What question did this study set out to answer?

To improve predictive performance on small-sample industrial data through a novel virtual sample generation technique.

May 16, 2026Journal of King Saud University - Computer and Information Sciences0 citationsOpen Access

Enhancing small data prediction performance via Bézier-interpolated virtual sample generation with bagging U-net GAN

LLLiang-Sian LinNational Taipei University of Nursing and Health Science HCHui-Chi ChuangNational Cheng Kung University YCYi-Chung ChengTainan University of Technology

Key Points

To improve predictive performance on small-sample industrial data through a novel virtual sample generation technique.
Developed Bézier-Unet-GAN as a generative adversarial network (GAN) for virtual sample generation.
Utilized a manifold algorithm for feature extraction from industrial datasets.
Conducted experiments on six datasets comparing performance against eight state-of-the-art VSG techniques.
Proposed method improved MAE by 11.928%, MAPE by 19.067%, and RMSE by 10.442% for a back propagation neural network at a sample size of 10.
Bézier-Unet-GAN demonstrated superior performance across MAE, MAPE, and RMSE metrics compared to current VSG techniques.
Statistical tests confirmed significant improvements in prediction accuracy with the Bézier-Unet-GAN method.

Abstract

In recent years, industrial context deep learning algorithms have been widely applied to support decision-makers. However, in early-stage decision support systems, due to limited data collection periods and demanding experimental conditions, decision-makers often face the dilemma of data scarcity and insufficient feature representation. Moreover, with increasing complexity in industrial processes, it becomes more challenging to extract representative information from collected data exhibiting high dimensionality and non-linear characteristics. As a result, achieving strong predictive performance is difficult with conventional deep learning algorithms. To tackle this issue, we propose a novel virtual sample generation (VSG) technique, termed Bézier-Unet-GAN, an auto-encoder U-net generative adversarial network (GAN) model designed to enhance prediction accuracy on small-sample industrial data. The proposed method aims to produce high-quality virtual samples resembling original data. Additionally, we employ a manifold algorithm to distill representative features from the original data. Based on the features, we create Bézier interpolation point and calculate its membership function (MF) to train a bagging-based U-net generator. To evaluate efficacy of the proposed approach, we conduct experiments on six industrial datasets. Compared with eight state-of-the-art VSG techniques, the proposed approach achieves promising results in terms of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) evaluation metrics. For example, at the sample size of 10, for a back propagation neural network (BPNN) predictive model, the proposed method improves the metrics by 11.928%, 19.067%, and 10.442%, respectively. Additionally, we use the Siegel-Castellan post hoc test and Wilcoxon test to elucidate whether prediction accuracy using the Bézier-Unet-GAN method significantly outperforms the other eight VSG techniques across MAE, MAPE, and RMSE metrics. The experimental results indicate that the Bézier-Unet-GAN method effectively improves the performance of the predictive model on small sample sets while consistently outperforming these state-of-the-art VSG techniques. Overall, the proposed approach delivers superior predictive performance on limited industrial datasets relative to the compared methods.

问 AI

Bookmark

View Full Paper