An AutoML framework for near-infrared spectral analysis in food quality assessment and origin traceability

Near-infrared (NIR) spectroscopy has emerged as a rapid and non-destructive analytical technique for food quality assessment; however, its predictive performance strongly depends on the appropriate combination of spectral preprocessing, wavelength selection, and modeling methods. In this study, a complexity-aware automated machine learning (AutoML) framework was developed for NIR spectral analysis to systematically construct compact and high-performing modeling pipelines. The framework integrates preprocessing optimization, wavelength subset selection, model evaluation, and complexity-aware post-selection within a unified workflow, enabling efficient exploration of candidate pipelines with varying structural complexity. Extensive experiments were conducted on multiple datasets covering both regression and classification tasks. The results demonstrate that the framework consistently identifies lightweight yet effective pipelines, achieving competitive or superior predictive performance compared with conventional manually designed approaches. Furthermore, the selected pipelines exhibit good generalization ability while maintaining relatively low computational complexity, highlighting the potential of well-optimized traditional machine learning approaches for practical NIR spectral modeling tasks. In addition, SHAP-based interpretability analysis revealed that the selected wavelengths were consistently concentrated within chemically meaningful spectral regions associated with characteristic absorption bands. Overall, this study provides a practical and interpretable AutoML-based framework for NIR spectral modeling, offering methodological and application value for rapid and non-destructive food quality evaluation.