An Adaptive User-Genre-Item Model for Collaborative Filtering

Collaborative filtering provides personalized recommendations based on individual user preferences as well as those of other users with similar interests. In collaborative filtering, memory-based approaches make predictions by measuring the whole similarity between two users. When a user has multiple interest genres, those methods seem too optimistic in making correct predictions in some situations. In addition, minor genres are often inhibited due to their minute share of the whole similarity. In this paper, we present a novel approach that combines the advantages of item-item similarity and user-user similarity by introducing a genre component to the relation between user and item. In our approach, the direct user-item relevance is developed into the combination of genre similarity and preference similarity, thus capturing more accurately the relevance between items as well as between user and item. Experimental results from EachMovie and MovieLens datasets show that our approach outperforms four other state-of-the-art collaborative filtering algorithms.