FraudGraph — Credit Card Fraud Detection with GNNs
Built on the Kaggle Credit Card Fraud Detection dataset (284,807 transactions, 492 fraud), the pipeline subsamples to 15,000 nodes while keeping every fraud case, builds a transaction graph, and trains a hand-rolled GCN in plain PyTorch with no torch_geometric dependency. A FastAPI backend exposes stats, graph, per-transaction and prediction endpoints, and the FraudGraph React dashboard renders normal and fraudulent nodes, live model metrics, and per-node inspection that highlights graph neighbours to show how connected transactions influence each prediction. On held-out data the model reaches 97.6% accuracy, 0.97 ROC-AUC and 85% fraud recall.
Built on the Kaggle Credit Card Fraud Detection dataset (284,807 transactions, 492 fraud), the pipeline subsamples to 15,000 nodes while keeping every fraud case, builds a transaction graph, and trains a hand-rolled GCN in plain PyTorch with no torch_geometric dependency. A FastAPI backend exposes stats, graph, per-transaction and prediction endpoints, and the FraudGraph React dashboard renders normal and fraudulent nodes, live model metrics, and per-node inspection that highlights graph neighbours to show how connected transactions influence each prediction. On held-out data the model reaches 97.6% accuracy, 0.97 ROC-AUC and 85% fraud recall.