Nonparametric statistics inspired similarity measure for accuracy in collaborative filtering

Similarity computation is especially significant in collaborative filtering algorithms. In the existed literatures or large recommender systems, researchers generally use cosine similarity or Pearson correlation coefficient to compute the similarities. However, on one hand, both two measures are more applicable in linear space while the ratings cannot linearly reflect users' interest intensity. On the other hand, different systems have different rating restricts, so it's unreasonable to treat varied systems with the same computation method. In this work, inspired by nonparametric statistics, a sound memory-based method with a novel similarity measure is proposed to eliminate the affection of rating distributions and system diversities on similarity computation. Experiments based on Jester are set to evaluate the performance of the approach. And the highly competitive results show that our approach is able to significantly improve the recommendation quality. © 2013 Binary Information Press.