Face to Fairness: Mitigating Influence Bias for Fair Graph Anomaly Detection in Dynamic Graphs

Yifan Hong, Jiao Luo

Keywords: Graph Anomaly Detection, Influence Bias, Graph Neural Networks, Dynamic Graphs, Wasserstein GAN, Adversarial Training, Temporal Modeling

Abstract:

Graph anomaly detection plays a vital role in applications such as fraud detection, cybersecurity, and social network analysis. Despite recent advances in graph neural networks (GNNs) for dynamic graphs, existing methods often suffer from an overlooked issue: influence bias caused by nodes with high structural centrality. Such nodes dominate message passing and feature aggregation, leading to biased anomaly detection where anomalies around high-influence regions are overemphasized, while anomalies on low-centrality nodes remain under-detected. Severe class imbalance further exacerbates this challenge by limiting the model's ability to learn from scarce, structurally marginal anomalies.

In this work, we propose FairGAD, a fairness-aware framework for dynamic graph anomaly detection that mitigates influence bias from a causal representation learning perspective. Our model integrates a Graph Expert Mixture Encoder that combines complementary structural experts and temporal modeling to produce robust node representations, a Representation-Space Anomaly Generator that learns the distribution of anomalous embeddings and synthesizes diverse tail anomalies to alleviate class imbalance, and an Influence Debiasing Classifier that employs adversarial training with a gradient reversal layer to learn influence-invariant representations by attenuating the spurious path from node influence to anomaly prediction.

Extensive experiments on multiple real-world dynamic graph datasets demonstrate that FairGAD not only improves anomaly detection performance but also substantially reduces influence bias, yielding more equitable detection across nodes with varying influence levels. For instance, on the UCI and ALPHA datasets, our approach achieves relative F1-score improvements of 0.89% and 1.53% over strong baselines while narrowing performance gaps between high- and low-centrality nodes.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 5