The invention discloses a feature selection method and a traditional Chinese medicine main-symptom selection method based on feature groups. The feature selection method based on the feature groups comprises the following steps: 1, screening an original feature set; 2, utilizing a feature clustering algorithm to cluster the screened feature set to obtain the corresponding feature groups; 3, introducing a latent variable to each feature group to obtain a corresponding latent class model (LCM), and calculating correlation between the latent variables and labels; 4, sorting the feature groups sequentially with decreasing according to the correlation between the latent variables and the labels; 5, adding the sorted feature groups into selected feature subsets in turn, and establishing a Bayesian network with the latent variables; and 6, calculating classification accuracy rates of the Bayesian network, then obtaining a curve of the numbers of the added feature groups and the classification accuracy rates, and obtaining the corresponding optimal feature subset through judging the convergence of the curve or the highest accuracy rate. According to the method, the feature groups are used as selected targets, and the feature groups composed of a plurality of features have better representation capacity for original data.