IntFair:Graph Neural Networks for Fair Recommendations with Interest Awareness

With the rapid development of recommender systems, it has been found that accuracy is not the only standard for evaluating the utility of recommender systems. In real-world recommendation scenarios, users' interests are typically diverse. However, due to the influence of popular bias, recommender systems vary significantly in their recommendation utility for different interests, which is unfair to users with multiple interests. In this paper, we propose the concept of interest-aware fairness, which suggests that the recommendation accuracy should be as consistent as possible between users' different interests. Additionally, we introduce a metric to quantify this fairness problem. To achieve interest-aware fairness, we propose IntFair, which contains three modules: The Balanced Category Sampling mitigates information redundancy during aggregation, Multi-Interest Capture Network encode the diverse interests of users, and Hybrid Expert System dynamically adjusts to different combinations of user interests. Experimental results on real datasets demonstrate that compared to other recommendation models, IntFair can maintain a comparable level of accuracy while achieving interest-aware fairness recommendation.