naidu9678/ieee-fraud-ft-transformer-gnn

FT-Transformer + GraphSAGE — IEEE Fraud Detection

Architecture

• —FT-Transformer: dmodel=64, nheads=4, nlayers=2, dffn=128
• —GraphSAGE: 2 layers, hidden=128, out=64
• —Fusion MLP: 96 → 128 → 2

Training

• —Dataset: IEEE-CIS Fraud Detection (590,540 train, 506,691 test)
• —Features: 432
• —Loss: FocalLoss(gamma=2.0, fraud_weight=25.0)
• —FT pre-training: 15 epochs × 1-fold CV
• —Joint training: 20 epochs

Results (OOF)

• —AUC: 0.5166
• —F1: 0.2037
• —Precision: 0.4290
• —Recall: 0.1336

Usage

import torch
from huggingface_hub import hf_hub_download
ckpt  = torch.load(hf_hub_download("naidu9678/ieee-fraud-ft-transformer-gnn", "gnn_ft_transformer_complete.pt"),
                   map_location="cpu", weights_only=False)
model = FraudDetectionSystem(ckpt["config"])
model.load_state_dict(ckpt["model_state"])
model.eval()