Mutual information based similarity measure for Collaborative Filtering

Abstract

Collaborative Filtering(CF) is one of the most successful recommender systems. The most critical step in CF is similarity computation. In CF, similarity is used for neighbor search. In addition, it will be used as a weighted coefficient during the prediction step. Typically, three different similarity measures are used: cosine based similarity, Pearson correlation coefficient based similarity and adjusted cosine based similarity. However, as these methods are based on linear correlations, they are also limited. Indeed the linear correlation only takes into account the linear part of the correlation. This paper introduces a new similarity measure based on mutual information to avoid the above limitation. The experiments, done on MovieLens data sets, show that this new method outperforms traditional similarity measures under the nonlinear correlation circumstance.