A Deep Graph-based Toolbox for Fraud Detection

Introduction

May 2021 Update: The DGFraud has upgraded to TensorFlow 2.0! Please check out DGFraud-TF2

DGFraud is a Graph Neural Network (GNN) based toolbox for fraud detection. It integrates the implementation & comparison of state-of-the-art GNN-based fraud detection models. The introduction of implemented models can be found here.

We welcome contributions on adding new fraud detectors and extending the features of the toolbox. Some of the planned features are listed in TODO list.

If you use the toolbox in your project, please cite one of the two papers below and the algorithms you used :

CIKM'20 (PDF)

@inproceedings{dou2020enhancing,
  title={Enhancing Graph Neural Network-based Fraud Detectors against Camouflaged Fraudsters},
  author={Dou, Yingtong and Liu, Zhiwei and Sun, Li and Deng, Yutong and Peng, Hao and Yu, Philip S},
  booktitle={Proceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM'20)},
  year={2020}
}

SIGIR'20 (PDF)

@inproceedings{liu2020alleviating,
  title={Alleviating the Inconsistency Problem of Applying Graph Neural Network to Fraud Detection},
  author={Liu, Zhiwei and Dou, Yingtong and Yu, Philip S. and Deng, Yutong and Peng, Hao},
  booktitle={Proceedings of the 43nd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2020}
}

Useful Resources

• PyGOD: A Python Library for Graph Outlier Detection (Anomaly Detection)
• UGFraud: An Unsupervised Graph-based Toolbox for Fraud Detection
• Graph-based Fraud Detection Paper List
• Awesome Fraud Detection Papers
• Attack and Defense Papers on Graph Data
• PyOD: A Python Toolbox for Scalable Outlier Detection (Anomaly Detection)
• PyODD: An End-to-end Outlier Detection System
• DGL: Deep Graph Library
• Outlier Detection DataSets (ODDS)

Table of Contents

• Installation
• Datasets
• User Guide
• Implemented Models
• Model Comparison
• TODO List
• How to Contribute

Installation

git clone https://github.com/safe-graph/DGFraud.git
cd DGFraud
python setup.py install

Requirements

* python 3.6, 3.7
* tensorflow>=1.14.0,<2.0
* numpy>=1.16.4
* scipy>=1.2.0
* networkx<=1.11

Datasets

DBLP

We uses the pre-processed DBLP dataset from Jhy1993/HAN You can run the FdGars, Player2Vec, GeniePath and GEM based on the DBLP dataset. Unzip the archive before using the dataset:

cd dataset
unzip DBLP4057_GAT_with_idx_tra200_val_800.zip

Example dataset

We implement example graphs for SemiGNN, GAS and GEM in data_loader.py. Because those models require unique graph structures or node types, which cannot be found in opensource datasets.

Yelp dataset

For GraphConsis, we preprocessed Yelp Spam Review Dataset with reviews as nodes and three relations as edges.

The dataset with .mat format is located at /dataset/YelpChi.zip. The .mat file includes:

• net_rur, net_rtr, net_rsr: three sparse matrices representing three homo-graphs defined in GraphConsis paper;
• features: a sparse matrix of 32-dimension handcrafted features;
• label: a numpy array with the ground truth of nodes. 1 represents spam and 0 represents benign.

The YelpChi data preprocessing details can be found in our CIKM'20 paper. To get the complete metadata of the Yelp dataset, please email to [email protected] for inquiry.

User Guide

Running the example code

You can find the implemented models in algorithms directory. For example, you can run Player2Vec using:

python Player2Vec_main.py 

You can specify parameters for models when running the code.

Running on your datasets

Have a look at the load_data_dblp() function in utils/utils.py for an example.

In order to use your own data, you have to provide:

• adjacency matrices or adjlists (for GAS);
• a feature matrix
• a label matrix then split feature matrix and label matrix into testing data and training data.

You can specify a dataset as follows:

python xx_main.py --dataset your_dataset 

or by editing xx_main.py

The structure of code

The repository is organized as follows:

• algorithms/ contains the implemented models and the corresponding example code;
• base_models/ contains the basic models (GCN);
• dataset/ contains the necessary dataset files;
• utils/ contains:
  • loading and splitting the data (data_loader.py);
  • contains various utilities (utils.py).

Implemented Models

Model	Paper	Venue	Reference
SemiGNN	A Semi-supervised Graph Attentive Network for Financial Fraud Detection	ICDM 2019	BibTex
Player2Vec	Key Player Identification in Underground Forums over Attributed Heterogeneous Information Network Embedding Framework	CIKM 2019	BibTex
GAS	Spam Review Detection with Graph Convolutional Networks	CIKM 2019	BibTex
FdGars	FdGars: Fraudster Detection via Graph Convolutional Networks in Online App Review System	WWW 2019	BibTex
GeniePath	GeniePath: Graph Neural Networks with Adaptive Receptive Paths	AAAI 2019	BibTex
GEM	Heterogeneous Graph Neural Networks for Malicious Account Detection	CIKM 2018	BibTex
GraphSAGE	Inductive Representation Learning on Large Graphs	NIPS 2017	BibTex
GraphConsis	Alleviating the Inconsistency Problem of Applying Graph Neural Network to Fraud Detection	SIGIR 2020	BibTex
HACUD	Cash-Out User Detection Based on Attributed Heterogeneous Information Network with a Hierarchical Attention Mechanism	AAAI 2019	BibTex

Model Comparison

Model	Application	Graph Type	Base Model
SemiGNN	Financial Fraud	Heterogeneous	GAT, LINE, DeepWalk
Player2Vec	Cyber Criminal	Heterogeneous	GAT, GCN
GAS	Opinion Fraud	Heterogeneous	GCN, GAT
FdGars	Opinion Fraud	Homogeneous	GCN
GeniePath	Financial Fraud	Homogeneous	GAT
GEM	Financial Fraud	Heterogeneous	GCN
GraphSAGE	Opinion Fraud	Homogeneous	GraphSAGE
GraphConsis	Opinion Fraud	Heterogeneous	GraphSAGE
HACUD	Financial Fraud	Heterogeneous	GAT

TODO List

• Implementing mini-batch training
• The log loss for GEM model
• Time-based sampling for GEM
• Add sampling methods
• Benchmarking SOTA models
• Scalable implementation
• Pytorch implementation

How to Contribute

You are welcomed to contribute to this open-source toolbox. The detailed instructions will be released soon. Currently, you can create issues or email to [email protected] for inquiry.