Dual-relation heterogeneous graph approach with positive-unlabeled learning for mineral prospectivity mapping

Mineral Prospectivity Mapping (MPM) is a key mineral exploration technique driven by multi-source geoscientific information. Its central challenge lies in stably characterizing the spatial controls and geochemical responses associated with mineralization under realistic conditions where positive samples are scarce and reliable negative instances are difficult to obtain. To overcome the limitations of existing approaches, which are highly dependent on supervised information, this paper proposes a dual-relation heterogeneous graph approach with positive-unlabeled learning (DHPUG) for mineral prospectivity mapping. The proposed framework consists of two complementary modules designed to adaptively weight dual-relation graph features and reduce the influence of missing negative samples. In the representation-learning module, a heterogeneous graph relation-gating mechanism is introduced to achieve adaptive weighting and semantic encoding of dual-relation graph features. In the information-supervision module, a positive-unlabeled learning strategy with spatial-buffer constraints is used to mitigate the instability of decision boundaries under missing-negative conditions. Experiments conducted in the Shannan area of the eastern segment of the Tethyan Himalaya show that the proposed method achieved the best overall performance among the compared baseline models under paradigm-consistent experimental settings, with an AUROC of 0.948 and an accuracy of 0.875. The generated prospectivity patterns are highly consistent with the distribution of known deposits, the regional tectonic framework, and metallogenic types, indicating that the proposed method provides a technically viable pathway with a clear structure, a complete mechanism, and geoscientifically interpretable outputs for mineral prospectivity mapping and target prioritization under scarce labels and complex geological settings.