Optuna-optimized boosting models for predicting quality traits in multiple juice types using NIRS: Interpretability analysis via SHAP

Near-infrared spectroscopy (NIRS) is a potential rapid and reagent-free technique for assessing the quality of fruit juices. However, most existing models focus on single juice type and rely on linear algorithms such as partial least squares regression (PLSR), which are often inadequate for handling...

Full description

Bibliographic Details
Authors: Ding, Fangchen, Rivero Delgado, Miguel Ángel, Zha, Rili, García-Martín, Juan Francisco
Format: article
Status:Published version
Publication Date:2026
Country:España
Institution:Universidad de Sevilla (US)
Repository:idUS. Depósito de Investigación de la Universidad de Sevilla
OAI Identifier:oai:idus.us.es:11441/180177
Online Access:https://hdl.handle.net/11441/180177
https://doi.org/10.1016/j.foodcont.2025.111878
Access Level:Open access
Keyword:Boosting models
Hyperparameter optimization
Juice quality
Near-infrared spectroscopy (NIRS)
Rapid analysis
SHapley additive ex
Planations (SHAP)
Description
Summary:Near-infrared spectroscopy (NIRS) is a potential rapid and reagent-free technique for assessing the quality of fruit juices. However, most existing models focus on single juice type and rely on linear algorithms such as partial least squares regression (PLSR), which are often inadequate for handling the nonlinear and heterogeneous characteristics of diverse juice matrices. To address this challenge, this study developed boosting models optimized by Optuna, including XGBoost, AdaBoost, and CatBoost, to predict four key quality traits, namely acidity, total phenolic compounds (TPC), total flavonoid content (TFC), and vitamin C across 4 types of fruit juice. The boosting models consistently outperformed PLSR, particularly for acidity, TPC, and vitamin C, achieving Rp2 values above 0.95 and RPD values exceeding 4.93. SHAP-based interpretability analysis further revealed that, in addition to typical NIRS absorption bands such as 1163 nm, 1169 nm, and 1193 nm located within the 1150–1210 nm region, non-classical regions including 1104 nm and several wavelengths between 1264 and 1322 nm also contributed positively to the model outputs. This demonstrates the capacity of boosting algorithms to capture informative spectral features from non-classical regions that are often overlooked by traditional linear models. Overall, this study demonstrates the value of combining automated hyperparameter optimization with interpretable machine learning, offering a robust and scalable framework for high-throughput, non-invasive quality control in the juice industry by NIRS.