Research on grade identification of Chinese strong-flavor raw liquors based on machine learning under feature selection

Using different grades of strong-flavor raw liquors as the research object, the trace components in the original spirits were analyzed qualitatively and quantitatively by gas chromatography-mass spectrometry (GC-MS), the random forest (RF) feature importance combined with sequence forward selection (SFS) method was used for feature selection, and the selected feature subsets were substituted into random forest, support vector machine (SVM) and extreme gradient boosting (XGBoost) to establish the grade identification model of strong-flavor raw liquors. The experimental results show that the optimal feature subset selected by RF+SFS in the SVM model is the first 17 features of the feature importance sequence, and its classification accuracy is 87.23%; the optimal feature subset selected by RF+SFS in the RF model is the first 26 features of the feature importance sequence, and its classification accuracy is 97.87%; the optimal feature subset selected by RF+SFS in the XGBoost model is the first 20 features of the feature importance sequence, and its classification accuracy is 100%. The optimal feature subset selected by RF+SFS in the XGBoost model was the first 20 features of the feature importance sequence, and its classification accuracy reached 100%. Studies have shown that the content of trace components in the raw liquars exhibits different trends in different grades. Therefore, by selecting the characteristic compounds with significant changes in different grades can effectively achieve the raw liquars grade identification and provide a new idea for the quality control of raw liquars.