Searching the forest: using decision trees as building blocks for evolutionary search in classification databases

Abstract

A new evolutionary search algorithm, called BGP (Building-block approach to Genetic Programming), to be used for classification tasks in data mining, is introduced. It is different from existing evolutionary techniques in that it does not use indirect representations of a solution, such as bit strings or grammars. The algorithm uses decision trees of various sizes as individuals in the populations and operators, e.g. crossover, are performed directly on the trees. When compared to the C4.5 and CN2 induction algorithms on a benchmark set of problems, BGP shows very good results.