ENDASh: Embedding Neighbourhood Dissimilarity with Attribute Shuffling for Graph Anomaly Detection

Recent unsupervised GNN based graph anomaly detection (GAD) methods adopt specific mechanisms designed for anomaly detection. This is in contrast to earlier methods that utilise components such as graph autoencoders that were designed for more general use-cases. However, these newer methods only lead to a modest increase in detection accuracy at the cost of complicated optimisation schemes and higher runtimes. To overcome these issues, we propose Embedding Neighbourhood Dissimilarity (END) with Attribute Shuffling (ENDASh), a simple but scalable and effective GAD framework. ENDASh utilises our proposed END measure to quantify the degree of abnormality of nodes using GraphSAGE embeddings that were optimised with Attribute Shuffling, a data augmentation method designed to project anomalies away from inliers in the latent space. Extensive experiments on real-world benchmarks demonstrate the competitive GAD performance of our ENDASh while being computationally efficient and capable of operating in an inductive environment.