ANIMATE: Unsupervised Attributed Graph Anomaly Detection with Masked Graph Transformers

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Attributed graphs are crucial for representing real-world sensor data.
• Unsupervised graph anomaly detection (UGAD) identifies abnormal nodes without labels, vital for IoT and sensor reliability.
• Traditional Graph Neural Networks (GNNs) face limitations like local aggregation and over-smoothing, hindering anomaly detection.

Purpose of the Study:

• To introduce a novel unsupervised attributed graph anomaly detection method.
• To overcome the limitations of traditional GNNs in capturing global graph structures.
• To enhance the detection of anomalies in class-imbalanced datasets.

Main Methods:

• Proposed unsupervised attributed graph Anomaly detectioN wIth Masked grAph TransformErs (ANIMATE).
• Utilized Graph Transformers (GTs) for a global receptive field to capture distinguishable abnormal characteristics.
• Employed masked auto-encoders for node feature reconstruction, focusing the model on normal patterns.
• Implemented a self-paced enhancement scheme tailored for UGAD tasks.

Main Results:

• ANIMATE demonstrated effectiveness on real-world benchmark datasets with organic anomalies.
• The method outperformed state-of-the-art competitors in unsupervised graph anomaly detection.
• Global perspective from Graph Transformers improved discrimination capacity compared to traditional GNNs.

Conclusions:

• ANIMATE offers a robust solution for unsupervised attributed graph anomaly detection.
• The integration of Graph Transformers and masked auto-encoders enhances anomaly identification.
• The proposed method contributes to improving the reliability of intelligent sensor systems.