Unbiased and augmentation-free self-supervised graph contrastive learning for detecting mineralization-related geochemical anomalies

Graph contrastive learning (GCL) provides a self-supervised learning technique, which utilizes the node (sample) features, neighborhood information of node pairs and sparse label information to build a self-supervised model for identifying mineralization-related geochemical anomalies. However, using the GCL algorithm to construct a self-supervised model needs to augment graph-structured data by changing the features or views between nodes, which can result in the loss of some key neighborhood information in the original graph-structured data. Unbiased and augmentation-free self-supervised graph contrastive learning (USAF-GCL) is a novel GCL learning technique that does not need to augment graph-structured data when building a self-supervised model. Thus, it is more reliable than the GCL algorithm in detecting high-dimensional geochemical anomalies through self-supervised node representation of graph-structured data. To show the superiority of the USAF-GCL technique in the detection of mineralization-related geochemical anomalies, a USAF-GCL model and a GCL model were built on the geochemical data set of the 1:200,000 stream sediment survey covered the Baishan area, Jilin Province, China. A comparison between the USAF-GCL model and the GCL model shows that the former is obviously superior to the latter in the detection of mineralization-related geochemical anomalies. The receiver operating characteristic (ROC) curve of the USAF-GCL model overwhelmingly dominates the ROC curve of the GCL model, the area under the ROC curves (AUCs) of the USAF-GCL and GCL models are 0.9545 and 0.8601, respectively. The precision recall (PR) curve of the USAF-GCL model dominates that of the GCL model. The area under the PR curves (AUPRCs) of the USAF-GCL and GCL models are 0.9312 and 0.026, respectively. F1 scores of the USAF-GCL and GCL models are 0.7360 and 0.0672, respectively. These statistical results indicate that the USAF-GCL model is much better than the GCL model in identifying mineralized anomaly samples from the geochemical exploration data set. The anomaly areas detected by the USAF-GCL model occupy only 17.1 % of the entire exploration area, identifying 100 % of the known polymetallic deposits; while those identified by the GCL model occupy 25 % of the entire area, identifying only 93 % of the known polymetallic deposits. Geologically, the mineralization-related anomalies identified by the USAF-GCL model are strongly consistent with the polymetallic metallogenic characteristics in the exploration area. Therefore, the USAF-GCL technique is a reliable and powerful tool for the detection of mineralization-related geochemical anomalies.