MetaLProjection: An Approach for Recommending Dimensionality Reduction Algorithms Using Meta-Learning
Information visualization, multidimensional projections, meta-learning, dimensionality reduction
Data visualization techniques have significant potential for analysis, summarization, and comprehension to facilitate information extraction. Recent areas such as Visual Analytics and Data Science underscore their importance, especially in analyzing complex datasets. In this sense, Multidimensional Projection techniques are particularly used for visual analysis of high-dimensional datasets because they perform dimensionality reduction and, consequently, exhibit better scalability in terms of the number of attributes/dimensions. However, there is a wide variety of these projection techniques, and determining the most suitable one for discovering visual patterns of information in one or multiple datasets is not a trivial task. While works in the literature test and compare different techniques on datasets with distinct characteristics, they do not do so systematically to assist user decision-making. In this context, this research leverages meta-learning to classify and recommend multidimensional projections, considering specific evaluation metrics from a knowledge base comprising over 500 distinct datasets. To evaluate the approach, we observed i) the relationship between the meta-attributes of all datasets, ii) the generation of a ranking containing the performance of the selected projection techniques, and iii) the accuracy of the recommendation of these techniques. Finally, the results obtained demonstrate that the developed approach effectively contributes to the selection and recommendation of multidimensional projection techniques.