This study develops an interpretable machine learning framework combining extreme gradient boosting (XGBoost) and SHapley Additive exPlanations (SHAP) to explain the seasonal and regional variations in tropical cyclone (TC)-induced sea surface temperature (SST) cooling in the Northwest Pacific. The framework adopts TC characteristics (e.g., intensity, translation speed, size) and pre-storm ocean conditions (e.g., mixed layer depth, ocean thermal structure) as predictors, and skillfully reproduces the spatial structure of cooling in different seasons and regions. The method identifies the drivers of the cooling in both the marginal and the open seas, and quantifies their respective contributions by explicitly accounting for both the seasonal and regional variations. The drivers vary with seasons, with TC characteristics explaining 43-56% of the variance in cooling in the open sea and 47-52% in the marginal sea while oceanic drivers contribute 23-41% and 31-35%, respectively. Our results reveal that the seasonality of TC intensity and mixed layer depth are the most important contributors to the seasonal variations in the cooling over the marginal and open seas.