Revisiting Tendency based Collaborative Filtering for Personalized Recommendations

Abstract

Recommender systems alleviates the problem of information overload by providing personalized suggestions to the users. In this context, recently introduced tendency based recommendation technique is proven to be more simple, intuitive and accurate than the traditional collaborative ltering (CF) techniques. This approach computes two important statistics namely, user tendency and item tendency from the rating dataset in order to predict the fi nal rating of an unrated item. Tendency of an item is computed using all the ratings received by it. However, these ratings might include the ratings provided by ambiguous (unstable) users. Another prominent drawback of the tendency based approach is that the tendency of an item is generic and remains unchanged across the users leading to a non-personalized recommendations. In this paper, we propose to compute item tendency in two different aspects. In the first aspect, we use an information theoretic approach to discover the most unambiguous users and utilize their ratings to compute the item tendency. In the second aspect, we compute the item tendency with respect to the active user, thereby making the tendency personalized to the users. Finally, we propose to obtain stable neighbor sets for each active user, thus making the recommendations more appropriate and accurate. Real-world datasets (Yahoo! Music, Netflix and MovieLens) are used to evaluate our approach. Experimental results show that the proposed techniques outperform the tendency based approach and traditional CF approaches across standard performance metrics.