Fraud Detection with GNN

The code of this project is based on Streaming Graph Neural Networks via Continual Learning（CIKM 2020).

Usage

Installation

Install dependencies

pip install -r requirements.txt

Install PyTorch 1.12.1 (CPU only)

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 -c pytorch

With CUDA support

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge

Generate dataset

Please first download the Amazon Review Dataset (2014) from here, and place the json files under /data/origin/.

Generate the preprocessed dataset. Here, we are using reviews_Amazon_Instant_Video_5.json as an example.

python preprocess.py --filename=reviews_Amazon_Instant_Video_5.json --dataset-name=amazon_instant_video --num-streams=15 --corpus-sim-percentile=95 --usu-interval=259200 --fradulent-threshold=0.5 --feature-schema=sentence_embeddings

Training

ContinualFraudre

python main_amazon_stream.py --data=amazon_instant_video --new-ratio=0.8 --memory-size=1000 --ewc-lambda=80.0 --max-detect-size=8 --batch-size=512 --num-epochs=60 --learning-rate=0.1

ContinualGNN

python main_stream.py --data=amazon_instant_video --new-ratio=0.8 --memory-size=1000 --ewc-lambda=80.0 --max-detect-size=8 --batch-size=512 --num-epochs=60 --learning-rate=0.001

Cora

ContinualGNN (proposed model) on Cora:

cd src/
python main_stream.py --data=cora --new-ratio=0.8 --memory-size=250 --ewc-lambda=80.0 

OnlineGNN (lower bound) on Cora:

python main_stream.py --data=cora

If using cuda, set --cuda.