Mineral prospectivity modelling (MPM) is essential in identifying areas with high potential for mineralization in data-scarce and geologically complex regions. Traditional rule-based and statistical approaches often require strong assumptions, manual weighting, and struggle with non-linear interactions among predictors. Here, we conduct a systematic comparison between classification and regression paradigms on a Canadian magmatic Ni (±Cu ± Co ± PGE) sulphide dataset. We apply domain-aware preprocessing—morphological dilation and Gaussian smoothing—to enrich sparse occurrence labels with spatial context. Our classification pipeline uses a Random Forest classifier with bootstrapped balanced sampling to mitigate extreme class imbalance, while our regression pipeline employs Ridge and Lasso to predict continuous prospectivity scores. We evaluate classification via precision–recall analysis and confusion matrices, and regression via cross-validated R2. Results show that regression models produce smoother, geologically coherent prospectivity surfaces, whereas classification, despite high specificity, suffers low minority recall. We discuss metric selection under imbalance, the impact of domain-aware labeling, and implications for exploration targeting, advocating for regression-driven MPM workflows.