Similarity Computation Method for Collaborative Filtering Based on Optimization

Abstract

The recommender system provides personalized recommendations at many e-commerce websites. Collaborative filtering is one of the most popular and effective recommendation algorithms. User-based collaborative filtering, the conventional approach in collaborative filtering, uses user similarity computed based on user item rating. Recommendations are provided by calculating rating predictions based on similarity. Pearson’s correlation coefficient or cosign distance is used as similarity. Until now, a lot of discussions for efficient similarity computation were given by many researchers. Despite active discussion of similarity computation, little computation has been made for optimal similarity in collaborative filtering. In this research, similarity optimization problem was formulated by defining similarities between an active user and other users as a vector variable. The quasi-optimal solution was obtained based on Particle Swarm Optimization (PSO) approach, compared to Pearson’s correlation coefficient. Experimental results based on agentbased simulation and sample dataset show that similarity based on PSO improves recommendation accuracy. We also found that PSO-based similarity computation provides rating predictions for unknown ratings more accurately than conventional similarity computation.