Improving Performance Factors of an Imbalanced Credit Risk Dataset Using SMOTE

In this paper, a supervised machine learning approach for predicting the credit risk of a loan applicant has been presented. Due to the imbalance of real-world datasets, often high accuracy but low precision, recall, and F1-scores is obtained. This paper aims to solve these issues by using oversampling technique called Synthetic Minority Over-sampling Technique (SMOTE). Lending Club, a real-world imbalanced dataset has been considered to train the models. The dataset consists of 151 features among which 113 columns are numerical and the rest are categorical. It had a total of 2260701 instances. SMOTE is applied to the dataset to make it balanced improving accuracy, precision, recall, and F1-score. Four supervised classifier algorithms such as Logistic Regression, XGBoost, Decision Tree, and Random Forest were used to predict the loan defaulters. Among them, Random Forest achieves the highest accuracy of 87% with precision, recall, and F1-score of 0.81, 0.97, and 0.89 respectively. Also, the ROC-AVC score of Random Forest was 0.93.