Aprendendo funções de previsão de notas em métodos de filtragem colaborativa baseada em usuário

Abstract

The large offer of contents nowadays makes it hard to find relevant information. Recommender systems (RS) have been developed to tackle with such information overloading. Such systems are tools that recommend, from a large number of alternatives, the ones that the users will probably be interested in. The main RS applications are based on two approaches, content based filtering and collaborative filtering. Among them, collaborative filtering is the most used one since, in general, it employs a more effective strategy to capture user preferences: to determine groups of users with similar likes and dislikes. The recommendation problem, as viewed by collaborative filtering, can be viewed as the problem of predicting the preference of the user, normally represented as a rating. Traditional systems predict such ratings by means of manually-crafted regression equations obtained by combining different evidences such as: users reputation and its strictness level. As with any other heuristic strategy, there is no guarantee that the used equations are the best for a particular dataset in the sense of minimizing the prediction error. Thus, in this work, we intend to determine if it would be better to learn regression equations instead of using heuristically built ones. Such learned equations should be obtained by using a machine learning regression task to find the most effective combination of evidence on minimizing error. According to our experiments, a simple regression method is able to significantly outperform the best traditional equations using only evidence explored by those equations. Further, features like ratings that neighbors give to item (as all or individually) and user, item and neighbors average ratings have the best performance. Finally, we obtained gain of until 7% over the baseline with trust feature and gain of 6% over baseline without it.