Item weighting techniques for collaborative filtering

Abstract

Collaborative Filtering (CF) recommender systems generate rating predictions for a target user by exploiting the ratings of similar users. Therefore, the computation of user-to-user similarity is an important element in CF; it is used in the neighborhood formation and rating prediction steps. In this paper we investigate the role of item weighting techniques. An item weight provides a measure of the importance of an item for predicting the rating of another item and it is computed as a correlation coefficient between the two items' rating vectors. In this paper we analyze a wide range of item weighting schemas. Moreover, we introduce an item filtering approach, based on item weighting, that works by discarding in the user-touser similarity computation the items with the smallest weights. We assume that the items with smallest weights are the least useful for generating the prediction. We have evaluated the proposed methods using two datasets (MovieLens and Yahoo!) and identified the conditions for their best application in CF. © 2009 Springer-Verlag Berlin Heidelberg.