GADAL: An Active Learning Framework for Graph Anomaly Detection

Abstract

Graph Neural Networks (GNNs) have been widely used in graph-based anomaly detection tasks, and these methods require a sufficient amount of labeled data to achieve satisfactory performance. However, the high cost for data annotation leads to some well-designed algorithms in low practicality in real-world tasks. Active learning has been used to find a trade-off between labeling cost and model performance, while few prior works take it into anomaly detection. Therefore, we propose GADAL, a novel Active Learning framework for Graph Anomaly Detection, which employs a multi-aspects query strategy to achieve high performance within a limited budget. First, we design an abnormal-aware query strategy based on the scalable sliding window to enrich abnormal patterns and alleviate the class imbalance problem. Second, we design an inconsistency-aware query strategy based on the effective degree to capture the most specificity nodes in information aggregation. Then we provide a hybrid solution for the above query strategies. Empirical studies demonstrate that our query strategy significantly outperforms other strategies, and GADAL achieves a comparable performance to the state-of-art anomaly detection methods within less than 3% of the budget.