Collaborative Machine Learning for Risk Ranking Under Concurrent Class Imbalance and Distribution Shift

This study addresses modeling requirements in real business scenarios where data distributions are complex and continuously evolving. It investigates a unified modeling problem under the coexistence of class imbalance and distribution shift. To tackle practical challenges such as scarce high-risk event samples, majority class dominance in learning, and decision bias amplified by environmental change, a collaborative modeling framework is developed. The framework introduces class structure awareness and distribution stability constraints within a shared representation space. This design enables the model to preserve discriminative capability while reducing dependence on a single training distribution. Structural bias caused by imbalance is alleviated through explicit modeling of class proportion differences. Interference from environmental change is suppressed through distribution consistency constraints during representation learning. As a result, model stability and consistency across different data conditions are improved. To examine controllability and robustness, the study conducts a systematic analysis of key hyperparameters, data quality degradation, and environmental perturbations on risk ranking performance. Special attention is given to sensitivity patterns induced by feature noise injection and varying missing value ratios. The results indicate that when class imbalance and distribution shift interact, collaborative modeling effectively mitigates performance degradation. The model maintains a stable risk recognition capability in complex business data. This provides a practically adaptive modeling approach for data-driven decision-making in high-risk scenarios.