Official demo code for the paper "Cluster-Aware Graph Anomaly Detection"
This repository provides the implementation and example experiments for reproducing results on benchmark datasets such as Amazon, BlogCatalog, IMDB, and DBLP.

CARE (Cluster-Aware Graph Anomaly Detection) is designed to detect anomalous nodes in graphs by leveraging both structural and cluster-level semantic information. The method jointly models local neighborhood consistency and global cluster awareness to effectively identify anomalies across diverse graph datasets by augmenting the graph adjacency matrix with pseudo-labels derived from soft cluster memberships. To mitigate potential bias from these pseudo-labels, CARE introduces a similarity-guided loss, theoretically shown to be a variant of contrastive learning loss connected to spectral clustering.

CARE Overview

Before running the demo, please ensure you have the following packages installed:

• PyTorch
• scikit-learn
• scipy

Install all dependencies with:

  pip install torch scikit-learn scipy

Before running any experiments, please unzip the dataset archive so the code can access the data files:

  unzip data.zip

This will create a data/ directory (or add files into it) containing the datasets used by the demo, including:

• Amazon
• BlogCatalog
• imdb
• dblp

Make sure the directory structure matches what main.py expects (e.g., data/Amazon/...).

To reproduce the experimental results on each dataset, run:

Each command will run training and evaluation for CARE on the chosen dataset and should save results (check your results/ or logs/ directory if present).

CARE demonstrates strong performance on benchmark datasets by capturing both cluster-level and structural anomalies.

🧩 Results on Multi-View Graphs (CERT, IMDB, DBLP)

🧮 Results on Single-View Graphs (BlogCatalog, Amazon, YelpChi)

If you use this code or ideas from the paper, please cite:

@inproceedings{DBLP:conf/www/ZhengBWZH25,
  author       = {Lecheng Zheng and
                  John R. Birge and
                  Haiyue Wu and
                  Yifang Zhang and
                  Jingrui He},
  editor       = {Guodong Long and
                  Michale Blumestein and
                  Yi Chang and
                  Liane Lewin{-}Eytan and
                  Zi Helen Huang and
                  Elad Yom{-}Tov},
  title        = {Cluster Aware Graph Anomaly Detection},
  booktitle    = {Proceedings of the {ACM} on Web Conference 2025, {WWW} 2025, Sydney,
                  NSW, Australia, 28 April 2025- 2 May 2025},
  pages        = {1771--1782},
  publisher    = {{ACM}},
  year         = {2025},
}