In current Baijiu extraction processes, the classification of base Baijiu grades is primarily performed using sensory evaluation, and the method is hampered by low detection efficiency and susceptibility to subjective influences. Therefore, near-infrared spectroscopy is applied to base Baijiu grade detection, and the feasibility of using the Shapley additive explanation (SHAP) algorithm from interpretable artificial intelligence for selecting characteristic spectral points is explored. It was found that when the number of features was 36, an accuracy of 97.08% was achieved by the LightGBM predictive model. To further improve model performance, a hybrid strategy combining interval partial least squares (iPLS) with SHAP was proposed, and an accuracy of 99.27% was achieved by the LightGBM model when the number of features was 9. Analysis of the spatial distribution of iPLS interval partitioning and SHAP contribution values indicated that the ranking of SHAP contributions does not strictly correspond to predictive performance. That model's performance can be improved by carefully designing feature selection strategies.