OR-Gate Mixup Multiscale Spectral Graph Neural Network for Node Anomaly Detection.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph neural networks (GNNs) are widely used for node anomaly detection.
• Existing GNNs often act as low-pass filters, suppressing crucial high-frequency signals and leading to over-smoothing.
• This limitation hinders the distinction between normal and anomalous nodes and can negatively impact data augmentation.

Purpose of the Study:

• To address the limitations of existing GNNs in node anomaly detection.
• To develop a GNN architecture capable of effectively mining high-frequency graph signals.
• To improve model generalization and reduce computational cost in graph anomaly detection.

Main Methods:

• Proposed a multiscale spectral GNN (MMGNN) utilizing a double-parallel structure.
• Designed multiorder multiscale bandpass filters by superposing polynomial spectral filters.
• Introduced or-gate mixup for spectral space data augmentation.

Main Results:

• MMGNN effectively mines high-frequency signals in graph data.
• The proposed method reduces computational cost compared to traditional serial GNN structures.
• Experimental results on four real-world datasets show MMGNN outperforms state-of-the-art methods.

Conclusions:

• MMGNN offers an effective approach to node anomaly detection by leveraging high-frequency graph signals.
• The spectral filtering and data augmentation strategies enhance detection performance and model generalization.
• MMGNN provides a promising direction for advancing graph anomaly detection techniques.