Fraud Detection in Credit Risk Assessment Using Supervised Learning Algorithms

Abstract

Our study systematically evaluates the performance of various supervised learning algorithms in credit risk assessment and fraud detection, including Logistic Regression, Decision Tree, Support Vector Machine, Random Forest, Gradient Boosting Tree, and Neural Network. The results show that in credit risk assessment, the Gradient Boosting Tree performed best with an accuracy of 90.5% and a ROC-AUC of 0.84, followed by Random Forest and Neural Network, with accuracies of 89.2% and 88.8%, and ROC-AUCs of 0.82 and 0.81, respectively. In the fraud detection task, the Neural Network performed best with an accuracy of 97.5% and a ROC-AUC of 0.88, while Gradient Boosting Tree and Random Forest achieved accuracies of 97.1% and 96.3%, and ROC-AUCs of 0.87 and 0.85, respectively. Feature importance analysis indicates that repayment history, credit limit, bill amount, and repayment amount are key features in credit risk assessment, while transaction amount, transaction time, and location are crucial for fraud detection. Data preprocessing and feature engineering played critical roles in enhancing model performance. Further optimization of model hyperparameters and addressing data imbalance issues will help improve model performance. In conclusion, ensemble learning methods and Neural Networks exhibit significant advantages in credit risk assessment and fraud detection. By employing scientific data preprocessing and feature engineering, combined with advanced machine learning algorithms, financial institutions can significantly enhance their risk management effectiveness.