A decision-tree approach to the discovery of empirical regularities

One of the standard problems of data mining is the discovery of empirical regularities in large amounts of data. In pattern recognition terminology, empirical regularities are known by the name Non-Reducible Descriptors (NRDs), especially when Boolean variables are used in pattern descriptions. In this paper, a decision-tree model for the construction of an empirical regularity or an NRD is presented. A computational procedure for the construction of all empirical regularities (or NRDs) for a given pattern is developed using this model. The procedure is illustrated with an application to the pattern recognition problem of recognizing Arabic numerals. Finally, an analysis of the computational complexity of the procedure is also presented.