An Integrated Framework for Data-Driven Mineral Prospectivity Mapping Using Bagging-Based Positive-Unlabeled Learning and Bayesian Cost-Sensitive Logistic Regression

Mineral prospectivity mapping (MPM) is a spatial quantitative approach to delineation of exploration targets. The input data and the data integration approaches are sources of critical uncertainties in MPM. Mineralization is a rare event. Hence, the imbalanced geosciences datasets used in data-driven MPM are usually composed of sparse positive samples and abundant unlabeled data. Selecting reliable negative samples is more challenging than selecting positive samples for MPM. Another issue in MPM is cost-sensitive classification because false positive errors if followed-up could result in increasing exploration costs, whereas false negative errors if not recognized could result in missing undiscovered mineral deposits. Few data-driven methods for MPM that simultaneously address systemic uncertainty from imbalanced data and cost-sensitive issues are reported in the literature. This study proposes an integrated framework for data-driven MPM, which is denoted as BPUL-BCSLR, with Bagging-based positive-unlabeled learning (BPUL) and Bayesian cost-sensitive logistic regression (BCSLR) in order to address simultaneously the issues mentioned above. We utilized the BPUL-BCSLR framework for MPM in the Wulong Au district, China. The performance of the BPUL-BCSLR framework was compared with that of logistic regression (LR) and with the BCSLR. The prediction-area plot was utilized to evaluate the performance of the LR, BCSLR, and BPUL-BCSLR predictive models and to outline exploration targets in the study area. Return-risk analysis was carried out to determine low risk-high return targets obtained by the BCSLR and BPUL-BCSLR frameworks. The results demonstrate that the BPUL-BCSLR framework outperformed the LR and BCSLR. The low risk-high return exploration targets obtained through the BPUL-BCSLR framework can benefit further exploration for Au in the Wulong Au district.