Graph Dimension Attention Networks for Enterprise Credit Assessment

TL;DR

GDAN introduces dimension-level attention within a heterogeneous enterprise graph to improve credit risk assessment, addressing the limitation of traditional entity-level attention in capturing dimension-specific risk signals. The framework includes DistShift, a data-centric edge explainer that uses distribution shifts during message passing to provide interpretable edge-level insights, and a real-world ECAD dataset released to the public. Empirical results on ECAD, SMEsD, and DBLP show GDAN consistently outperforms state-of-the-art baselines, with ablations confirming the value of dimension attention and DistShift. The work advances interpretable, dimension-aware graph learning for finance and provides a practical, scalable approach for enterprise risk evaluation.

Abstract

Enterprise credit assessment is critical for evaluating financial risk, and Graph Neural Networks (GNNs), with their advanced capability to model inter-entity relationships, are a natural tool to get a deeper understanding of these financial networks. However, existing GNN-based methodologies predominantly emphasize entity-level attention mechanisms for contagion risk aggregation, often overlooking the heterogeneous importance of different feature dimensions, thus falling short in adequately modeling credit risk levels. To address this issue, we propose a novel architecture named Graph Dimension Attention Network (GDAN), which incorporates a dimension-level attention mechanism to capture fine-grained risk-related characteristics. Furthermore, we explore the interpretability of the GNN-based method in financial scenarios and propose a simple but effective data-centric explainer for GDAN, called GDAN-DistShift. DistShift provides edge-level interpretability by quantifying distribution shifts during the message-passing process. Moreover, we collected a real-world, multi-source Enterprise Credit Assessment Dataset (ECAD) and have made it accessible to the research community since high-quality datasets are lacking in this field. Extensive experiments conducted on ECAD demonstrate the effectiveness of our methods. In addition, we ran GDAN on the well-known datasets SMEsD and DBLP, also with excellent results.

Figures (10)

• Figure 1: An Illustrative Example of Enterprise Heterogeneous Graph (EHG). The depth of color shading conveys the significance of each feature dimension, while the thickness of the edges reflects the relative importance of the respective relationships.
• Figure 2: Model Architecture of GDAN. (1) Input Projection Layer aims to map original risk features into a shared latent space; (2) Graph Convolution Layer learns structural information of EHG; (3) Heterogeneous Dimension Attention Layer captures risk-related information regarding different dimensions; (4) Risk Merging Layer fuse former risk information to generate final risk representations; (5) Classifier gives the final assessed risk level.
• Figure 3: Illustration of the difference between entity-level attention and dimension attention.
• Figure 4: Distribution Shift. Following the message passing procedure intrinsic to GNNs, the revised node attributes exhibit enhanced distinguishability.
• Figure 5: Visualization of node representations for different GNN models on ECAD. Green points, red points, and black points denote enterprises with "AAA", "AA" and "Others" credit levels, respectively. We can observe that GDAN outperforms baseline models.

Theorems & Definitions (2)

• Definition 1
• Definition 2