A Dynamic Cost-Adjusted AdaCost Model for Credit Prediction of Smallholder Farmers

Abstract

Data imbalances constrain credit prediction models. This study proposes a dynamic cost-adjusted AdaCost credit prediction model to improve prediction accuracy. Our model enriches existing research methods on the credit problems of smallholder farmers and opens new avenues for establishing credit prediction models for smallholder farmers. The present study is characterized by two distinctive research innovations: Firstly, it introduces a variable cost-sensitive function that enables adaptively adjusted misclassification costs for each sub-model generated by the AdaCost framework. This advancement effectively addresses the inherent limitation of static cost-sensitive function values in conventional AdaCost models. Secondly, the dynamic nature of the proposed cost-sensitive function induces corresponding variations in both sample weights and model weights within interconnected subsequent sub-models. The mechanism fundamentally resolves the critical oversight in traditional AdaCost methodologies that failed to account for the dynamic interdependencies between cost-sensitive functions and weight adaptation processes. Using data on agricultural loans from a commercial bank in China as empirical data and comparing them with seven baseline models, including AdaBoost, AdaCost, Cost RF, Cost XGBoost, Cost SVM, Cost GBDT, and Cost DT, we found that the proposed dynamic cost-adjusted AdaCost model outperformed the other models. Robustness tests were conducted using two publicly available loan datasets from UCI. They showed that the dynamic cost-adjusted AdaCost model performed better than the AdaBoost and AdaCost models.