the activities performed by participants of an organization, as well as providing support for specifying system requirements and defining services that support them. In this sense, it is important to define methodologies that facilitate the business process models reuse to identify relevant services for the development of service-oriented applications. Several proposals in the literature follow in this direction, however, there is still no consensus on which it is the best way to use these models to identify the services, because the transference of information between models of different levels of abstraction is an arduous task. One of the problems found in these proposals is that the derivation occurs straight from business level to service level, without considering, for instance, system requirements that support them. Therefore, it generates basic services that requires more information to reach an architectural level that can properly generate code. To find a solution to this problem, this dissertation proposes DERIVA (Service Derivation on Business Process), which is a process based on Model-Driven Development approach, and considers not only business processes, but also system requirements that support them to achieve service derivation. The DERIVA process performs service derivation through a semiautomatic transformations’ chain, where process models are transformed into less abstract models until code generation is accomplished. The process transformation’s chain is heuristic based, which served as source for the mapping. This process was evaluated through a case study and a controlled experiment by developers of different knowledge in Service Oriented Architecture and Model-Driven Development. The results of this study demonstrated that these professionals, from a process model and using DERIVA process, were able to execute the entire process, produce a service architecture and its Java’s code, with parameters and data types, highlighting the proposal viability and efficiency. Service derivation from the business process can be facilitated through model-driven development. In addition, the inclusion of system requirements modeling, as part of the derivation process, results in a more detailed service architecture.

Concept Location is the step of software modification where the developer
search for the chunks of code that must be modified. In industrial environments, this
step consumes most of the time and effort. In essence, the concept location is a problem
of software comprehension. In reading the source code, cognitive models point out that
the developers refine their knowledge by elaborating and verifying hypotheses from the
tips suggested by the identifiers. Problem - In the literature, Dependence Search (DepS)
is one of the most recognized techniques for concept location. After selecting an initial
code snippet called seed, DepS suggests inspecting static dependencies, providing only
the immediate dependencies of the seed. However, we did a study of the dependency
graph of five softwares and found that in some cases the nearest dependencies use terms
similar or equal to the terms of the seed. For those who do not know the terms of the seed,
the terms similar to or equal to those used by the seed will hardly refine the hypotheses
about the concepts implemented. Proposal - In order to refine the DepS, we developed
a new technique called Microdomain-based Concept Location (MCL). The objective of
the technique is to provide methods with terms different from those used by the seed and
that they were at the same time statically related to each other. To achieve this goal,
we developed two concepts: Microdomain and Seed Context. A microdomain is a cluster
of statically related methods, built on the basis of dominance analysis. When selecting
a seed, the MCL presents graphically the seed context, constituted by the microdomain
of the seed and the microdomains neighboring the microdomain of the seed. In order to
support the use of MCL, we developed a tool called the Context Viewer. Evaluation - In
order to collect evidence of the applicability of the technique, we assigned a task of concept
location to a developer, who used the Context Viewer. At the end of the evaluation, we
conducted an interview. Conclusion - The tips provided by seed context, built by
Context Viewer, helped the developer to realize the relevance of a seed, completing the
task with the desired code snippets. Despite the evidence collected in our study, further
research needs to be conducted in order to measure the similarity between the terms of
one microdomain and the difference between the terms of microdomain neighbors.

-As part of the unsupervised machine learning area, time series clustering aims at designing methods to extract patterns from temporal data in order to organize dierent series according to their similarities. According to the literature, most of researches either perform a preprocessing step to convert time series into an attribute-value matrix to be later analyzed by traditional clustering methods, or apply measures specically designed to compute the similarity among time series. Based on such studies, we have noticed two main issues: i) clustering methods do not take into account the stochastic and the deterministic in uences inherent of time series from real-world scenarios; and ii) similarity measures tend to look for recurrent patterns, which may not be available in stochastic time series. In order to overcome such drawbacks, we present a new clustering approach that considers both in uences and a new similarity measure to deal with purely stochastic time series. Experiments provided outstanding results, emphasizing time series are better clustered when their stochastic and deterministic in uences are properly analyzed.

The use of the Internet has undergone major transformations, the number of devices and connections is with the pace of accelerated growth, unplanned facts in the network architecture project. As a result, the popularization of mobile devices is leading to increased demand for mobile Internet and an overload in access networks. Faced with the challenges posed by emerging communications, Named Data Networking (NDN) were proposed to solve the problems faced in the current network, through a content-centric communication model, in which consumers request interests in the network in search of data that are generated by the producers. NDN supports the mobility of consumers through the architecture itself, however, support for the mobility of producers still poses a challenge, due to damage to the network, especially during the period of unavailability of the producer. In this way, this work proposes the SCaN-Mob, a new strategy to support the mobility of producers in NDN. SCaN-Mob uses caching on local network devices to help producers disseminate their data between neighboring nodes, to maintain in the device cache the data from the producers most likely to perform handoff . Thus, a greater diversity of contents in the network is obtained, which increases the chances of responding to the interests that arrive in the absence period of the producer. The main contributions of this work are: i) a cache replacement policy that can act with different data prioritization criteria; and ii) a packet forwarding strategy to control the effect of packet flooding on the network. The proposal was implemented in the ndnSIM simulator and was submitted to a detailed analysis, comparing it with the standard NDN, under different metrics. The results obtained show gains of up to 35.32% in the interest satisfaction rate, even with the producer's unavailability of the producer, and with a reduction of 56.23% in the time to obtain data.

Data collected in recent years show that the amount of information traveling in the network core has been increasing at an exponential rate, evidencing the inability of Wavelength-division Multiplexing networks to meet such demand. Elastic Optical Networks represent a new approach to dealing with high data traffic, offering bandwidth closer to that requested by users, providing greater use of available optical resources. In scenarios with high traffic volume, requests can be blocked due to the inability of the network to allocate new resources. One of the strategies used to avoid blocking new requests is service degradation, where certain requests may receive less than requested bandwidth so that they can operate without loss. In the current literature it is possible to perceive that most of the studies approach the degradation of service using a static scenario where all the requests are known in advance, resulting in the absence of online strategies that present greater conformity with real scenarios. This paper proposes an online strategy for service degradation using the proportional QoS model. This strategy aims to reduce the likelihood of blockage caused by lack of optical resources and offer better treatment for different classes of service. In addition, pairs of source-destination nodes were modeled as a queuing system operating on the Generalized Processor Sharing policy with Leaky Bucket admission control in order to quantify the impact of degradation on lightpaths. The results show that the proposed strategy can reduce the probability of blocking and provide network operators with greater control among different classes of degraded services.

In software engineering, the concept of Technical Debt (DT) contextualizes the problem of technical tasks not performed in the present, which may incur costs and adversities in the future. Exposure to debt can bring short-term benefits, such as increasing current productivity and reducing the time to launch a product, but this benefit is obtained at the cost of extra effort in the future to resolve the debts incurred. In this context, it is necessary to develop mechanisms to manage technical debts. Identifying and monitoring DT is an important and necessary step in the debt management process. Unfortunately these tasks are complex because they require the analysis of large amounts of information. Different works have been carried out in the development of approaches to identify DT either automatically or semi-automatically. Most of these works are based on code metrics analysis. However, studies have already pointed out that the use of metrics only for the identification of DT items limits the scope of this activity. It is also necessary to consider the developer's point of view. In this direction, the combination of an approach that simultaneously uses metrics, static code analysis, comment analysis, and information visualization, has been shown to be promising for DT identification and monitoring. This work presents the development and evolution of two tools to materialize this approach. The first, called RepositoryMiner, mines software repositories, calculates metrics, detects code smells, analyzes comments and source code, and combines these different means as indicators of technical debts of software modules. The second, called VisminerTD, uses visual techniques to, based on the calculated indicators, support software engineers in the identification and monitoring of technical debts throughout the life cycle of a software project. To show the usability and effectiveness of the approach, two feasibility studies were conducted showing that the tools can support the management of DT in medium-sized software projects.

MOOC is an acronym for Massive Open Online Course, an online education modality
oering video courses, with dierent ages, backgrounds and nationalities. The MOOCs
have emerged as technological innovations that idealize a democratization of education,
through a course of free courses. This modality of teaching has as characteristic a population
of dierent ages, formations and nationalities. However, MOOCs are subject to a
high rate of their students, corresponding to their main criticisms. You can participate
at any time, a completion rate for most courses of 13%. Some of the conventions students
attend are those related to the technology platform. In this sense, the research had
to investigate the quality of semiotic metacommunication inuencing the permanence or
the abandonment of education in MOOC courses. The research was carried out in four
stages. The rst one consisted in the execution of a literature review, in order to identify
the elements of the nature of the genetics that infuenced the permanence and the abandonment
in MOOC. With the review vote, twenty-four general causes of abandonment in
MOOC were found. The second stage was held a study of documentary events moments:
development of a platform of MOOC; evaluation of the quality of the metacommunication
of the interaction design of a course in this platform; course implementation; and, nally,
to identify and evaluate the ruptures of communicability existing in the case studied. In
the design-time evaluation, a MOOCLICAR technique was applied, which was mitigated.
The data sheet with the data of the course of analysis and analysis of the course was not
carried out. Finally, the Semiotic Inspection (SIM) and Communality Assessment (CEM)
methods were used to identify communication disruptions in the MOOC platform. The
third phase of the session had an analysis of the collected data, relating the evaluation
of the interaction design, the questionnaires, the analysis of forums and the assessment
of communicability. The results can be useful for the development of a set of interaction
design and evaluation guidelines for the MOOC from a communication point of view.