Graph-Based Discovery of Implicit Corporate Relationships Using Heterogeneous Network Learning

Abstract

This paper proposes a method based on deep heterogeneous graph neural networks and contrastive learning mechanisms to identify related transactions and hidden associations in enterprise networks. First, a heterogeneous economic network is constructed by integrating ownership relationships, supply chain transactions, and management overlaps. Multi-source data are transformed into a unified graph structure for modeling. Next, a heterogeneous graph neural network framework is designed. It combines node features, edge features, and attention aggregation. Through multi-layer feature learning, the model captures complex association patterns. To enhance the model's ability to distinguish hidden associations, a graph contrastive learning strategy is introduced. This further optimizes the discriminative power of node representations. In multiple experiments, systematic comparisons with existing heterogeneous graph learning methods validate the superior performance and robustness of the proposed method under different noise perturbations, enterprise size variations, and negative sampling strategies. The results show that this method outperforms traditional approaches in terms of Precision, Recall, and AUC. It effectively improves the accuracy of hidden related transaction detection and provides technical support for intelligent auditing and risk identification in complex enterprise network environments.