AI-Driven Hybrid Graph-Ensemble Approach for Credit Risk Classification

Credit risk rating is crucial for financial institutions to assess default probabilities and manage lending risks effectively. While traditional machine learning (ML) and deep learning (DL) models offer predictive capabilities, they often lack the ability to capture complex inter-customer relationships and ensemble synergies. This study proposes a hybrid AI-driven framework that integrates Random Forest (RF), Social Network Analysis (SNA), and Graph Convolutional Networks (GCNs) within a stacking classifier. The SNA component enriches the model with structural insights using centrality measures—PageRank, eigen-vector, betweenness, closeness, and degree centrality—which quantify borrower influence, connectivity, brokerage roles, proximity to others, and overall network integration within the credit ecosystem. The stacking classifier strategically combines multiple base learners to capture diverse data perspectives, reduce prediction variance, and improve resistance to overfitting, resulting in a more stable and robust risk assessment. To ensure transparency, the model includes feature attribution techniques that quantify the exact impact each input feature has on a given prediction and localized approximation methods that explain how the model arrives at decisions for specific individuals. This dual-layer inter-pretability framework empowers stakeholders with both global and instance-level 1 insights into the decision-making process. Extensive experimentation shows that integrating centrality-based relational features enables the model to distinguish between direct and indirect risk influencers, detect potential default cascades, and prioritize high-risk profiles based on network position. Combined with ensemble-level learning, this approach delivers enhanced precision, context-aware risk stratification, and actionable interpretability—making it highly suitable for deployment in real-world financial systems where both trust and performance are non-negotiable.