Classification is one of the important fields in data analysis. Generating concept-based (JSM) hypotheses is a well-known approach to this task. Although the accuracy of this approach is quite good, the coverage is often insufficient. In this paper a new classification approach is presented. The approach is based on the similarity of an object to be classified to the current set of hypotheses: it attributes the new object to the class that minimizes the set of new hypotheses when a new object is added to the training set. The proposed approach provides a better coverage in comparison with the classical approach.