A hybrid multi-layered ensemble model based on heterogeneous information network for small and medium-sized enterprise default prediction

For small and medium-sized enterprises (SMEs) with relatively limited financial and disclosure information, constructing networks using external relational data creates new opportunities to enhance default prediction of SMEs. To address the limitations of existing network-based approaches, which are typically limited to single-type structures and fail to fully mine the network risk information, this paper proposes an SME default prediction method based on a heterogeneous information network. Specifically, we construct the Small and Medium-sized Enterprise Relationship Network (SMERN), which integrates multiple types of relationships between enterprises and people, and extract structured network features based on SME credit business expertise to represent network-linked risk information. Furthermore, we design a hybrid model that integrates a multi-layered gradient boosting decision tree (GBDT) component to learn structured network features, and a graph neural network (GNN)-based label propagation module to capture dynamic network contagion risks. A joint optimization mechanism is designed to enable information exchange and fusion between the multi-layered GBDT and GNN components, allowing the model to better capture interactions among different risk signals. Empirical results on real-world data show that SMERN significantly improves SME default prediction, and the proposed model achieves superior performance by effectively exploiting network information in a comprehensive manner.