A belief propagation based recommender system for online services

Abstract

In this paper we report our progress in the first application of iterative probabilistic algorithms in the design and evaluation of recommender systems. The proposed iterative recommender system (referred to as BPRS) is based on the belief propagation, a powerful decoding algorithm for turbo codes and Low-Density Parity-Check (LDPC) codes. The belief propagation algorithm relies on a graph-based representation of an appropriately chosen factor graph for the recommender systems. The factor graph representation of the recommender systems turned out to be a bipartite graph, where the users and products are arranged as two sets of variable and factor nodes that are connected via some edges. Recommendations (predicted ratings) for each particular user can be computed by probabilistic message passing between nodes in the graph. We provide an evaluation of BPRS via computer simulations using the MovieLens dataset. We observed that BPRS iteratively reduces the error in the predicted ratings of the users until it converges. Further, our initial results indicate an improvement in the Mean Average Error (MAE) and Root Mean Square Error (RMSE) over the Item Averaging. Therefore, we are confident that the belief propagation is a new promising approach which will offer robustness and accuracy for the recommender systems.