AntiFraud

A unified research framework for financial fraud detection. Implementations of representative deep learning and graph-based fraud detection methods, with unified data processing, training, and evaluation pipelines.

🔬 Methods · 🚀 Usage · 🗃️ Datasets · 📊 Results · 🗂️ Repository · 📚 Citation

🔬 Implemented Methods

AntiFraud provides implementations of representative financial fraud detection methods spanning convolutional, spatio-temporal, graph-based, and graph augmentation approaches.

• MCNN: Credit card fraud detection using convolutional neural networks in ICONIP 2016.
• STAN: Spatio-temporal attention-based neural network for credit card fraud detection in AAAI 2020.
• STAGN: Graph Neural Network for Fraud Detection via Spatial-temporal Attention in TKDE 2020
• GTAN: Semi-supervised Credit Card Fraud Detection via Attribute-driven Graph Representation in AAAI 2023.
• RGTAN: Enhancing Attribute-driven Fraud Detection with Risk-aware Graph Representation in TKDE 2025
• HOGRL: Effective High-order Graph Representation Learning for Credit Card Fraud Detection in IJCAI 2024.
• Grad: Grad: Guided Relation Diffusion Generation for Graph Augmentation in Graph Fraud Detection in WWW 2025.

🚀 Usage

📦 Requirements

python           3.7
scikit-learn     1.0.2
pandas           1.3.5
numpy            1.21.6
networkx         2.6.3
scipy            1.7.3
torch            1.12.1+cu113
dgl-cu113        0.8.1

📥 Data processing

1. Run unzip /data/Amazon.zip and unzip /data/YelpChi.zip to unzip the datasets;
2. Run python feature_engineering/data_process.py to pre-process all datasets needed in this repo.
3. Run python feature_engineering/get_matrix.py to generate the adjacency matrix of the high-order transaction graph.Please note that this will require approximately 280GB of storage space. Please be aware that if you intend to run HOGRL , you should first execute the get_matrix.py script.

Grad is run on the DGL's Amazon and Yelp dataset.

🧠 Training & Evalutaion

To test implementations of MCNN, STAN and STAGN, run

python main.py --method mcnn
python main.py --method stan
python main.py --method stagn

Configuration files can be found in config/mcnn_cfg.yaml, config/stan_cfg.yaml and config/stagn_cfg.yaml, respectively.

Models in GTAN and RGTAN can be run via:

python main.py --method gtan
python main.py --method rgtan

For specification of hyperparameters, please refer to config/gtan_cfg.yaml and config/rgtan_cfg.yaml.

Model in HOGRL can be run via:

python main.py --method hogrl

For specification of hyperparameters, please refer to config/hogrl_cfg.yaml.

To run the Grad, please read the README of Grad, including the environments and pipelines.

🗃️ Datasets

There are three datasets, YelpChi, Amazon and S-FFSD, utilized for model experiments in this repository.

YelpChi and Amazon datasets are from CARE-GNN, whose original source data can be found in this repository.

S-FFSD is a simulated & small version of finacial fraud semi-supervised dataset. Description of S-FFSD are listed as follows:

Name	Type	Range	Note
Time	np.int32	from $\mathbf{0}$ to $\mathbf{N}$	$\mathbf{N}$ denotes the number of trasactions.
Source	string	from $\mathbf{S_0}$ to $\mathbf{S}_{ns}$	$ns$ denotes the number of transaction senders.
Target	string	from $\mathbf{T_0}$ to $\mathbf{T}_{nt}$	$nt$ denotes the number of transaction reveicers.
Amount	np.float32	from 0.00 to np.inf	The amount of each transaction.
Location	string	from $\mathbf{L_0}$ to $\mathbf{L}_{nl}$	$nl$ denotes the number of transacation locations.
Type	string	from $\mathbf{TP_0}$ to $\mathbf{TP}_{np}$	$np$ denotes the number of different transaction types.
Labels	np.int32	from 0 to 2	2 denotes unlabeled

We are looking for interesting public datasets! If you have any suggestions, please let us know!

📊 Results

The performance of five models tested on three datasets are listed as follows:

	YelpChi			Amazon			S-FFSD
	AUC	F1	AP	AUC	F1	AP	AUC	F1	AP
MCNN		-	-	-	-	-	0.7129	0.6861	0.3309
STAN	-	-	-	-	-	-	0.7446	0.6791	0.3395
STAGN	-	-	-	-	-	-	0.7659	0.6852	0.3599
GTAN	0.9241	0.7988	0.7513	0.9630	0.9213	0.8838	0.8286	0.7336	0.6585
RGTAN	0.9498	0.8492	0.8241	0.9750	0.9200	0.8926	0.8461	0.7513	0.6939
HOGRL	0.9808	0.8595	-	0.9800	0.9198	-	-	-	-
Grad	0.9908	-	0.9644	0.9789	-	0.8968	-	-	-

MCNN, STAN and STAGN are presently not applicable to YelpChi and Amazon datasets.

HOGRL and Grad are presently not applicable to S-FFSD dataset.

🗂️ Repository

The repository is organized as follows:

AntiFraud/
├── config/               # Configuration files for different models
├── data/                 # Dataset files
├── feature_engineering/  # Data preprocessing and graph construction
├── methods/              # Implementations of fraud detection methods
├── models/               # Pre-trained models for each method
├── main.py               # Main entry point for model execution
└── requirements.txt      # Package dependencies

🤝 Contributors :

📚 Citation

If you find Antifraud is useful for your research, please consider citing the following papers:

@inproceedings{yang2025grad,
  title={Grad: Guided relation diffusion generation for graph augmentation in graph fraud detection},
  author={Yang, Jie and Zhang, Rui and Cheng, Ziyang and Cheng, Dawei and Yang, Guang and Wang, Bo},
  booktitle={Proceedings of the ACM on Web Conference 2025},
  pages={5308--5319},
  year={2025}
}
@inproceedings{zou2024effective,
  title={Effective High-order Graph Representation Learning for Credit Card Fraud Detection.},
  author={Zou, Yao and Cheng, Dawei},
  booktitle={International Joint Conference on Artificial Intelligence},
  year={2024}
}
@ARTICLE{xiang2025enhancing,
  author={Xiang, Sheng and Zhang, Guibin and Cheng, Dawei and Zhang, Ying},
  journal={IEEE Transactions on Knowledge and Data Engineering}, 
  title={Enhancing Attribute-Driven Fraud Detection With Risk-Aware Graph Representation}, 
  year={2025},
  pages={1-12},
  doi={10.1109/TKDE.2025.3543887}
}
@inproceedings{Xiang2023SemiSupervisedCC,
    title={Semi-supervised Credit Card Fraud Detection via Attribute-driven Graph Representation},
    author={Sheng Xiang and Mingzhi Zhu and Dawei Cheng and Enxia Li and Ruihui Zhao and Yi Ouyang and Ling Chen and Yefeng Zheng},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
    year={2023}
}
@article{cheng2020graph,
    title={Graph Neural Network for Fraud Detection via Spatial-temporal Attention},
    author={Cheng, Dawei and Wang, Xiaoyang and Zhang, Ying and Zhang, Liqing},
    journal={IEEE Transactions on Knowledge and Data Engineering},
    year={2020},
    publisher={IEEE}
}
@inproceedings{cheng2020spatio,
    title={Spatio-temporal attention-based neural network for credit card fraud detection},
    author={Cheng, Dawei and Xiang, Sheng and Shang, Chencheng and Zhang, Yiyi and Yang, Fangzhou and Zhang, Liqing},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
    volume={34},
    number={01},
    pages={362--369},
    year={2020}
}
@inproceedings{fu2016credit,
    title={Credit card fraud detection using convolutional neural networks},
    author={Fu, Kang and Cheng, Dawei and Tu, Yi and Zhang, Liqing},
    booktitle={International Conference on Neural Information Processing},
    pages={483--490},
    year={2016},
    organization={Springer}
}