DualShiftNet: Joint Class-Imbalance and Distribution-Shift Aware Learning for Business Risk Prediction

Business risk prediction tasks such as fraud detection, credit default prediction, and equipment failure forecasting face two fundamental challenges simultaneously: severe class imbalance where anomalous events are extremely rare, and distribution shift where data patterns evolve over time due to changing business conditions or adversarial behavior. While existing approaches address these challenges in isolation, real-world deployment requires handling both simultaneously. We propose DualShiftNet, a unified framework that jointly addresses class imbalance and distribution shift through a two-stage architecture. The first stage learns imbalance-aware representations using synthetic minority oversampling, focal loss optimization, and class-balanced contrastive learning to create discriminative embeddings. The second stage employs Maximum Mean Discrepancy (MMD) based drift detection coupled with importance reweighting to adapt predictions under distribution shift. Additionally, we introduce an uncertainty-driven threshold calibration mechanism that dynamically adjusts decision boundaries based on detected shift intensity. Experiments on three benchmark datasets demonstrate that DualShiftNet achieves relative improvements of approximately 3–4% in AUC-ROC scores and 10–22% in F1-scores compared to state-of-the-art methods that address only one challenge. Our ablation studies confirm that both stages contribute meaningfully to performance, with the joint approach outperforming sequential or isolated solutions.