An enhancement of support vector machine for credit risk assessment applied in loan application

ABSTRACT: This study focuses on enhancing the Support Vector Machine (SVM) for credit risk classification by addressing limitations in training time, and data imbalance. Traditional SVM works by identifying the optimal hyperplane that linearly separates two data classes, but it can be resource-intensive and less effective with imbalanced data. To improve performance, the researchers proposed three key enhancements: improving the Quadratic Programming (QP) optimization using Dual Coordinate Descent (DCD) to reduce training time, applying the Adaptive Synthetic (ADASYN) technique to address data imbalance, and using Bayesian Optimization (BO) for hyperparameter tuning. These enhancements were tested on datasets of varying sizes: 500, 1k, 10K, 15k, and 30k samples. The DCD method significantly reduced training time across all sample sizes, averaging a 93.84% decrease (117.96 seconds). ADASYN improved classification accuracy by generating synthetic data for minority classes, effectively addressing imbalance, BO enabled optimal by hyperparameter selection, further increasing accuracy. The optimized “C” values for the linear kernel ranged from 0.1160 to 0.4031 across the datasets. Overall, the enhanced SVM----combining DCD, ADASYN, and BO---outperformed traditional SVM, showing a 7% improvement in accuracy, an 82.78% reduction in training time, and an 8% increase in precision. This comprehensive approach demonstrates that the improved SVM model offers a more accurate and efficient solution for credit risk analysis, making it a highly suitable tool for this application.

Filipiniana