Machine learning algorithms have been widely used for mineral prediction. However, these algorithms are hard to deal with the geological data with the characteristics of high-dimensional, sparse and unbalanced samples. It is important to study some new mineral prediction models suited for geological big data. In this paper, a semi-supervised co-training model was proposed for mineral prediction in Haobugao district, Inner Mongolia. Firstly, nine prospecting factors were extracted consisting the faults, the Permian formation, the Yanshanian intrusions, the contact zone between Yanshanian intrusions and Permian formation, the zones of skarn alteration and Pb, Zn, Sn, Cu geochemical anomalies. Secondly, the feature selection method based RF-RFE was used to optimize the factors combinations. The eight factors were selected as the final prospecting factors excepted Sn geochemical anomaly. Then SVM and RF model were used as basic classifier for co-training model to predict mineral probability. The analysis of the ROC curves and predictability curves showed that the semi-supervised co-training model was more accurate than single SVM or RF model. It is suggested that this method has a certain feasibility for mineral prediction in the big data environment.