Machine learning model interpretability using SHAP values: Applied to the task of classifying and predicting the nutritional content of different cuts of mutton
Li Wang, Xuchun Sun, Jing Liang, Zhiyuan Ma, Fei Li, Shengyan Hao, Baocang Liu, Long Guo, Xiuxiu Weng
Abstract
, hindshank, and foreshank. An SVM-SHAP model predicted crude fat, protein, and fatty acids, while interpreting feature contributions. PUFA were significantly higher in the hindshank than in the longissimus lumborum and foreshank. The SVM model achieved a classification accuracy of 92.5 % and successfully predicted key nutritional parameters such as EE, CP, MUFA and PUFA with RPD values exceeding 2.7 in the test set. SHAP value analysis revealed that lipid-related variables and wavelengths in the 2300-2500 nm region were major contributors to the model. Vis-NIR-based SVM modeling technology is a fast, non-destructive, and accurate tool for evaluating fresh mutton.