Applying Data Augmentation Method to Improve Anomaly Detection with Graph Neural Network

Tóm tắt

While deep learning on graphs has gradually become one of the mainstream for mining real-world graph data, detecting anomaly nodes still remains a challenging issue. The rooting cause is the natural over-smoothing and over-squashing of graph deep learning which becomes more severe in graph anomaly detection tasks. Beta-Wavelet filters are introduced to combat this issue, however, their performance is still limited by both the under-reaching and sparse training data issues. Nodes that are topologically distant from labeled nodes suffer from a lack of supervision, as the influence of labeled nodes diminishes with increasing graph distance. In this work, we first analyze the under-reaching issue of Beta-Wavelet Graph Neural Networks under the context of anomaly node classification in a semi-supervised learning paradigm. This problem arises from the limited propagation of supervision signals from labeled nodes, particularly to distant nodes, which restricts the model’s ability to capture relevant features across the graph. The challenge is further exacerbated in anomaly detection tasks, where anomalies are sparsely distributed and often surrounded by a majority of normal nodes, making their identification more difficult. Based on these insights, we propose Beta-Wavelet Mixup, a simple and flexible data augmentation approach tailored for anomaly detection. Our method effectively utilizes latent domain information to identify potential anomalies, generate synthetic nodes, and establish connections on them. The introduced mixup operator is not only to solve the sparseness in the training nodes but also allows more signal propagation by adding mixing edges. Our findings highlight the need for novel strategies to extend the effective receptive field of BWGNNs while preserving the discriminative power of node embeddings, motivating the development of our proposed method. To evaluate our approach, we experiment on five popular graph anomaly detection datasets where our method outperforms the other baselines significantly