Although GIS-based mineral prospect mapping (MPM) is widely used, the lack of correlation interpretation between the final prospect map and metallogenesis makes the prediction results unreliable. Therefore, in this study, we utilized a machine learning technique that combines knowledge embedding and explainable ensemble learning to improve the interpretability of the results. Best-worst method (BWM) was adopted to embed prior knowledge. Stacking ensemble learning integrated these weights into a predictive model to generate targets. Apply Permutation Importance (PI), Partial Dependency Plot (PDP), and Locally Interpretable Model-Independent Explanation (LIME) to calculate global and local output weights for features and thereby target interpretability. Increased sex. The experiment was conducted at the Kelin ore dump in Sichuan, China. Experiments have demonstrated the effectiveness of this method, with 84% of the samples falling within the high and very high lithification probability zones, covering 6.58% of the total area. Na-feldspar spectrum, Na2O+K2O, ring structure, Li/La, and dimicite granite appear one after another as important predictive features, indicating orderliness and validating the close correlation with the metallogenic environment and exploration indicators of Keelin's pegmatite-type lithium deposit. doing.
