Graph-Based Collaborative Filtering Using Rating Nodes: A Solution to the High Ratings/Low Ratings Problem

Graph-based random walk models have recently become a popular approach to collaborative filtering recommendation systems. Under the conventional graph-based approach, a user node and item node are connected if the user has rated the item, and the value of the rating is represented as the weight of the connection. Commencing from some target user, a random walk is performed on the graph, and the results used to perform useful tasks such as ranking items in order of their importance to the user. Because random walk favors large-weighted connections, walk is more likely to proceed through two users that share a high rating for some item, than through users who share a low rating. This is a problem because there are similarity relations implicit in the data that are not being captured under this representation. We refer to this as the ‘High Ratings/Low Ratings’ problem. This paper proposes a novel graph representation scheme in which item ratings are represented using multiple nodes, allowing flow of information through both low-rating and high-rating connections. Empirical results on the MovieLens dataset show that recommendation rankings made using the proposed scheme are much better correlated with results in the test ratings, and that under a top-k evaluation, there is an improvement of up to 15 % in precision and recall. An attractive feature of the approach is that it also associates a confidence value with a recommendation.