Impact of Data Pre-Processing Techniques on Supervised Machine Learning Models for Credit Risk Classification
Abstract
Credit risk classification is a critical task to reduce default risk and support informed lending decisions. Nonetheless, the execution of machine learning in the predicting credit risks is greatly influenced by characteristics of the data including data heterogeneity, variance of the feature scales, and class imbalance. This paper investigates thoroughly the effect of data pre-processing approaches on supervised machine learning models to address credit risk classification problems. Specifically, well-known classifier namely Logistic Regression and Decision Tree, Random Forest and Support Vector Machine and K-Nearest Neighbor will be assessed by applying various data pre-processing techniques on a benchmark credit dataset. To do so, feature scaling and class imbalance handling via SMOTE technique will be conducted in a controlled manner, and their influences on classifier performance will be observed separately as well as together. Model performances will be measured based on accuracy, recall, precision, F1-score, and the area in ROC curve metrics. Experimental results show that feature scaling helps to enhance the performance of distance-based classifiers without influencing tree-based models. On the other hand, handling class imbalance yields inconsistent results, implying that data oversampling approaches depend highly on machine learning classifiers and must be applied wisely.