A method for predicting the aroma type of cigarette recipe module based on near-infrared spectral feature dimensionality reduction was proposed to classify and identify the aroma type of cigarette recipe modules with near-infrared spectroscopy. The near-infrared spectral data of 238 cigarette recipe module samples from 2017 to 2019 were selected to construct an aroma prediction model based on feature variables through combining the recursive feature elimination method in feature engineering and three machine learning techniques including BP neural network， random forest and XGBoost. Compared with the classification effect of full spectrum data training， the spectral feature variables filtered by recursive feature elimination method effectively improved the recognition accuracy of aroma type of cigarette recipe module. Among them， the algorithm of XGBoost had the best classification performance， with a model recognition accuracy of 90.41% for the test set. It is indicated that the prediction method of aroma type based on the recursive feature elimination of near-infrared spectral features has a certain role in assisting decision-making in the rapid positioning， scientific evaluation and cigarette formulation design of cigarette recipe modules.