Anomaly Detection on Attributed Networks via Contrastive Self-Supervised Learning.

Area of Science:

• Computer Science
• Machine Learning
• Network Analysis

Background:

• Anomaly detection on attributed networks is crucial for understanding complex systems.
• Deep learning methods show promise but existing graph autoencoder approaches have limitations.
• Current methods struggle with suboptimal performance, scalability, and lack of direct anomaly detection objectives.

Purpose of the Study:

• To propose a novel Contrastive self-supervised Learning framework for Anomaly detection on attributed networks (CoLA).
• To overcome limitations of existing methods by fully exploiting network information and enabling scalable training.
• To develop an anomaly detection-aware learning objective for improved performance.

Main Methods:

• CoLA samples novel contrastive instance pairs to capture node-substructure relationships unsupervised.
• A graph neural network (GNN)-based contrastive learning model learns embeddings from attributes and local structure.
• Anomaly scores are derived from multiround predictions and statistical estimation, enabling anomaly detection-aware training.

Main Results:

• The proposed CoLA framework fully exploits local network information.
• CoLA effectively learns informative embeddings from high-dimensional attributes and local structure.
• The framework demonstrates scalability to large networks by processing instance pairs in batches.

Conclusions:

• CoLA significantly outperforms state-of-the-art baseline methods on seven benchmark datasets.
• The novel contrastive instance sampling and GNN-based contrastive learning contribute to superior anomaly detection.
• CoLA offers a scalable and effective solution for anomaly detection in attributed networks.