The graph attention network and its post-hoc explanation for recognizing mineralization-related geochemical anomalies

Abstract

Deep learning algorithms have become a cutting-edge technology for mining geochemical survey data to identify geochemical patterns related to mineralization. Similarities in the origins of the same types of mineral deposits may result in similarities in the observed geochemical anomalies to some extent. However, image-based models have limited ability to model the relationships between samples because fixed-size pixel patches are not connected. Graphs have enormous potential to capture spatial information, which can model the complex nonlinear spatial relationship between vertices and edges and effectively measure the spatial relationships between metallogenic information and geochemical survey sites. The graph-based model considers both labeled and unlabeled data as inputs, allowing the relationships between individual samples to be incorporated into the network. In this study, we used a graph-based model, the graph attention network (GAT), to recognize geochemical anomalies associated with gold mineralization in the Xiaoqinling–Xiong'ershan region of the western Henan Province, China. The GNNExplainer, a method for explaining the predictions of graph-based deep learning tasks, was used to determine the importance of the input features of the trained GAT model. The results indicated that the model regards Bi, Cd, Mn, Zn, and Mo as important geochemical elements for mineralization. Meanwhile, a comparative study of GAT with a convolutional neural network suggests that the graph-based model can be applied to effectively identify mineralization-related geochemical patterns, and geochemical anomalies recognized by GAT are more representative of gold mineralization.