<p>Due to the rarity of metallogenic processes and the diversity of ore deposit types, data-driven mineral prospectivity mapping in large, geologically complex regions faces two major challenges: insufficient representativeness of negative samples and limited capacity for differentiated interpretation under complex metallogenic patterns. To address these challenges, this paper proposes a novel analysis framework for data-driven mineral prospectivity mapping based on dynamic collaborative sample optimization and multi-pattern analysis. Firstly, a dual-constraint negative sample selection strategy is developed by integrating Mahalanobis and Euclidean distances through a covariance-weighted mechanism, combined with a clustering-based quota sampling approach, which enables the adaptive selection of high-confidence negative samples. This strategy significantly outperformed traditional methods in terms of AUC and F1-score metrics. Secondly, a spatially constrained adaptive Bayesian Gaussian mixture model was developed, which incorporates a Dirichlet process prior to automatically determine the optimal number of mineralization patterns. The model integrates feature-space joint distribution modeling with boundary-adaptive regularization, enabling precise modeling of spatial probability field. This model successfully delineated 24 distinct mineralization patterns. Finally, by integrating random forest with SHapley Additive exPlanations, the contribution of ore-controlling factors was quantitatively assessed, which revealed that fault structures and gold concentration are dominant controlling factors, and a strong coupling relationship between characteristic geochemical elements and mineralization is significant for patterns 0, 2, 15, and 23. The experimental analysis proved that the framework is capable of successfully delineating prospective mineralization targets and it offers a novel solution for data-driven mineral prospectivity mapping in complex geological settings.</p>