Utilize Item Correlation to Improve Aggregate Diversity for Recommender Systems

Abstract

Nowadays, more and more researchers have noticed that recommenders performing well with respect to accuracy do not necessarily mean that they produce the most effective recommendations in practice. Aspects such as aggregate diversity were identified as additional factors that can positively or negatively affect the quality of recommendations. The major challenge for improving aggregate diversity for a recommender system lies in the data imparity problem for popular items and niche items. Recommender systems are more inclined to recommend popular items that maintain sufficient user feedback while ignoring niche items that have been rated by a small portion of users. To address this problem, we associate popular items and niche item via item features. The correlation between items are calculated by meta-path similarity measurement in a heterogeneous information network. Finally, we propose an Item Correlation based Probabilistic Matrix Factorization method, incorporating multiple item features and collaborative data. Empirical studies based on a real world data set demonstrate the effectiveness of this method and show that it outperforms state-of-the-art methods for improving aggregate diversity.