This repository contains the code and datesets used for the experiments described in the paper entitled with
STM: A Spatio-Temporal Model for Dynamic Graph Fraud Detection
The experiments were conducted on a 2 GHz Linux server with an RTX 4090D (24GB)
Datasets:
scikit-learn==1.5.2
pandas==1.1.0
numpy==2.0.2
torch==2.4.0+cu124
torch-geometric==2.6.1
torch-scatter==2.1.2+pt24cu124
torch-sparse==0.6.18+pt24cu124
torch-spline-conv==1.2.2+pt24cu124
Make sure that the environment is set up before the experiment.
The relevant datasets have been uploaded to the repository, and the code can be run directly.
python run_FTM_anomalousDetection.py -k 1 --data wikipedia --n_epoch 15 --embedding_module HopTransformer --use_time_line --time_line_length 2 --sample_mode time --hard_sample --n_layer 1 --bs 100 --use_ST_dist
| Parameter | Type | Description |
|---|---|---|
| data | str | Dataset name |
| bs | int | Batch_size |
| n_degree | int | Number of neighbors to sample |
| n_head | int | Number of heads used in attention layer |
| time_line_length | int | Number of frame |
| epoch | int | The number of rounds of model training |
| embedding_module | str | Type of embedding module |
| k | int | number of Aggregated hop |
This repo is built upon the following work:
TGN: Temporal Graph Networks
https://github.com/twitter-research/tgn
FTM：A Frame-level Timeline Modeling Method for Temporal Graph Representation Learning
https://github.com/yeeeqichen/FTM
Many thanks to the authors and developers!
The repository will be continuously updated.