Scholarlink

<p>Recognizing mineralization-associated geochemical anomalies is essential for geochemical data-based mineral exploration targeting. However, the inherent spatial anisotropy of element migration is often overlooked in geochemical anomaly recognition. This anisotropy stems from various process-driven factors genetically related to mineralization, such as fluid flow direction, structural conduits, lithological boundaries, and thermodynamic gradients. To make use of this anisotropy information in geochemical anomaly recognition, an anomaly identification framework is proposed by the integration of directed graph neural network (DIGNN) and support vector machine (SVM) and named directed graph neural network support vector machine (DIGSVM). The approach employs a graph aggregator of directional flow imbalance to generate node embeddings from a directed graph constructed upon adjacency relationships. This method simultaneously encodes geochemical element concentrations and directional topological relationships, and then the integrated SVM model classifies the node embeddings. The framework combines the capability of DIGNNs in capturing directional relationships among neighboring data points with the advantages of SVMs in binary classification. A case study is presented for modeling a geochemical dataset obtained from within the Baishan region in Jilin Province (China). Different classification models were built for the identification of mineralization-associated geochemical anomalies. The features of their receiver operating characteristic curves (ROC) revealed that the novel DIGSVM model achieved the most notable effectiveness in the separation of anomalous and normal geochemical samples. The values of the areas under the curves (AUCs) revealed that the DIGSVM method ranked the highest with 0.927, followed by DIGNN with 0.915, both of them outperformed the undirected graph convolutional network (GCN) with 0.891 and the SVM classifier with 0.872. Additionally, the lift index was employed to characterize target enrichment levels. The DIGSVM model (8.19) significantly outperformed the DIGNN (6.71), GCN (4.69), and SVM (3.39) models. All these models revealed, however, spatial correlations between anomalous zones and known polymetallic deposits, with the DIGSVM model demonstrating the strongest correlation. These findings indicate that the proposed DIGSVM framework performs exceptionally well in the identification of geochemical anomalies, as it fully utilizes the spatial anisotropy information of element migration. This framework can serve as an efficient and reliable method for identifying mineralization-associated geochemical anomalies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Directed Graph Neural Network Support Vector Machine for Interpretability-Enhanced Recognition of Geochemical Anomalies

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Hang Su,
  • Yongliang Chen

摘要

Recognizing mineralization-associated geochemical anomalies is essential for geochemical data-based mineral exploration targeting. However, the inherent spatial anisotropy of element migration is often overlooked in geochemical anomaly recognition. This anisotropy stems from various process-driven factors genetically related to mineralization, such as fluid flow direction, structural conduits, lithological boundaries, and thermodynamic gradients. To make use of this anisotropy information in geochemical anomaly recognition, an anomaly identification framework is proposed by the integration of directed graph neural network (DIGNN) and support vector machine (SVM) and named directed graph neural network support vector machine (DIGSVM). The approach employs a graph aggregator of directional flow imbalance to generate node embeddings from a directed graph constructed upon adjacency relationships. This method simultaneously encodes geochemical element concentrations and directional topological relationships, and then the integrated SVM model classifies the node embeddings. The framework combines the capability of DIGNNs in capturing directional relationships among neighboring data points with the advantages of SVMs in binary classification. A case study is presented for modeling a geochemical dataset obtained from within the Baishan region in Jilin Province (China). Different classification models were built for the identification of mineralization-associated geochemical anomalies. The features of their receiver operating characteristic curves (ROC) revealed that the novel DIGSVM model achieved the most notable effectiveness in the separation of anomalous and normal geochemical samples. The values of the areas under the curves (AUCs) revealed that the DIGSVM method ranked the highest with 0.927, followed by DIGNN with 0.915, both of them outperformed the undirected graph convolutional network (GCN) with 0.891 and the SVM classifier with 0.872. Additionally, the lift index was employed to characterize target enrichment levels. The DIGSVM model (8.19) significantly outperformed the DIGNN (6.71), GCN (4.69), and SVM (3.39) models. All these models revealed, however, spatial correlations between anomalous zones and known polymetallic deposits, with the DIGSVM model demonstrating the strongest correlation. These findings indicate that the proposed DIGSVM framework performs exceptionally well in the identification of geochemical anomalies, as it fully utilizes the spatial anisotropy information of element migration. This framework can serve as an efficient and reliable method for identifying mineralization-associated geochemical anomalies.