Enhancing recommender systems : a comprehensive study from effectivenss to fairness

Abstract

In the age of information overload, recommender systems play a crucial role in helping users discover relevant and personalized content in various areas, such as e-commerce, entertainment, and social media. In this thesis, we study how to improve the effectiveness and fairness of existing recommender systems. Specifically, we propose new techniques for the effectiveness of meta-path-based and multi-domain recommendations. Furthermore, we propose a new framework to handle the popularity bias issue to improve the fairness of recommender systems.

First, we investigate meta-path-based recommender systems (MPRs) and how to automatically find effective meta-paths as input to these MPRs. We observe that the performance of MPRs is highly sensitive to the meta-paths they use, but existing works manually select the meta-paths from many possible ones. Thus, we propose the Reinforcement learning-based Meta-path Selection (RMS) framework to discover effective meta-paths automatically. We also propose a new MPR called RMS- HRec, which uses an attention mechanism to aggregate information from the meta-paths. We conduct extensive experiments on real datasets. Compared with the manually selected meta-paths, the meta-paths identified by RMS consistently improve recommendation quality, and RMS-HRec outperforms state-of-the-art recommender systems in terms of recommendation accuracy.

Second, we study how to conduct more effective multi-domain rec- recommendation (MDR). We find that existing MDR models are difficult to disentangle knowledge that generalizes across domains (e.g., a user likes cheap items) and knowledge specific to a single domain (e.g., a user likes blue clothing but not blue cars) due to their model-level disentangling mechanism. Therefore, we propose an embedding disentangling recommender focusing on disentangling at the embedding level. We also find that existing MDR methods have limited ability to transfer knowledge across domains with small overlaps, so we propose a random walk-based domain alignment strategy to identify similar users/items from different domains, which further helps knowledge sharing. Extensive experiments show that our method consistently outperforms the baselines on all datasets and domains.

Finally, we focus on the popularity bias issue in recommender systems, which is important to the fairness of recommendations. Popularity bias is the phenomenon that popular items are recommended much more frequently than they should be. This goes against the goal of providing personalized recommendations and harms user experience and recommendation accuracy. Therefore, we aim to increase the personalization of recommendations to handle it. We find that existing debiasing methods overlook that popularity is defined in a "global" manner (thus, we call it global popularity (GP) in this thesis), which is defined as the proportion of all users who have interacted with the item. Hence, we propose "Personal popularity" (PP), which considers the interests of individual users. Furthermore, we propose the Personal Popularity Aware Counterfactual (PPAC) framework that integrates PP and GP to handle popularity bias. Experiment results show that our method can effectively reduce the recommendations of popular items, enhancing the fairness of recommendations.