Deep learning regression model based on data pairing and pseudo-label fusion for NIR predictive modeling in food and...

Near-infrared (NIR) spectroscopy combined with artificial intelligence (AI) is widely used in food and pharmaceutical industries for rapid analysis, but its reliability is limited by small labeled datasets due to the difficulty of measuring sample chemical compositions. This study proposes DeepSS, a novel deep learning-based chemometrics approach, to address this challenge. DeepSS employs a pyramid encoder to extract low-dimensional latent features from NIR spectra and introduces a data pairing strategy to augment training data by combining latent features of sample pairs and subtracting their labels. During testing, pseudo labels are generated by pairing test samples with training features, followed by a decision strategy for final predictions. Benchmarked against four models using datasets from tablets, apples, and American ginseng, DeepSS reduced prediction errors by 1.51 %-14.29 % for key components. DeepSS demonstrates reliable performance on small labeled NIR datasets, improving prediction accuracy and efficiency in food and pharmaceutical analysis.