Semi-supervised Graph Edge Convolutional Network for Anomaly Detection

In recent years, with deep learning development, graph-based deep anomaly detection has attracted more and more researchers' attention due to graph data's strong expression ability. However, at present, graph-based methods mainly focus on node-level anomaly detection, while edge-level anomaly detection is relatively minor. Anomaly detection at the edge level can distinguish the specific edges connected to nodes as detection objects, so its resolution granularity is more detailed than that of the node-based method. Second, the rules of anomalies are challenging to learn. At present, most of the algorithms adopt the unsupervised method to train the model. As a result, the detected result is likely to be noise data. In this paper, we propose a Graph Edge Anomaly Detection model based on a Semi-supervised auto-encoder (GEADS). In this model, we first adjust the traditional mini-batch training strategy to train the model on a large-scale graph. It improves the scalability of the model. Second, we design an edge convolutional neural network layer to realize the fusion of edge neighborhood information. We take the reconstruction error as the evaluation criterion after stacking multiple edge convolutional neural network layers that encode and decode the edges. Third, the few abnormal samples with known labels are utilized to guide the model's parameter optimization process. While ensuring the generalization ability of the model, it also improves the pertinence to specific anomalies. Finally, we show the effectiveness of the proposed algorithm through experiments on two real-world datasets.