A developed collaborative filtering similarity method to improve the accuracy of recommendations under data sparsity

Abstract

This paper presented a new similarity method to improve the accuracy of traditional Collaborative Filtering (CF) method under sparse data issue. CF provides the user with items, that what they need, based on analyses the preferences of users who have a strong correlation to him/her preference. However, the accuracy is influencing by the method that use to find neighbors. Pearson correlation coefficient and Cosine measures, as the most widely used methods, depending on the rating of only co-rated items to find the correlations between users. Consequently, these methods have lack of ability in addressing the sparsity. This paper presented a new proposed similarity method based on the global user preference to address the sparsity issue and improve the accuracy of recommendation. Thus, the novelty of this method is the ability to solve the similarity issue with a capability of finding the relationship among non-correlated users. Furthermore, to determine the right neighbors during the process of computing the similarity between a pair of users, the developed method considered two main factors (fairness and proportion of co-rated). The MovieLens 100K benchmark dataset is used to evaluate the developed method accuracy. The experiments' result showed that the accuracy of the developed method is improved compared to the traditional CF similarity methods using a specific common CF evaluation metrics.