FR: A Novel Strategy to Improve Item-Oriented Collaborative Filtering

Abstract

Traditional Item-oriented Collaborative Filtering (ICF) suffers the problem of low recommendation accuracy in spite of its popularity. In this paper we propose two novel strategies respectively named Recommendation Bias Removal (RBR) and Folksonomy Network Based Filtering (FNBF) to address this issue; with a special focus on the accuracy of individual recommendations. RBR statistically estimates the Recommendation Bias (RB) between system predictions and actual user ratings, and then modifies the system predictions via removing the corresponding RB estimations. FNBF first builds the Folksonomy Network (FN) model based on the folksonomy data, and then applies the FN model to filtering out the irrelative recommendation results of ICF. We further combine FNBF and RBR to form the FR strategy which can further improve the recommendation quality based on information from both ratings and folksonomy data. The empirical studies on large real datasets showed that both FNBF and RBR were effective on the accuracy issue; furthermore, a recommender with FR could make more accurate recommendations due to the aggregating effect of FNBF and RBR.