Abstract:Selecting reliable predictors has always been crucial in classification. Especially decision trees are very popular for solving supervised variable selection and classification problems. When variable selection has to be performed with regard to acquisition costs, which have to be paid whenever the respective variable is extracted for a new observation, the problem of balancing the predictive power of the model against its costs describes a multi-objective optimization problem which can be solved with meta-heuristics such as evolutionary multi-objective algorithms. In this paper, we present a non-hierarchical evolutionary multi-objective tree learner (NHEMOtree) based on genetic programming using a binary decision tree representation to handle multi-objective optimization problems with equitable optimization criteria. This tree learner is applied to a multi-objective classification problem from medicine as well as to simulated data to evaluate its performance relative to two wrapper approaches based on either NSGA-II or SMS-EMOA with bitstring representation and CART as the enclosed classification algorithm. Moreover, a novel crossover operator based on a multi-objective variable importance measure is introduced. Using this crossover operator, NHEMOtree can be improved.

A novel crossover operator based on variable importance for evolutionary multi-objective optimization with tree representation