<p>In Mineral Prospectivity Mapping (MPM), multi-source geological data are heterogeneous. However, conventional homogeneous graph neural networks are unable to effectively integrate heterogeneous geological data and mine complex geospatial correlations, leading to the limitation in prediction performance. To address this, this study proposed Heterogeneous Graph Convolutional Neural Networks (HGCN) and conducted mineral prospectivity prediction targeting copper deposits in the Zhongtiaoshan area (the southern of the North China Craton). Based on the geological background of the study area, four types of nodes were constructed, including geochemical elements, faults, high-density fault areas, and ore-controlling strata. Corresponding edge types were defined to characterize heterogeneous semantic relationships among these nodes, such as inter-element correlation, fault-controlled geochemical distribution, and mineralization spaces constrained by strata. Along with multiple meta-paths built on these relationships, the HGCN can aggregate node features through the heterogeneous graph convolution layer, effectively identifying the geospatial characteristic relationships. Experimental results showed, the HGCN model demonstrated strong robustness under different ratios of positive and negative samples. Compared with the homogeneous GCN and CNN, the proposed HGCN had stronger predictive performance and generalization ability. The prediction results indicated that the HGCN captured more geospatial details from the data, especially a high degree of consistency with the geometric morphology of the ore-controlling strata. Comparative experiments confirmed that geochemical spatial topological relationships were critical for multi-source heterogeneous data fusion in MPM.</p>