Evolving decision-tree induction algorithms with a multi-objective hyper-heuristic

Multi-objective optimization has been widely used in evolutionary computation for solving problems in which two or more conflicting objectives need to be optimized in a simultaneous fashion. This paper presents a multi-objective hyper-heuristic based on evolutionary algorithms that automatically designs complete decision-tree induction algorithms. Such algorithms are designed to generate decision trees that present an interesting trade-off between predictive performance and complexity. The proposed approach is tested over 20 UCI datasets, and it is compared with a single-objective hyper-heuristic as well as with traditional decision-tree induction algorithms. Experimental results show that the proposed approach can match the top predictive performance achieved by the baseline methods, without significant loss in model comprehensibility.