// the find
safe-graph/DGFraud
A Deep Graph-based Toolbox for Fraud Detection
A collection of GNN-based fraud detection model implementations, covering nine papers from 2017–2020 across financial fraud, spam review, and cyber criminal detection. It's aimed at researchers who want runnable baselines without re-implementing everything from scratch. Not a production library.
Nine distinct model implementations in one place with per-model READMEs and BibTeX citations makes it genuinely useful as a research starting point. The Yelp and DBLP datasets are preprocessed and included, so you can run comparisons immediately. GraphConsis gets the most complete treatment — it has minibatch training, custom samplers, and aggregators rather than just a toy script. The model comparison table is honest about what graph type and base architecture each method uses.
Abandoned since April 2022 and requires TF1 (tensorflow<2.0, networkx<=1.11) — these dependency pins are years out of date and will fight you on any modern Python environment. The TF2 successor repo is where active work went, making this one a dead end. Mini-batch training is listed as TODO for most models, meaning several implementations will OOM on anything larger than the bundled toy datasets. No unified evaluation harness — you can't run all nine models against the same dataset and get a comparison table; that's left entirely to you.