Utilizing Effective Dynamic Graph Learning to Shield Financial Stability from Risk Propagation

TL;DR

Financial networks exhibit rapid risk propagation across intertwined spatial and temporal dimensions, with many risky signals remaining unlabeled. GraphShield combines a sandwich-style dynamic graph learning module with separable kernel attention, a semi-supervised risk recognition mechanism based on Gaussian mixtures, and a visualization tool to quantify propagation via PCC/PDC and Granger-style causality. The approach achieves state-of-the-art performance on multiple real-world and open datasets, robust to low labeling and high unlabeled-ratio settings, and provides actionable insights into influential nodes and propagation paths. Its successful deployment in a Sichuan bank demonstrates practical impact for financial stability and potential extensions to supply chain finance and risk management domains.

Abstract

Financial risks can propagate across both tightly coupled temporal and spatial dimensions, posing significant threats to financial stability. Moreover, risks embedded in unlabeled data are often difficult to detect. To address these challenges, we introduce GraphShield, a novel approach with three key innovations: Enhanced Cross-Domain Infor mation Learning: We propose a dynamic graph learning module to improve information learning across temporal and spatial domains. Advanced Risk Recognition: By leveraging the clustering characteristics of risks, we construct a risk recognizing module to enhance the identification of hidden threats. Risk Propagation Visualization: We provide a visualization tool for quantifying and validating nodes that trigger widespread cascading risks. Extensive experiments on two real-world and two open-source datasets demonstrate the robust performance of our framework. Our approach represents a significant advancement in leveraging artificial intelligence to enhance financial stability, offering a powerful solution to mitigate the spread of risks within financial networks.

Figures (6)

• Figure 1: An Example to Illustrate Financial Risk Propagation. (a) Financial risks are intricately interconnected across spatial and temporal domains. For instance, the risk status of a node (e.g., $u_t$) is shaped by both its immediate neighbors and its previous state ($u_{t-1}$). (b) Risk nodes (e.g., $u_2$) hidden in unlabelled data drive the propagation of risk. Its risk status influences both its immediate neighbors and its future vulnerabilities. Failing to effectively identify these nodes can result in uncontrolled risk propagation. (c) Risk samples frequently display clustering patterns in both temporal and spatial dimensions. For example, if certain nodes are identified as risky, their neighboring nodes and previous states might also pose potential risks.
• Figure 2: Overall Architecture of Our Proposed GraphShield Framework. (a) We construct the dynamic graph learning module integrating spatial and temporal operations within each layer. The sandwich-style stacking structure ensures thorough learning of spatial and temporal information. (b) We represent each risk sample as originating from a Gaussian mixture distribution and employ a fully-connected neural network to enhance the hidden risk recognition. (c) We offer a financial risk propagation visualization analysis tool to quantifying and validating the impact effects between risks.
• Figure 3: Sensitivity of Embedding Dimension and Layer Number in Dynamic Graph Learning Module.
• Figure 4: Sensitivity of Balance Weight $\tau_3$ and Layer Number in Semi-supervised Risk Detecting Module.
• Figure 5: Examples of Estimation and Verification of Impact Effects in Stock Sell-off Risk Propagation. Such examples are extracted from the Shareholding dataset, which serves as the basis for analyzing risk propagation mechanisms. PCC is used to quantify bidirectional effects of risk propagation in the spatial domain, whereas PDC focuses on unidirectional effects in the temporal domain. PCC and PDC values with $p\text{-value} \geq 0.05$ are considered statistically insignificant and are thus excluded from further analysis. The shareholder nodes $\mathrm{S}_{4}$, $\mathrm{S}_{8}$, $\mathrm{S}_{10}$, $\mathrm{S}_{11}$, and $\mathrm{S}_{14}$ have been identified as risk nodes by our proposed approach and are highlighted in orange. Notably, these nodes align with the actual labels and play a critical role in risk propagation, warranting special attention.