This paper addresses the problem of corporate credit risk assessment under incomplete transaction data. Traditional models rely primarily on firm-level financial indicators, neglecting inter-company dependencies and latent financial relationships. We propose a graph-based framework for reconstructing economically meaningful structures from partial transaction networks. The approach employs multiplex graph representations and derives stable features capturing connectivity, flow concentration, and dependency patterns. The methodology is designed to remain robust under missing data, noise, and unobserved links. Empirical evaluation demonstrates that graph-derived features significantly improve predictive performance compared to traditional financial-ratio-based models. The proposed framework provides a practical and interpretable foundation for integrating network-based analytics into credit risk assessment in corporate banking.
Dmitry Kazakov (Tue,) studied this question.