A memetic algorithm for global induction of decision trees

Abstract

In the paper, a new memetic algorithm for decision tree learning is presented. The proposed approach consists in extending an existing evolutionary approach for global induction of classification trees. In contrast to the standard top-down methods, it searches for the optimal univariate tree by evolving a population of trees. Specialized genetic operators are selectively applied to modify both tree structures and tests in non-terminal nodes. Additionally, a local greedy search operator is embedded into the algorithm, which focusses and speeds up the evolutionary induction. The problem of over-fitting is mitigated by suitably defined fitness function. The proposed method is experimentally validated and preliminary results show that the proposed approach is able to effectively induce accurate and concise decision trees. © Springer-Verlag Berlin Heidelberg 2008.