Banca de DEFESA: DIEGO CORREA DA SILVA

Exploring Weighted Calibration, Balancing, and Metrics for Justice in Recommender Systems

calibration, fairness, balancing, recommendation systems, accuracy

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.