Improving Accuracy of Multi-Criteria Collaborative Filtering By Normalizing User Ratings

Abstract

Multi-criteria collaborative filtering schemes allow modeling user preferences in a more detailed manner by collecting ratings on various aspects of a product or service. Although preferences are expressed by numerical ratings within a predetermined scale, it is not guaranteed that users comprehend such scale identically. As a result, profiles of users with similar tastes might turn out to be unrelated. Besides, distinct criteria might have different rating scales creating an essential incompatibility with the rating schemes of users which in turn conceals proper relation between main criterion and sub-criteria. Since users rate items based on their personal rating habits, it is essential to determine user similarities according to their rating patterns by normalizing ratings to an identical scale. In this paper, two different normalization methods are studied, i.e., z-score normalization and decoupling normalization, in order to improve accuracy of multi-criteria collaborative filtering systems. In particular, two normalization methods are employed by modifying the state-of-the-art memory-based multi-criteria recommender schemes so that similarities among users are calculated based on preference models rather than pure numerical ratings. Real world data-based experimental results show that both methods, especially decoupling normalization method, provide significant improvements on accuracy of estimated multi-criteria predictions and outperform previous pure numerical ratings-based approach.

Multi-criteria collaborative filtering schemes allow modeling user preferences in a more detailed manner by collecting ratings on various aspects of a product or service. Although preferences are expressed by numerical ratings within a predetermined scale, it is not guaranteed that users comprehend such scale identically. As a result, profiles of users with similar tastes might turn out to be unrelated. Besides, distinct criteria might have different rating scales creating an essential incompatibility with the rating schemes of users which in turn conceals proper relation between main criterion and sub-criteria. Since users rate items based on their personal rating habits, it is essential to determine user similarities according to their rating patterns by normalizing ratings to an identical scale. In this paper, two different normalization methods are studied, i.e., z-score normalization and decoupling normalization, in order to improve accuracy of multi-criteria collaborative filtering systems. In particular, two normalization methods are employed by modifying the state-of-the-art memory-based multi-criteria recommender schemes so that similarities among users are calculated based on preference models rather than pure numerical ratings. Real world data-based experimental results show that both methods, especially decoupling normalization method, provide significant improvements on accuracy of estimated multi-criteria predictions and outperform previous pure numerical ratings-based approach.