Discovery of decision rules from databases: An evolutionary approach

An EA-based system called EDRL, which for each class label sequentially generates a disjunctive set of decision rules in propositional form, which is evaluated by comparing its classification accuracy with that of C4.5 learning algorithm on six datasets from UCI repository.

Abstract

Decision rules are a natural form of representing knowledge. Their extraction from databases requires the capability for effective search large solution spaces. This paper shows, how we can deal with this problem using evolutionary algorithms (EAs). We propose an EA-based system called EDRL, which for each class label sequentially generates a disjunctive set of decision rules in propositional form. EDRL uses an EA to search for one rule at a time; then, all the positive examples covered by the rule are removed from the learning set and the search is repeated on the remaining examples. Our version of EA differs from standard genetic algorithm. In addition to the well-known uniform crossver it employs two non-standard genetic operators, which we call changing condition and insertion. Currently EDRL requires prior discretization of all continuous-valued attributes. A discretization technique based on the minimization of class entropy is used. The performance of EDRL is evaluated by comparing its classification accuracy with that of C4.5 learning algorithm on six datasets from UCI repository.