Contribution Level Calculation to Similarity Measuring in Collaborative Filtering

Abstract

Collaborative filtering is the well-known recommendation technique to make a prediction for the given user/item pair based on the rating of similar users/items. Hence, it largely relies on the similarity measure of the users/items. Pearson similarity and Cosine similarity are two of the most famous similarity measures but these measures just take into account the similarity of common ratings and thus have some defects when applying to the realistic application. In addiction to plain similarity measures, we propose a contribution level calculation to similarity measures and discuss the neighborhood selection of similar user/items to further improve the rating prediction. Experiments are evaluated over the classic datasets of recommendation problems and the result shows that the newly proposed measure on the similarity calculation is better than the original one.