Abstract
Credit risk assessment for agricultural small and medium-sized enterprises (SMEs) is challenging due to information asymmetry, rendering traditional models ineffective. We propose a novel knowledge-informed neural network to assess credit risk in this data-scarce environment. Our two-stage model first uses Fuzzy Analytic Hierarchy Process (Fuzzy AHP) to quantify judgments from a survey of 202 stakeholders into risk scores, which serve as learning labels. Subsequently, after using Principal Component Analysis (PCA) for dimensionality reduction, a neural network is trained to replicate the experts' decision logic. The AHP analysis identified 'Financial Status' and 'Development Planning' as the most critical risk themes. On an independent test set, the trained network showed exceptional predictive accuracy, achieving a Root Mean Squared Error (RMSE) of approximately 0.0100. This study validates that fusing expert knowledge with machine learning is a highly effective approach, providing an operational decision-support tool and a new paradigm for AI applications in information-asymmetric environments.