<![CDATA[Class imbalance in peer-to-peer (P2P) lending data can bias machine learning predictions toward the majority class, limiting the identification of loans that are not fully repaid. This study evaluates three resampling techniques: Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), and Conditional Tabular Generative Adversarial Networks (CTGAN), combined with four classifiers: Logistic Regression, Ridge Classifier, XGBoost, and LightGBM. The Lending Club dataset was divided into training and testing sets, with resampling applied only to the training data. Models were evaluated on the original, unresampled test set. Without resampling, LightGBM achieved an F1-score of 33.37%. LightGBM with ADASYN obtained the highest Recall (96.95%), while LightGBM with CTGAN achieved the highest F1-score (83.28%) and Accuracy (84.54%). These results demonstrate that resampling effectiveness depends on the classifier and evaluation metric. Statistical significance testing and synthetic data quality assessment were also conducted to examine result reliability. Overall, CTGAN combined with LightGBM provided the most balanced performance among the tested combinations, whereas ADASYN with LightGBM was more suitable for prioritizing minority-class detection. Further research should validate these findings using additional datasets.]]>
Copyrights © 2026
