🌐 CARE: Cluster-Aware Graph Anomaly Detection

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.

---

🧠 Overview

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

⚙️ Requirements

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

---

📦 Dataset Preparation

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/...).

---

🚀 Run the Demo

To reproduce the experimental results on each dataset, run:

Dataset	Command
Amazon	python main.py --dataset Amazon
BlogCatalog	python main.py --dataset BlogCatalog
IMDB	python main.py --dataset imdb
DBLP	python main.py --dataset dblp

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).

---

📊 Experimental Results

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

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

Method	CERT AUPRC	CERT AUROC	IMDB AUPRC	IMDB AUROC	DBLP AUPRC	DBLP AUROC
MLRA	0.0379 +/- 0.001	0.3829 +/- 0.003	0.2695 +/- 0.007	0.5926 +/- 0.005	0.2211 +/- 0.005	0.5568 +/- 0.005
NSNMF	0.0704 +/- 0.001	0.4578 +/- 0.001	0.0634 +/- 0.000	0.4969 +/- 0.001	0.1436 +/- 0.007	0.6418 +/- 0.001
NCMOD	0.0749 +/- 0.001	0.5133 +/- 0.001	0.6629 +/- 0.013	0.8030 +/- 0.007	0.4809 +/- 0.006	0.7271 +/- 0.004
SRSLP	0.0806 +/- 0.007	0.5405 +/- 0.003	0.5552 +/- 0.017	0.7343 +/- 0.003	0.0643 +/- 0.002	0.5228 +/- 0.001
TAM	0.0771 +/- 0.007	0.5400 +/- 0.005	0.6521 +/- 0.016	0.8233 +/- 0.013	0.3466 +/- 0.016	0.6690 +/- 0.005
CARE	0.1198 +/- 0.003	0.6056 +/- 0.001	0.8968 +/- 0.038	0.9370 +/- 0.029	0.8495 +/- 0.005	0.8696 +/- 0.006

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

Method	BlogCatalog AUPRC	BlogCatalog AUROC	Amazon AUPRC	Amazon AUROC	YelpChi AUPRC	YelpChi AUROC
ANOMALOUS	0.0652 +/- 0.005	0.5652 +/- 0.025	0.0558 +/- 0.001	0.4457 +/- 0.005	0.0519 +/- 0.002	0.4956 +/- 0.003
Dominant	0.3102 +/- 0.011	0.7590 +/- 0.010	0.1424 +/- 0.002	0.5996 +/- 0.002	0.0395 +/- 0.020	0.4133 +/- 0.100
CoLA	0.3270 +/- 0.000	0.7746 +/- 0.009	0.0677 +/- 0.001	0.5898 +/- 0.011	0.0448 +/- 0.002	0.4636 +/- 0.001
SLGAD	0.3882 +/- 0.007	0.8123 +/- 0.002	0.0634 +/- 0.005	0.5937 +/- 0.005	0.0350 +/- 0.000	0.3312 +/- 0.035
HCM-A	0.3139 +/- 0.001	0.7980 +/- 0.004	0.0527 +/- 0.015	0.3956 +/- 0.014	0.0287 +/- 0.012	0.4593 +/- 0.005
ComGA	0.3293 +/- 0.028	0.7683 +/- 0.004	0.1153 +/- 0.005	0.5895 +/- 0.010	0.0423 +/- 0.000	0.4391 +/- 0.000
CONAD	0.3284 +/- 0.004	0.7807 +/- 0.003	0.1372 +/- 0.009	0.6142 +/- 0.008	0.0405 +/- 0.002	0.4588 +/- 0.003
TAM	0.4182 +/- 0.225	0.8248 +/- 0.003	0.2634 +/- 0.008	0.7064 +/- 0.008	0.0778 +/- 0.009	0.5643 +/- 0.007
CARE	0.4043 +/- 0.010	0.8194 +/- 0.003	0.6563 +/- 0.011	0.8656 +/- 0.002	0.1218 +/- 0.003	0.7516 +/- 0.003

---

🧩 Citation

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},
}