Deep Fair Graph Clustering via Adversarial Clustering Structure Debiasing

Deep fair graph clustering (DFGC) aims to learn unbiased graph representations that prevent clustering results from being dominated by specific sensitive groups. Existing methods typically mitigate sensitive-attribute bias by training an adversarial discriminator to predict sensitive attribute labels from node representations. However, this node-wise prediction objective overlooks the sensitive-group clustering structure encoded by graph topology. Consequently, such structural bias may persist in the learned node representations, yielding clusters that are overly aligned with sensitive-group boundaries. To address this issue, we propose a novel DFGC framework termed **A**dversarial **C**lustering **S**tructure **D**ebiasing **(ACSD)**, which learns fair graph representations by adversarially debiasing sensitive clustering structure. Specifically, a sensitive adversarial encoder is designed to extract sensitive clustering structure through modularity maximization and sensitive cluster-membership reconstruction. The clustering backbone is then optimized to hinder the extraction of sensitive clustering structures while preserving clustering quality, promoting graph clustering embeddings that are less dependent on sensitive-group structures. Extensive experiments on benchmarks demonstrate the effectiveness and superiority of ACSD over existing state-of-the-art methods.