An Empirical Study on the Effect of Different Similarity Measures on User-Based Collaborative Filtering Algorithms

Abstract

Computation of similarity between user profiles (user rating vectors) is one of the core components of user-based (k-nearest-neighborhood based) collaborative filtering algorithms. Present techniques work by identifying or selecting a similarity function by the designer of the recommendation engine and keeping it fixed throughout the collaborative filtering process and using the same function to compute the neighborhood of every user. However, we found that there is no single similarity measure that gives best predictive accuracy for all users. We see this as a limitation of current systems. For the same user, applying different similarity functions results in different predictive accuracy. We propose that the accuracy of user-based collaborative filtering recommendation engines can be further increased by learning an optimal similarity function for a particular user and by applying different similarity measure for different users. We present an empirical study on the effect of eleven different similarity measures on the predictive accuracy of user-based collaborative filtering algorithms.