An Empirical Comparison of Outlier Detection Methods for Identifying Grey-Sheep Users in Recommender Systems

Collaborative filtering is a popular recommendation technique predicting user preferences through the analysis of similar users’ historical behaviors, offering personalized recommendations based on shared interests. While collaborative filtering algorithms are widely used, they face well-known challenges like rating sparsity, cold-start problems, and the presence of grey-sheep users. The grey-sheep users are the users with uncommon item preferences, and they can be treated as outliers. It is surprising that the outlier detection technologies were not fully examined to identify the grey-sheep users. In this paper, our study addresses this gap by empirically comparing multiple state-of-the-art outlier detection methods and also introducing novel user representations to enhance the outlier detection process in recommender systems.