Dynamic knowledge graph and graph neural network-based early-warning method

Abstract

Early warning of corporate debt default puts forward higher requirements for multi-source credit data access, relationship risk identification and early warning level judgment. Aiming at the problems of lagging credit indicators, insufficient description of relational risks and inadequate utilization of external events in traditional early warning, this paper proposes an early warning method of corporate debt default driven by dynamic knowledge graph. This method integrates credit, bond, rating, guarantee, litigation, public opinion and transaction data, and forms an updatable knowledge graph through subject matching, indicator standardization, event labeling and relationship mapping. It constructs 1200 enterprise nodes, 3120 bond nodes, 2760 associated subject nodes, 6380 event nodes and 29300 relationship edges. Identify debt service pressure, maturity pressure, market credit and external shock signals. On this basis, the early warning scoring mechanism is introduced to transform the risk characteristics into the early warning state of the enterprise, and the four types of judgment are normal, concern, warning and serious warning. Experimental results show that the proposed method achieves 93.4%, 92.6%, 91.8%, 92.2% and 94.1% in Accuracy, Precision, Recall, F1 and AUC, respectively. The amount of warning advance, stability and false alarm control reach 91.7%, 90.6% and 90.8%, respectively. The research can provide engineering support for graph-based risk signal extraction, model inference and early-warning deployment.