Mineral Prospectivity Mapping (MPM) is an pivotal methodology for identifying prospective deposits across large regions using complex geophysical datasets. The application of machine learning could significantly improve these processes. However, a critical challenge in data-driven mineral prospectivity mapping is the class imbalance between the mineralized locations and large background, which can severely limit model performance. To address this, this study systematically evaluates two machine learning workflows: a supervised Multilayer Perceptron (MLP) and a contrastive representation learning with radius classifier. The algorithm applied to a geophysical dataset from Finland included integrated data balancing (M=N), nested cross-validation, and methods for uncertainty quantification (radius distance and Shannon entropy) and interpretability (Shapley Additive exPlanations(SHAP)). The supervised MLP performed well with an Area Under the Curve (AUC) of 0.99, perfect of recall 100%, and Geometric Mean (G-mean) of 0.9937. The Shapley Additive explanations analysis showed that magnetic and pseudo-gravity anomalies are among those more significant features. Findings indicate that a well developed MLP can address significant data imbalance, successfully reducing the investigation footprint to around 1% of the total area while detecting all known deposits. The use of uncertainty maps showed that such deposits are found in high-confidence zones (low-uncertainty) along transitional corridors at geological boundaries, providing a reliable and economical framework for directing mineral exploration.