Chewing gum is frequently found as physical evidence at crime scenes, necessitating an efficient method for gum category identification. The spectral data were normalized, followed by Principal Component Analysis for dimensionality reduction, selecting the top 100 principal components that accounted for a cumulative explained variance of 92.32%. After preprocessing, the entire dataset was divided into a 70% training set and a 30% test set. These sets were input into three machine learning models—Random Forest, Support Vector Machine, and K-Nearest Neighbors-combined with the Bayesian Optimization. After 100 iterations, the three models obtained optimal hyperparameter combinations, achieving classification accuracies of 98.03%, 88.72%, and 89.21%, respectively. Notably, the Bayesian Optimization - Random Forest model exhibited the highest classification accuracy, reaching 98.03%. K-fold Cross-Validation was subsequently applied to evaluate the classification accuracy and stability of the Bayesian Optimization-Random Forest model.