<p>This paper proposes a domain-adaptive hybrid machine learning framework aimed at addressing the challenges of risk identification and claim amount prediction in logistics claim decision-making. This framework systematically builds an end-to-end risk control model by integrating business rule constraints, adaptive imbalanced learning, and constrained ensemble optimization. The research first achieved a three-level fine-grained risk annotation of “reasonable claim,” “excessive claim,” and “severe excessive claim” through a dynamic threshold segmentation method combining constrained K-means clustering and ordinal regression, with an identification accuracy rate of 92.5%. Secondly, an XGBoost regression model embedded with business rule constraints was developed. By applying logarithmic transformation of the target variable and feature interaction optimization, the root mean square error (RMSE) of claim amount prediction was reduced by 37.2%. Experimental results show that this model significantly outperforms the baseline model in key indicators such as recall rate and prediction stability, effectively resolving the conflict between the imbalance of claim data and complex business constraints, and has extremely high industrial application value.</p>