Abstract
Nearest-neighbor collaborative filtering (CF) algorithms are gaining widespread acceptance in recommender systems and e-commerce applications. These algorithms provide recommendations for products, based on suggestions of users with similar preferences. One of the most crucial factors in the effectiveness of nearest-neighbor CF algorithms is the similarity measure that is used. The most popular measures are the Pearson correlation and cosine similarity. In this paper, we identify existing fallacies in the calculation of these measures. We propose a novel approach, which addresses the problem and substantially improves the accuracy of CF results. Moreover, we propose an evaluation procedure that produces reliable conclusions about the performance of nearest-neighbor CF algorithms. Through the proposed evaluation procedure, our experimental results identify the problems of existing approaches (which could not be revealed with existing evaluation procedures) and illustrate the superiority of the proposed approach.