A semi-supervised approach for mineral prospectivity mapping via weighted positive-unlabeled learning and tree-structured parzen estimator for hyperparameter optimization
Mahsa Hajihosseinlou, Abbas Maghsoudi, Reza Ghezelbash
Abstract
Mineral prospectivity mapping (MPM) has been approached through supervised, unsupervised, and semi-supervised learning methods. Supervised models often yield high accuracy but require both positive and negative labels, which are rarely available in mineral exploration. Unsupervised methods avoid the need for labels but typically lack geological interpretability and cannot integrate known mineral occurrences. Semi-supervised techniques, particularly Positive Unlabeled (PU) learning, offer a balanced alternative by reducing dependence on negative labels. Beyond addressing label scarcity, PU learning enhances generalization, mitigates labeling bias, and utilizes known mineral information more effectively. In this study, Weighted PU learning was employed. Unlike earlier PU strategies that treated all unlabeled data uniformly, this method assigns differentiated weights to positive and unlabeled instances, reducing class imbalance and improving model robustness. Since Weighted PU learning requires a classifier, we selected Gaussian Naive Bayes (GNB) due to its probabilistic nature and ability to handle uncertainty in mineral prospectivity modeling. Additionally, the parameters of Weighted PU learning were optimized using the Tree-structured Parzen Estimator (TPE) to achieve the best performance. The method was applied to Mississippi Valley-Type (MVT) Pb-Zn mineralization in the Malayer-Esfahan metallogenic belt, Iran. A conceptual deposit model guided the selection and extraction of relevant exploration layers, which were then integrated into the modeling workflow. Comparative evaluation against Semi Supervised Random Forest learning demonstrated the superior predictive performance of the Weighted PU model. These findings highlight the practical effectiveness of the Weighted PU approach in real-world mineral prospectivity mapping.