Collaborative filtering with a graph-based similarity measure

Collaborative filtering is a technique widely used in recommender systems. Based on behaviors of users with similar taste, the technique can predict and recommend products the current user is likely interested in, thus alleviates the information overload problem for Internet users. The most popular collaborative filtering approach is based on the similarity between users, or between products. The quality of similarity measure, therefore, has a large impact on the recommendation accuracy. In this paper, we propose a new similarity measure based on graph models. The similarity between two users (or symmetrically, two products) is computed from connections on a graph with vertices being users and products. The computed similarity measure is then used with the k - nearest neighbor algorithm to generate predictions. Empirical results on real movie datasets show that the proposed method significantly outperforms both collaborative filtering with traditional similarity measures and pure graph-based collaborative filtering.