Accounting for metallogenic heterogeneity in ensemble learning-based mineral prospectivity mapping: A case study of copper mineralization in eastern Junggar, Xinjiang, China

While machine learning has greatly advanced data-driven mineral exploration, its effectiveness is often limited by geological heterogeneity arising from the coexistence of multiple mineralization types within the same area. For example, in eastern Junggar, Xinjiang, China, copper deposits occur within a broadly shared regional metallogenic framework and exhibit similar surface expressions, yet different deposit types vary significantly in ore-controlling factors and geochemical signatures. Treating these distinct deposit types as a single homogeneous target may obscure subtle, type-specific patterns crucial for mineral exploration. To address this issue, this study presents an interpretable ensemble learning approach that explicitly accounts for mineralization types and applies it to copper prospectivity mapping in eastern Junggar. Copper deposits in the study area are categorized into Porphyry, VMS, and Vein types, and separate models including random forest (RF), logistic regression (LR) and k-nearest neighbor (KNN) are evaluated to select the best model and identify the key controlling factors for each mineralization type. The probabilities predicted by the type-specific base models are then used as inputs for an LR meta -learner to construct a stacked ensemble, which produces an integrated copper prospectivity map. Shapley Additive Explanations (SHAP) analysis is further adopted to interpret the prediction mechanisms and factor contributions for each mineralization type. Results show that the type-specific models effectively capture the spatial distribution and ore-controlling patterns of each copper mineralization type. High-prospectivity zones in the integrated prospectivity map cover 20 % of the study area but contain 81 % of known copper deposits, outperforming baseline models that ignore mineralization types. SHAP analysis reveals distinct controlling factors and contribution patterns across the deposit types, consistent with their established ore-controlling characteristics. The proposed framework therefore improves predictive performance while providing more geologically consistent and interpretable prospectivity mapping.