RBIAS: A model of users' response biases in item ratings

Kyung-Wha Park, Byoung-Hee Kim, Tae-Suh Park, and Byoung-Tak Zhang

Abstract

Recommender systems in this information-overloading age have become mandatory tool for various online services on music, books, movies, etc. Although recommender systems have achieved highly personalized performances, they mostly overlook inherent biases within databases of rating values, which are the basic source for collaborative filtering approaches to recommendation. Previous studies on modeling biases of ratings have limitations in that they directly apply various statistical tools that deal with rating values as real-valued observations in interval or ratio scale although these ratings just represent ordinal levels of preferences. We argue that the first preprocessing step for rating data should be the bias reduction in ordinal scale and convert data into the interval or the ratio scale. To reduce biases in ratings, we need to consider user-specific cognitive biases that intrude during active action in selecting movies and grading them. We assume that the cognitive gap between intrinsic and extrinsic preferences of users may be the primary source of these biases. With this assumption, the RBIAS model on rating pattern is introduced which consists of six rating types and criteria to estimate these rating types of users based on descriptive statistics. As a bias-reduction method, RBIAS-based generative model is presented with a rescaling-based inference method on ‘normative’ ratings among people. We demonstrate that the RBIAS model provides an interesting movie profiling tool which uncovers preference patterns behind popular movies. Our bias-reduction method shows improved performances in various state-of-the-art collaborative recommender systems.