Inspiration des sondages d'opinion pour réduire la latence en filtrage collaboratif

Collaborative filtering is one of the most popular approaches in recommender systems. In collaborative filtering, the system exploits the ratings of the active user’s like-minded users: his neighbors, to estimate his ratings on the items he has not rated yet and recommend him the items with the highest estimated ratings. Collaborative filtering faces a latency problem: a new item cannot be recommended to any user while this item has not been rated a sufficiently high number of times. To alleviate this latency problem, we propose a new approach based on mentors, inspired from opinion polls. A mentor is a reliable, representative and trusted user. The knowledge of the ratings of the mentors on new items allows to estimate the ratings of the whole population. We show that, on the corpus tested, when using some of the mentor selection methods, only 6 mentors are sufficient, to reliably estimate the ratings of the whole population. MOTS-CLÉS : Systèmes de recommandation, Filtrage collaboratif, Sondages d’opinion.