Unlabeled Sample Selection for Mineral Prospectivity Mapping by Semi-supervised Support Vector Machine

Semi-supervised learning (SSL) algorithms can use unlabeled data to improve the performance of supervised learning algorithms for mineral prospectivity mapping with few known mineral deposits or mineralized blocks. However, SSL algorithms are sensitive to unlabeled samples and, in some cases, perform worse than supervised algorithms. In this study, a quasi-Newton method for semi-supervised support vector machine (QN–S3VM) was used in the 3D mineral prospectivity mapping of the Honghai volcanogenic massive sulfide Cu–Zn deposit in eastern Tianshan, northwestern China. Three Euclidean distance-based similarity measures of unlabeled samples to known mineral deposits or mineralized blocks were proposed to select unlabeled samples. The influence of the similarity and number of unlabeled samples on the performance of the QN–S3VM was investigated. The results showed that lower similarity in unlabeled samples yielded enhanced QN–S3VM performance. The performance of the QN–S3VM was affected by the number of unlabeled samples used. However, there was no consistent pattern among them. Compared with random selection, the QN–S3VM trained with unlabeled samples selected by the similarity measure had higher generalization and stability. Among the maximum, minimum, and average similarities, the minimum similarity had the best generalization while the average similarity had the best stability. Therefore, similarity to known mineral deposits or mineralized blocks is a good tool for unlabeled sample selection. This can effectively guarantee the performance of SSL for mineral prospectivity mapping.