Risk-Level Multi-Prototype Contrastive Learning for Financial Fraud Detection
Year: 2026
Type: conference-paper
Abstract: Financial fraud detection remains challenging because fraudulent transactions are rare, heterogeneous, and embedded in high-dimensional tabular feature spaces. Existing tree-based models are strong structured-data classifiers, but they mainly rely on raw features and lack explicit risk-aware representation learning. We propose RMCL, a risk-level multi-prototype contrastive learning framework for financial fraud detection. RMCL learns complementary static and dynamic representations from the same processed features, uses a train-only LightGBM teacher to construct risk groups, and applies prototype-guided contrastive learning to capture heterogeneous fraud-risk patterns beyond binary labels. The learned dual-channel representation is compressed by training-fitted TruncatedSVD and fused with processed raw features for downstream LightGBM prediction. Experiments on Credit Card Fraud Detection and IEEE-CIS Fraud Detection show best or competitive performance across multiple metrics under an aligned validation/test protocol. Ablation, sensitivity, and visualization analyses support the contribution of risk-level prototype learning and SVD-enhanced fusion. All threshold-dependent metrics use validation-selected thresholds, and the test set is used only for final evaluation.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7213998645