GraphGLA: Enhancing Graph Neural Networks with Generative Local Feature Augmentation for Anomaly Detection

Graph Neural Networks (GNNs) are commonly applied in graph anomaly detection. Despite their success in mining structured data, traditional GNNs often encounter issues like over-smoothing and information loss. when dealing with sparse and heterogeneous neighborhoods, limiting their effectiveness in anomaly detection. To tackle these challenges, we introduce a new framework, GraphGLA which improves node representation by enhancing homogeneous neighbor features. Specifically, we design a node discriminator to classify neighbors as homogeneous or heterogeneous, and then leverage a Conditional Variational Autoencoder (CVAE)<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]</sup> to generate additional homogeneous neighbor features. The generated features are subsequently merged with the original feature set and used for anomaly detection via a Graph Attention Network (GAT). Experiments on three public anomaly detection datasets show that GraphGLA greatly outperforms existing methods. This approach effectively mitigates the impact of sparse and heterogeneous neighborhoods on anomaly detection, providing robust and accurate detection capabilities.