Financial Fraud Detection on IEEE-CIS (out-of-time evaluation)
0.9289AUC-ROCLSTM+XGBoost Ensemble (0.6/0.4)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LSTM+XGBoost Ensemble (0.6/0.4)Ensemble Weighting=0.6/0.4, Classification Threshold (τ)=0.502026.04 | 0.9289 | 63.6 | 62.3 | 0.8898 | 0.0391 | |
| LSTM (this study)Classification Threshold (τ)=0.602026.04 | 0.9205 | 55.3 | 62.3 | 0.8579 | 0.0626 | |
| XGBoost (baseline)Classification Threshold (τ)=0.502026.04 | 0.9021 | 55.72 | 81.3 | 0.9004 | 0.0017 | |
| TransformerClassification Threshold (τ)=0.502026.04 | 0.8832 | 43.7 | 57 | — | — | |
| GNN / GraphSAGEClassification Threshold (τ)=0.50, Subgraph Sampling=50,000-transaction stratified subgraph2026.04 | 0.825 | 32 | 20 | — | — |