This paper introduces a probabilistic model of two-class pattern recognition. The measurable sets are defined by a similarity, which is a reflexive and symmetric binary relation. The heuristic information model is formulated by a type of data clustering called representative clustering. The heuristic information about a data record is a data subset containing the record, which is computed by comparing the record with all representative records. For the corresponding classifiers, both Bayes and Neyman-Pearson Theorems are proved in this paper. In application, the knowledge discovering process searches for similarity and representative clustering in a training data set. The evaluation is extended to records in a testing data set. The experiment shows the trade-off between the number of representatives and classifier performance.