AntiFraud

A Financial Fraud Detection Framework.

Source codes implementation of papers:

• 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 AAAI2020
• STAGN: Graph Neural Network for Fraud Detection via Spatial-temporal Attention, in TKDE2020
• GTAN: Semi-supervised Credit Card Fraud Detection via Attribute-driven Graph Representation, in AAAI2023
• RGTAN: Enhancing Attribute-driven Fraud Detection with Risk-aware Graph Representation,
• HOGRL: Effective High-order Graph Representation Learning for Credit Card Fraud Detection,

Usage

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.

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.

Data Description

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!

Test Result

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	-	-	-	-

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

HOGRL is presently not applicable to S-FFSD dataset.

Repo Structure

The repository is organized as follows:

• models/: the pre-trained models for each method. The readers could either train the models by themselves or directly use our pre-trained models;
• data/: dataset files;
• config/: configuration files for different models;
• feature_engineering/: data processing;
• methods/: implementations of models;
• main.py: organize all models;
• requirements.txt: package dependencies;

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

Contributors :

Citing

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

@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}
}
@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}
}