SparseGraphSage: A Graph Neural Network Approach for Corporate Credit Rating

Si Shi, Wuman Luo, Rita Tse, Giovanni Pau

Abstract: Corporate Credit Rating (CCR) remains a critical research problem. In the past few decades, various machine learning approaches have gradually replaced and surpassed traditional labor-consuming manual checking. In particular, Graph Neural Networks (GNNs) have shown their capabilities and potential due to their strong power of processing non-Euclidean data. However, the current GNNs methods have two issues: 1) the proper design and construction of graphs; 2) the slow running speed and vast consumption of computing power. To address these issues, we propose a method named ‘SparseGraphSage’, which incorporates randomness in graph construction and integrates diffusion and sparse techniques in the GraphSage model. We design a stochastic edge selection process in the construction stage and diffusion matrices acting as operators in the graph layers. Through sufficient experiments and ablation study on two open-source CCR datasets, we demonstrate that our method exceeds the current state-of-the-art GNNs baselines in performance and is proven efficient.