A critical aspect of graph neural networks (GNNs) is to enhance the node representations by aggregating node neighborhood information. However, when detecting anomalies, the representations of abnormal nodes are prone to be averaged by normal neighbors, making the learned anomaly representations less distinguishable. To tackle this issue, we propose an unsupervised counterfactual data augmentation method for graph anomaly detection (CAGAD) that introduces a graph pointer neural network as the heterophilic node detector to identify potential anomalies whose neighborhoods are normal-node-dominant. For each identified potential anomaly, we design a graph-specific diffusion model to translate a part of its neighbors, which are probably normal, into anomalous ones. At last, we involve these translated neighbors in GNN neighborhood aggregation to produce counterfactual representations of anomalies. Through aggregating the translated anomalous neighbors, counterfactual representations become more distinguishable and further advocate detection performance. The experimental results on four datasets demonstrate that CAGAD significantly outperforms strong baselines, with an average improvement of 2.35% on F1, 2.53% on AUC-ROC, and 2.79% on AUC-PR.

Graph anomaly detection is crucial in many high-impact applications across diverse fields. In anomaly detection tasks, collecting plenty of annotated data tends to be costly and laborious. As a result, few-shot learning has been explored to address the issue by requiring only a few labeled samples to achieve good performance. However, conventional few-shot models may not fully exploit the information within auxiliary sets, leading to suboptimal performance. To tackle these limitations, we propose a dual-level knowledge distillation-based approach for graph anomaly detection, DualKD, which leverages two distinct distillation losses to improve generalization capabilities. In our approach, we initially train a teacher model to generate prediction distributions as soft labels, capturing the entropy of uncertainty in the data. These soft labels are then employed to construct the corresponding loss for training a student model, which can capture more detailed node features. In addition, we introduce two representation distillation losses—short and long representation distillation—to effectively transfer knowledge from the auxiliary set to the target set. Comprehensive experiments conducted on four datasets verify that DualKD remarkably outperforms the advanced baselines, highlighting its effectiveness in enhancing identification performance.