Reproduce “Towards Self-Interpretable Graph-Level Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yixin Liu, Kaize Ding, Qinghua Lu, Fuyi Li, Leo Yu Zhang, Shirui Pan

A new framework for self-interpretable, unsupervised graph-level anomaly detection identifies anomalous graphs and explains their anomalous nature through vital subgraphs. The proposed SIGNET model demonstrated superior performance in both anomaly detection (average rank 1.7 across 16 datasets) and explanation quality (27.89% higher NX-AUC, 8.99% higher EX-AUC) compared to baselines.

Key insights

  • The Information Bottleneck principle can be effectively adapted into a multi-view setting to enable self-interpretable graph-level anomaly detection without explicit supervision.
  • Dual Hypergraph Transformation provides a robust and deterministic method for generating two distinct yet structurally related graph views, crucial for capturing comprehensive anomaly patterns.
  • A unified subgraph extractor that consistently derives explanations across different graph views significantly enhances both the quality of anomaly detection and the interpretability of results.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.