Risk Score Estimation and Features Ranking Using Regression based on Deep Learning Models

Abstract

This study explores the effectiveness of regression-based deep learning models for risk score estimation and feature ranking, comparing them with traditional machine learning approaches. A comprehensive meta-analysis and computational experiments were conducted on different datasets to assess the model performance. The study indicates that deep learning models, especially transformer-based architectures, can predict more accurately (R² = 0.95 and mean squared error (MSE) = 0.007), compared to more tried and traditional methods such as linear regression and random forest. Transformer models also achieve more refined and interpretable feature rankings, which help decision-making in high-risk domains such as finance and healthcare. While such advantages exist, computational efficiency and interpretability remain challenging and require further optimization techniques. This research demonstrates the likelihood of deep learning in risk assessment and reinforces the need for future enhancements to increase the scalability and real-world applicability.