<![CDATA[The People's Business Credit (KUR) program is vulnerable to Non-Performing Loan (NPL) risks [1]. Conventional credit scoring often fails to objectively map the complex financial variables of customers. This research aims to analyze the performance of the Support Vector Machine (SVM) algorithm in predicting KUR payment feasibility by comparing four kernel functions: Linear, Radial Basis Function (RBF), Polynomial, and Sigmoid. The dataset comprises 398 financial records from Bank XYZ with an extreme class imbalance (80:20). The model was trained using 9 initial financial capacity features, Z-score standardization, and GridSearchCV optimization [9]. Performance evaluation focused on the Confusion Matrix to detect the Accuracy Paradox on imbalanced data [6]. The results showed that the Polynomial Kernel achieved the highest Accuracy (73.75%) and F1-Score (84.67%) but suffered from the Accuracy Paradox by failing to detect the minority class, misclassifying 14 Default customers as Eligible. Conversely, the RBF Kernel (C = 0.1, γ = 2) demonstrated the best and most rational performance, balancing the margin to correctly detect 6 Default customers while maintaining high precision for the Eligible class.]]>
Copyrights © 2026
