Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection (NeurIPS'23)

This repository is the official implementation of NeurIPS'23 paper "Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection"

The full paper can be found at NeurIPS Portal or arXiv.

📋 We explore the property one class homophily to introduce a novel unsupervised anomaly scoring measure for GAD -- local node affinity -- that assigns a larger anomaly score to nodes that are less affiliated with their neighbors, with the affinity defined as similarity on node attributes/representations. We further propose Truncated Affinity Maximization (TAM) that learns tailored node representations for our anomaly measure by maximizing the local affinity of nodes to their neighbors. TAM is instead optimized on truncated graphs where non-homophily edges are removed iteratively to mitigate this bias. Extensive empirical results on six real-world GAD datasets show that TAM substantially outperforms seven competing models

Requirements

To install requirements:

pip install -r requirements.txt

📋 TAM is implemented in Pytorch 1.6.0 with Python 3.7 and all the experiments are run on an NVIDIA GeForce RTX 3090 24GB GPU. For large-scale graph datasets, like T-Finance and YelpChi-all, the experiments are run on NVIDIA A100 GPU for quickness. The experiments on OGB-Protein are run on the CPU.

Datasets

BlogCatalog and ACM were downloaded from https://github.com/yixinliu233/CoLA
Amazon and Yelpchi were downloaded from https://github.com/YingtongDou/CARE-GNN
Amazon-all and Yelpchi-all were downloaded from https://github.com/YingtongDou/CARE-GNN
Facebook is obtained from https://github.com/zhiming-xu/conad
Reddit is downloaded from https://github.com/pygod-team/data
T-finance is downloaded from https://drive.google.com/drive/folders/1PpNwvZx_YRSCDiHaBUmRIS3x1rZR7fMr
OGB-Protein is downloaded from https://ogb.stanford.edu/docs/nodeprop/#ogbn-arxiv
For convenience, all the datasets can be obtained from https://drive.google.com/drive/folders/1qcDBcVdcfAr_q5VOXBYagtnhA_r3Mm3Z?usp=drive_link

Training

To train the model(s) in the paper, run this command:

python train.py

📋 In TAM, each LAMNet is implemented by a two-layer GCN, and its weight parameters are optimized using Adam optimizer with 500 epochs and a learning rate of 1e-5 by default. T=3 and K=4 are used for all datasets. Datasets with injected anomalies, such as BlogCatalog, ACM and Facebook, require strong regularization, so $\lambda=1$ is used by default; whereas $\lambda=0$ is used for the four real-world datasets. For the larger datasets like Amazon-all, YelpChi-all, and OGB-Protein, they require larger truncation times due to the large number of edges. So we set K = 7.

Evaluation

To evaluate our model on datasets, run:

python train.py 

Dataset

Dataset	Nodes	Edges	Attributes	Anomalies(Rate)
BlogCatalog	5,196	171,743	8,189	300(5.77%)
ACM	16,484	71,980	8,337	597(3.63%)
Amazon(UPU)	10244	175,608	25	693(6.66%)
Facebook	1,081	55,104	576	27(2.49%)
Reddit	10,984	168,016	64	366(3.33%)
YelpChi (RUR)	24,741	49,315	32	1,217(4.91%)
Amazon-all	11,944	4,398,392	25	821(6.87%)
YelpChi-all	45,941	3,846,979	32	6,674(14.52%)
T-Finance	39,357	21,222,543	10	1,803 (4.58%)
OGB-Protein	132,534	39,561,252	8	6000(4.5%)

Results

Our model achieves the following performance on :

Metric	BlogCatalog	ACM	Amazon	Facebook	Reddit	YelpChi	Amazon-all	YelpChi-all	T-finance	OGB-Protein
AUROC	0.8248	0.8878	0.7064	0.9144	0.6023	0.5643	0.8476	0.5818	0.6175	0.7449
AUPRC	0.4182	0.5124	0.2634	0.2233	0.0446	0.0778	0.4346	0.1886	0.0547	0.2173

Main Contributions

We reveal an important anomaly-discriminative property, the one-class homophily, in GAD datasets with either injected or real anomalies. We utilize this property to introduce a novel unsupervised GAD measure, local node affinity, and further introduce a truncated affinity maximization (TAM) approach that end-to-end optimizes the proposed anomaly measure on truncated adjacency matrix with the non-homophily edges eliminated

Citation

If you use this package and find it useful, please cite our paper using the following BibTeX. Thanks! :)

@inproceedings{qiao2023truncated,
  title={Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection},
  author={Qiao, Hezhe and Pang, Guansong},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}