<p>The graph convolutional network (GCN) has proven to be a powerful tool for exploration criterion identification and mineral prospectivity analysis because of its excellent ability to integrate and capture complex spatial geo-anomalies associated with mineralization. However, the graph convolution of the GCN is essentially a special form of Laplacian smoothing that can result in the over-smoothing of output features. This over-smoothing can undermine the reliability of mineral prospectivity mapping (MPM), which aims to identify mineralization-related geo-anomalies typically characterized by a series of anomalous values in the GCN model. To solve this problem, we introduce a novel deep learning model, the self-ensembling graph convolutional network (SEGCN), which integrates the mean teacher model with a GCN for MPM by fully learning unlabelled geological features. To illustrate the superiority of the SEGCN model, a case study for tungsten polymetallic prospectivity analysis was carried out in the Nanling belt, China. The SEGCN model was used to build a classification model based on the fusion of multiple geospatial datasets, including geochemical data and data on faults and rock masses. The SEGCN model was compared with conventional GCN, convolutional neural network (CNN), random forest (RF), and support vector machine (SVM) models in terms of their classification performance and predictive accuracy, and was evaluated via receiver operating characteristic (ROC) curves and prediction–area (P–A) plots. The results demonstrated that the SEGCN model significantly outperformed the other four models in mapping tungsten polymetallic prospectivity in the Nanling region of China. We concluded that the SEGCN model can establish an efficient, robust, and high-performance classification model to effectively recognize favourable mineralization targets from multi-source geo-anomalies, thereby suppressing the over-smoothing effect in the GCN.</p>