Comprehensive Assessment of Robustness in Fairness of GNN-based Recommender Systems against Attacks

The robustness of recommendation models is typically measured by their ability to maintain the original utility when exposed to attacks. In contrast, robustness in fairness pertains to the resilience of fairness levels in the presence of such attacks. Despite its significance, this latter area remains largely underexplored. In this extended abstract, we evaluate the robustness of graph-based recommender systems with respect to fairness from both the consumer and provider perspectives, under attacks involving edge-level perturbations. We analyze the impact of these perturbations on fairness through an experimental protocol involving three datasets and three graph neural networks. Our findings reveal severe fairness issues, particularly on the consumer side, where fairness is compromised to a greater extent than on the provider side. Source code: https://github.com/jackmedda/CPFairRobust.