Imbalanced Credit Risk Prediction in Ensemble Learning Classifiers: A Comparative Analysis of SMOTE, ADASYN, SMOTETomek, and Cluster Centroids

Abstract

The improvement of financial institutions' ability to predict customers' credit risk has benefited from the continuous updating of machine learning algorithms. The ensemble algorithms represented by Random Forest and XGBoost carry forward the advantages of decision tree structures and perform well on datasets with complex characteristics, such as credit data. Research methodology is conducted via sampling design; we employ the German Credit dataset, a benchmark dataset comprising 1000 samples and 18 features, for empirical analysis. Measurement Design: various performance metrics such as accuracy, precision, recall, and F1-score are used to assess the efficacy of the balancing techniques. Analysis Design: A comparative analysis is conducted to evaluate the strengths and weaknesses of these balancing techniques in different ensemble learning classifiers. However, in real credit datasets, the number of defaulted samples is usually only a small percentage. Class imbalance greatly weakens the predictive performance of ensemble models. Therefore, this paper employed four different techniques for handling class imbalance problems, namely SMOTE, ADASYN, SMOTETomek, and ClusterCentroid. A comparison of different tree-based models in datasets is demonstrated where balancing techniques are applied. The conclusions show that all models perform much better on the datasets with balancing techniques than without balancing. Balancing the data in advance does improve the predictive ability of the models, and the over-sampling and integrated sampling methods outperform the under-sampling techniques on small and medium-sized datasets.