Rule discovery from data using decision matrices

This thesis presents a new method for computing all maximally general rules for target classification. Maximally general rules are rules with a minimized number of conditions. The method is based on the rough sets model and the concepts of decision matrices and decision functions. It is confined to the generation of rules for objects expressed in terms of symbolic or categorical feature variables. Under the assumption that all numeric feature variables have been discretized by some previous process, the method presented here guarantees that rules are always discovered in their maximally general form. All maximally general rules hidden in data can be discovered by this method.;With this method, the problem of finding all maximally general rules is reduced to the problem of computing prime implicants of a group of associated Boolean expressions. The method is particularly applicable to the identification of all potentially interesting rules in a knowledge discovery system.;The performance of the method is evaluated in regard to its classification accuracy for several data sets. Some of the results show that it has higher classification accuracies than other algorithms.