Explorando calibragem ponderada, balanceamentos e métricas para justiça em sistemas de recomendação

Abstract

Recommendation Systems are tools used to suggest items, which possibly will be interesting to the users. These systems are based on the user’s preference historic to generate a suggestion list that has higher similarity with the items in the user’s historic, providing a better precision and minor error. It is expected that when an item is recommended, the user will send a feedback, indicating if he/she likes or dislikes the recommendation. The user interaction with the system provides a better way to understand the preferences of the user, who, over time, adds more and more itens in its profile. The recommendation based on item similarity with the user preference searching for better precision, can cause collateral effects in the recommendation list, such as: superspecialization of the items in the list with a little kernel of items, few diversity of categories and unbalanced genres. Thus, this dissertation aims to explore the calibration, which is a way to produce recommendations that will be relevant to the users and at the same time considering all genres in the preferences, trying to avoid a misproportion in the recommendation list. To this end, ways are addressed to weight the tradeoff between the relevance of the recommendations and the calibration based on divergence measures, as well as a calibrated system framework and decision protocol. The hypothesis are that calibration can contribute positively to provide more fairer recommendations. The research is accomplished with a large approach proposing a framework and a decision protocol that contemplates nine recommender algorithms applied on the domains of movie and music, analysing three divergence measures, two personalized tradeoff weights and two tradeoff formulations between relevance-calibration. The evaluation is analysed with well-known metrics, as well as proposed metrics. The results indicate that the calibration produces positive effects to the precision and to the fairness, creating a recommendation list that respects the genre distribution. The results also indicate which is the best system combination to obtain a better performance when applying the proposed calibration