Understanding the Impact of Individual Users' Rating Characteristics on Predictive Accuracy of Recommender Systems

Abstract

In this study, we investigate how individual users' rating characteristics aect the user-level performance of recommendation algorithms. We measure users' rating characteristics from three perspectives: rating value, rating structure and neighborhood network embeddedness. We study how these three categories of measures in uence the predictive accuracy of popular recommendation algorithms for each user. Our experiments use ve real-world datasets with varying characteristics. For each individual user, we estimate the predictive accuracy of three recommendation algorithms. We then apply regression-based models to uncover the relationships between rating characteristics and recommendation performance at the individual user level. Our experimental results show consistent and signicant eects of several rating measures on recommendation accuracy. Understanding how rating characteristics aect the recommendation performance at the individual user level has practical implications for the design of recommender systems.