[Paper Review] Every Corporation Owns Its Structure: Corporate Credit Ratings via Graph Neural Networks

Bojing Feng, Haonan Xu+2 more|arXiv (Cornell University)|Nov 27, 2020

Financial Distress and Bankruptcy Prediction25 references4 citations

TL;DR

This paper proposes CCR-GNN, a novel graph neural network that models each corporation as a graph of interrelated financial features, enabling explicit learning of intra-corporate feature interactions. By constructing individual graphs via self-outer product and applying graph attention networks, CCR-GNN outperforms state-of-the-art methods on a Chinese corporate credit rating dataset, achieving an F1-score of 0.95177.

ABSTRACT

Credit rating is an analysis of the credit risks associated with a corporation, which reflects the level of the riskiness and reliability in investing, and plays a vital role in financial risk. There have emerged many studies that implement machine learning and deep learning techniques which are based on vector space to deal with corporate credit rating. Recently, considering the relations among enterprises such as loan guarantee network, some graph-based models are applied in this field with the advent of graph neural networks. But these existing models build networks between corporations without taking the internal feature interactions into account. In this paper, to overcome such problems, we propose a novel model, Corporate Credit Rating via Graph Neural Networks, CCR-GNN for brevity. We firstly construct individual graphs for each corporation based on self-outer product and then use GNN to model the feature interaction explicitly, which includes both local and global information. Extensive experiments conducted on the Chinese public-listed corporate rating dataset, prove that CCR-GNN outperforms the state-of-the-art methods consistently.

Motivation & Objective

• To address the limitation of existing graph-based models that treat corporations as nodes without modeling internal feature interactions.
• To improve corporate credit rating prediction by explicitly capturing high-order feature interactions within each corporation.
• To propose a new graph-level modeling approach where each corporation is represented as a graph of its own financial features.
• To demonstrate the effectiveness of graph neural networks in modeling intra-corporate feature dependencies for credit risk assessment.
• To provide a new perspective for applying GNNs in financial risk modeling by shifting from inter-firm networks to intra-corporate feature graphs.

Proposed method

• Construct individual graphs for each corporation using the self-outer product of financial features to model pairwise feature relations.
• Represent each corporation as a graph where nodes are features and edges represent the strength of interaction between features.
• Apply Graph Attention Networks (GAT) with multi-head attention to aggregate information across feature nodes, capturing both local and global dependencies.
• Use three graph attention layers with attention channels set to {8, 64, 9}, employing mean pooling for first two layers and max pooling for the last.
• Optimize the model using Adam with an initial learning rate of 0.001 and L2 regularization of 0.00001.
• Apply SMOTE for data augmentation to address class imbalance in the multi-label credit rating dataset.

Experimental results

Research questions

• RQ1 Can modeling intra-corporate feature interactions through a graph structure improve corporate credit rating prediction?
• RQ2 How does the proposed CCR-GNN model compare to traditional machine learning and deep learning baselines in credit rating performance?
• RQ3 To what extent do graph attention mechanisms enhance the modeling of high-order feature interactions in financial data?
• RQ4 Does representing each corporation as a graph of its own features lead to better generalization and robustness in credit risk assessment?
• RQ5 Can graph neural networks effectively capture both local and global feature dependencies in corporate financial profiles?

Key findings

• CCR-GNN achieves the highest F1-score of 0.95177 on the Chinese public-listed corporate rating dataset, outperforming all baseline methods.
• The model's performance surpasses XGBoost, the second-best baseline, which achieved an F1-score of 0.94133.
• CCR-GNN demonstrates superior recall (0.93437) and accuracy (0.95012), indicating strong performance across all credit rating classes.
• The use of graph attention mechanisms enables effective learning of high-order feature interactions, contributing to improved predictive performance.
• The proposed corporation-to-graph transformation effectively captures internal feature relations, validating the model's architectural innovation.
• Extensive ablation and comparison confirm that the graph-level modeling approach significantly enhances credit rating prediction beyond node-level graph models.