To address key challenges in credit risk prediction, including the instability of highdimensional feature selection under sample perturbations, limited structural interpretability and difficulties in ensuring model robustness, this paper proposes a robust hybrid structural learning (Ro-HCD) framework for feature selection. By integrating the Spirtes–Glymour noncommutative tensor exponential approach to regulatory structure (PC-NOTEARS) structural learning, adaptive transformation strategies tailored to financial data characteristics, domain knowledge constraints and a multilevel stability evaluation mechanism, the proposed framework identifies risk-related features that exhibit robust dependency patterns and improved interpretability under complex credit conditions. Unlike conventional feature selection methods that primarily rely on statistical correlations, Ro-HCD improves the consistency of selected features under data variations and structural perturbations by combining structural constraints with perturbation-based validation. We note that the structures learned in this study represent stable dependencies under specific modeling assumptions and conditions rather than definitive evidence of underlying causal mechanisms. Extensive experiments on multiple publicly available credit risk data sets demonstrate that Ro-HCD effectively reduces feature dimensionality while maintaining competitive predictive performance and improving the robustness of complex nonlinear models under perturbed conditions. The results suggest that the proposed framework offers an effective solution for credit risk modeling by achieving a balance between predictive accuracy, structural stability and model interpretability.