Integrating volatile organic compound profiles with machine learning algorithms to construct a discrimination model for jiangxiangxing baijiu

This study developed a targeted and efficient strategy for classifying Jiangxiangxing baijiu （JXXB）, specifically JXXB (daqu) (JXXB-D), JXXB (others) (JXXB-O), and edible alcohol-blended liquor (EABL). The methodology integrated chemical profiling of volatile compounds with machine learning algorithms. The concentrations of 82 volatile compounds were determined and converted to odor activity values (OAV) to better reflect their actual classification contributions. Random Forest, CatBoost, and Naïve Bayes classifiers were employed, all of which demonstrated excellent classification performance using the full component dataset and their OAV. To streamline the method, multi-model feature importance evaluation was applied, successfully identifying 20 core components (including alcohols, ethers, pyrazines, aldehydes/ketones, and esters). A simplified model based solely on these 20 markers retained classification metrics comparable to the full-model, achieving high accuracy (＞0.95), precision(＞0.95), and AUC (＞0.95) values. The distribution patterns of these key components aligned with the theoretical expectations of different production processes. The established method requires the detection of only a limited number of components, making it a practical, cost-effective, and rapid solution suitable for quality control in small-to-medium enterprises and routine inspection by regulatory bodies, thereby addressing a significant need in the baijiu industry.