Uncovering Hidden Financial Risk Chains through Temporal and Relational Knowledge Graph Learning

Abstract

This paper addresses the challenges in financial risk identification, such as the difficulty in uniformly representing multi-source heterogeneous information, the complexity and concealment of transaction relationship chains, and the dispersion and semantic inconsistency of risk evidence across entities. It proposes a financial knowledge graph-driven risk identification framework that integrates transaction entities, fund flow relationships, event context, and temporal attributes into a unified graph structure representation to support risk inference based on interconnected links. The method first normalizes the original data into entities and constructs relationships, forming a knowledge graph containing entity, relationship, and temporal information. Then, a relationship-aware graph representation learning mechanism is employed to propagate and aggregate differentiated information in multi-relational neighborhoods, enabling node representations to incorporate multi-hop paths and local structural patterns. Simultaneously, a time-decay evidence weighting strategy is introduced to strengthen the impact of recent behavior and key interactions on risk judgment, improving the ability to characterize dynamically evolving risks dynamically. At the risk output layer, the framework maps the learned entity representations to continuous risk scores and expresses them probabilistically, thereby achieving rankable risk intensity estimation and supporting link-level evidence localization. Comparative experimental results show that the method achieves consistent improvement on multiple evaluation metrics, can more effectively capture structural risk signals in complex transaction networks, and enhances the stability and interpretability of model output while maintaining discriminative ability.