Abstract: In order to improve mechanical and blind prediction behavior of some built models, a classifier prediction control algorithm was designed with flue-cured tobacco and burley tobacco in different producing areas as experimental samples. It had the characteristic of rejecting recognition and credibility analysis through integrating several theories and methods including hypothesis testing, convex hull, interior point analysis and sequence random testing. The results demonstrated that classifier could effectively accept test sample set and give predictive values and reliability reference values of test data in convex hull. In the meanwhile, classifier could also accurately reject burley tobacco sample and special type flue-cured sample, which was different from flue-cured sample set. The feasibility and validity of classifier were verified through different type of testing data, especially the practicality of cigarette sensory evaluation was based on expert experience or domain knowledge.