Induction of Multiple Decision Trees Using Multiobjective Particle Swarm Optimization

The decision tree is a popular and widely-used classification model. The two main objectives in decision tree induction are accurate predictions for unseen instances and human comprehensibility. In this paper, we use multiobjective optimization for the evolution of decision tree classifiers that are efficient both with respect to classification accuracy and classifier complexity. Simpler decision trees are generally more comprehensible to humans at the expense of accuracy. We employ the Multiobjective Particle Swarm Optimization using Crowding Distance (MOPSO-CD) algorithm to evolve a population of decision trees that are optimal on two objectives: classification accuracy and classifier complexity based on the Minimum Description Length Principle. The validity and performance of this approach is evaluated on several commonly-used benchmark datasets. The results show that our approach is indeed effective in inducing multiple decision trees that are accurate and simple.