An uncertainty quantification and accuracy enhancement method for deep regression prediction scenarios

Abstract

Accurate regression prediction is a critical objective in industry; however, epistemic and aleatoric uncertainties can significantly impact prediction accuracy. Existing research primarily focuses on either prediction or uncertainty estimation, with limited studies addressing joint prediction methods and accuracy enhancement. This paper proposes an uncertainty quantification and accuracy enhancement method, referred to as UQAE, for deep regression prediction scenarios. This approach enables the simultaneous provision of both point predictions and uncertainty estimation results for any deep regression task. Furthermore, it facilitates a quadratic improvement in point prediction accuracy. Specifically, threshold monitoring and parameter-sharing asynchronous training strategies are implemented to ensure the joint prediction of point estimates and distribution-free intervals. Fuzzy rules are incorporated to enhance the accuracy of point predictions, providing interpretability based on the joint predictions. The proposed method is rigorously compared and evaluated using nine generalized manufacturing-related datasets, demonstrating significant improvements in both point prediction accuracy and uncertainty prediction interval estimation. Additionally, this approach is expected to advance regression prediction research in manufacturing, promoting higher accuracy and interpretability. The UQAE method and the associated datasets are available at https://github.com/ZhangTeng-Hust/UQAE.