Bagging-based Positive–Unlabeled Data Learning Algorithm with Base Learners Random Forest and XGBoost for 3D Exploration Targeting in the Kalatongke District, Xinjiang, China

The Kalatongke district is located in northern Xinjiang, where the Kalatongke Cu–Ni sulfide deposit is one of the most famous early Permian Cu–Ni sulfide deposits. At present, 13 shallow mafic intrusions lying less than 500 m in the subsurface have been explored and almost all orebodies are hosted in these mafic intrusions. The ore-bearing intrusions are controlled by faults and are characterized by high magnetic, high density, low resistivity and low seismic wave velocity. With the depletion of proved resources, deep and peripheral prospecting is imminent in the study area. In this paper, three-dimensional (3D) mineral prospectivity modeling is implemented by integrating 3D geological–geophysical data and ore deposit geology. Relevant exploration criteria for prediction were acquired as follows: (1) The bi-dimensional empirical mode decomposition (BEMD) was applied to decompose magnetic and gravity data, and the decomposed images were used to interpret faults and construct fault buffering models; (2) the magnetic susceptibility and density models were obtained by inversion of gravity data and magnetic data; and (3) discrete smooth interpolation was used to interpolate frequency domain controlled-source electromagnetic method (FDCSEM) data and seismic data, and obtain resistivity and seismic wave velocity models. Taking fault buffering model, geophysical models (density, magnetic susceptibility, seismic wave velocity and resistivity) as exploration criteria, the bagging-based positive–unlabeled data learning (BPUL) algorithm that can effectively address the issue of imbalanced training data was used for 3D mineral prospectivity modeling. Random Forest (RF) and XGBoost were applied as the base learners of BPUL to develop novel BPUL–RF and BPUL–XGBoost models, respectively. The results indicate that the BPUL–XGBoost showed better performance than the BPUL–RF, RF and XGBoost prediction models. The prediction–area plot was applied to outline the targets, which are of great significance for further Cu–Ni exploration in the Kalatongke district.