EAGLE: Contrastive Learning for Efficient Graph Anomaly Detection

EAGLE, an efficient graph anomaly detection method, distinguishes abnormal nodes from normal ones using distances to local context. It employs graph autoencoder for unsupervised learning of informative node embeddings, combined with a discriminator to predict anomaly scores. EAGLE surpasses state-of-the-art methods on three heterogeneous network datasets. Graph anomaly detection models, particularly graph neural networks, have advanced, reducing human expertise reliance. Liu et al. applied a graph convolution network to account-device graphs, embedding structure and node attributes. Wang et al. proposed a framework for defending Water Treatment Networks against cyber attacks, while Zhang et al. developed a fraud detection model based on GNNs addressing graph inconsistencies. EAGLE, introduced in the paper, uses contrastive self-supervised learning to generate positive and negative instances at the meta path-level, preserving network semantics. It captures structure and attribute information, distinguishing between anomalous and normal nodes. EAGLE outperforms baseline methods on real-world graphs.