Personal Values Estimation Methods for Collaborative Filtering

Abstract

Personal values (PVs) have been used in recommender systems, and RMRate, lift, and Laplace value have been proposed to estimate the PVs. In this study, we propose a method for estimating PVs using coefficients of the linear regression model. The proposed method is applied to collaborative filtering (CF) based on k-nearest neighbor. The CF model with the proposed PVs estimation method forms clusters with users who have similar PVs and creates a recommendation list by using the evaluation score of items that users in the cluster have evaluated in the past. We evaluated such list using Mean Reciprocal Rank (MRR) and normalized Discounted Cumulative Gain (nDCG), and these accuracy scores were improved from 0.044 to 0.081 for MRR and from 0.052 to 0.066 for nDCG, compared to the conventional CF. In contrast, compared with the conventional PVs-based CF models, the scores of MRR and nDCG were same level because from -0.015 to 0.008 for MRR and from -0.008 to 0.022 for nDCG.