Factorisation en matrices non négatives pour le filtrage collaboratif

Collaborative filtering (CF) systems help a user find the information satisfying his needs by using other users’ judgements. These systems are mainly used in Web sites to provide recommendations. In general, CF systems ask users to provide a judgement to a number of items of their choice, expressed as a rating. The goal of the system is then to compare different user’s tastes to predict the unknown ratings. We propose a new algorithm for CF based on a dimensionality reduction method called non-negative matrix factorization, well-known in the domains of image recognition or text processing. We test our approach on a movie recommendation task, and we show that it has a performance comparable with the state-of-the-art. Moreover, the method is simple to implement, has low complexity and enables to identify groups of films which have highly correlated ratings among the users. MOTS-CLÉS : filtrage collaboratif, réduction dimensionnelle, apprentissage