Adaptive Multi-Information Fusion Model for Enhanced Anomaly Detection

Graph Neural Networks (GNNs) have gained increasing popularity for graph anomaly detection tasks in recent years. However, existing research has revealed a significant issue: they learn node representations through the message aggregation mechanism, essentially acting as low-pass filters. This approach has limitations in handling anomalous nodes, as it tends to overly smooth out the significant feature differences between anomalous nodes and their neighbors during the message passing process, resulting in a loss of anomalous characteristics. To overcome this limitation, we propose a novel adaptive multi-information fusion model. In this model, we introduce a graph deviation network layer specifically designed to capture the differential information between nodes, enhancing the identification of anomalous behaviors. By adaptively fusing multiple sources of information, the model can better represent nodes. As demonstrated by the experimental results on two widely used real-world datasets, our model showcases a significant performance improvement compared to several state-of-the-art baseline methods.