Normality Calibration in Semi-supervised Graph Anomaly Detection

---

Overview

We propose GraphNC, a graph normality calibration framework that leverages both labeled and unlabeled data to calibrate the normality from a teacher, namely a pre-trained semi-supervised GAD model, jointly in anomaly score and representation spaces. GraphNC includes two main components: anomaly score distribution alignment (ScoreDA) and perturbation-based normality regularization (NormReg). ScoreDA optimizes the anomaly scores of our model by aligning them with the score distribution yielded by the teacher. Since the teacher provides accurate scores for most normal nodes and a subset of anomaly nodes, this alignment effectively pulls the anomaly scores of the two classes toward opposite ends, resulting in more separable anomaly scores. To mitigate the misleading effects of inaccurate teacher scores, NormReg is designed to regularize normality in the representation space. Specifically, it encourages more compact representations of normal nodes by minimizing a perturbation-guided consistency loss solely on labeled nodes.

Environment Settings

This implementation is based on Python 3. The main dependencies are:

• torch==1.13.0
• torch-geometric==2.1.0
• torch-scatter==2.1.0+pt113cu117
• torch-sparse==0.6.15+pt113cu117
• dgl==0.5.3
• numpy==1.21.6
• scipy==1.7.3
• scikit-learn==1.0.2
• networkx==2.6.3
• seaborn==0.12.2
• tqdm==4.65.0

Example installation:

conda create -n graphnc python=3.7 -y
conda activate graphnc

pip install torch==1.13.0 --extra-index-url https://download.pytorch.org/whl/cu117
pip install torch-geometric==2.1.0
pip install torch-scatter==2.1.0+pt113cu117 torch-sparse==0.6.15+pt113cu117 \
  -f https://data.pyg.org/whl/torch-1.13.0+cu117.html
pip install dgl==0.5.3
pip install numpy==1.21.6 scipy==1.7.3 pandas==1.3.5 scikit-learn==1.0.2 \
  networkx==2.6.3 matplotlib==3.5.3 seaborn==0.12.2 tqdm==4.65.0

Alternatively, install from the provided package list:

pip install -r requirements.txt

Datasets

For convenience, some datasets can be obtained from google drive link. We sincerely thank the researchers for providing these datasets.

Put the .mat datasets under:

dataset/

Example:

dataset/photo.mat
dataset/tolokers.mat

Pretrained Teacher Models

This repository includes pretrained GGAD teacher checkpoints:

pretrained_model/Photo_GGAD_PreTrainModel.pth
pretrained_model/Tolokers_GGAD_PreTrainModel.pth

Use the corresponding checkpoint path with --teacher_path.

Run Experiments

Example for Photo:

python main.py --dataset photo --teacher_path pretrained_model/Photo_GGAD_PreTrainModel.pth

Example for Tolokers:

python main.py --dataset tolokers --teacher_path pretrained_model/Tolokers_GGAD_PreTrainModel.pth

Training logs are saved to:

ggad_labeledNormal/{dataset}.txt

Main Results

Project Structure

GraphNC/
|-- main.py                  # training and evaluation entry point
|-- model.py                 # GGAD teacher and OCGNN student models
|-- utils.py                 # dataset loading and preprocessing utilities
|-- requirements.txt         # Python dependencies
|-- pretrained_model/        # pretrained teacher checkpoints
|-- GraphNC logo.png
`-- framework.png

📖 Citation

If you find this work useful, please cite our paper:

@article{zeng2025normality,
  title={Normality Calibration in Semi-supervised Graph Anomaly Detection},
  author={Zeng, Guolei and Qiao, Hezhe and Ai, Guoguo and Guo, Jinsong and Pang, Guansong},
  journal={arXiv preprint arXiv:2510.02014},
  year={2025}
}