Locally Adaptive Neighborhood Selection for Collaborative Filtering Recommendations

User-to-user similarity is a fundamental component of Collaborative Filtering (CF) recommender systems. In user-to-user similarity the ratings assigned by two users to a set of items are pairwise compared and averaged (correlation). In this paper we make user-touser similarity adaptive, i.e., we dynamically change the computation depending on the profiles of the compared users and the item whose rating prediction is sought (target item). We propose to base the similarity between two users only on the subset of co-rated items which best describes users’ tastes with respect to the target item. These are the items which have the highest correlation with the target item. We have evaluated the proposed method using a range of error measures and showed that the proposed locally adaptive neighbor selection, via item selection, can significantly improve the recommendation accuracy compared to standard CF.