GeoGraph contrastive attention Network: Geological priors guided CoDA-aware graph contrastive attention network for geochemical anomaly recognition

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurately recognizing geochemical anomalies is critical for mineral exploration, yet regional geochemical data are often constrained by compositional closure, strong spatial heterogeneity, and scarce deposit labels. Existing graph learning methods rarely preserve both log-ratio geometric structure and geological control constraints during message passing. This study proposes a geological prior-guided, CoDA-aware graph learning framework, called the GeoGraph Contrastive Attention Network (GGCAN), for detecting mineralization-related geochemical anomalies. GGCAN builds a geological prior graph over 2,520 geochemical nodes in the Suolong area, Zhaishang-Mawu gold district, Western Qinling Orogenic Belt, incorporating spatial proximity, lithological continuity, fault buffer connectivity, fault distance, and favorable host rocks. Following centered log-ratio (CLR) transformation, the eleven geochemical elements are propagated across compositional channels under a CLR-guided zero-sum centering constraint, while geological priors are encoded in separate Euclidean channels. An anisotropic, edge-refined attention mechanism captures spatial neighborhood contributions, coupled with a two-stage weak-supervision strategy that combines supervised contrastive pretraining with non-negative PU risk optimization. Compared with GraphSAGE, GCN, GAT, MLP, and GAE, GGCAN achieves the best performance (PR-AUC 0.9166, F1 0.8702, MCC 0.8525, Precision 0.8615, Recall 0.8794). Prediction-Area analysis shows GGCAN identifies 89.9% of known deposits within only 10.1% of the study area. Ablation results confirm that geological priors, CoDA-aware attention, and CoDA centering jointly improve anomaly ranking, classification stability, and interpretability. Identified anomalies show strong spatial coherence with known deposits, favorable host rocks, granite outer contact zones, and mineralization-controlling structures, demonstrating GGCAN’s effectiveness as a geologically interpretable framework for geochemical anomaly detection.