Beyond ASP: An Evolutionary Multiobjective Decision Tree Framework for Binary Classification Tasks

The Algorithm Selection Problem (ASP) recommends the most appropriate algorithm for a particular instance. The EMODT-ASP framework, based on decision trees and a multiobjective approach, was originally proposed for ASP, with a focus on interpretability. We extended this framework to binary classification, adapting the evaluation function to use AUC, incorporating support for categorical features, and limiting the depth of trees to 7 levels to ensure interpretability. We evaluated our approach in 26 binary bases, obtaining a mean AUC of 0.792. A sensitivity study revealed that the original configuration of the parameters is robust and that shallow trees can be overfitted with excessive generations, and this effect is mitigated by higher minimum thresholds of instances per leaf. To improve generalization, we suggest automatic feature creation and cross-validation during training. We also want to extend our approach to multiclass and regression problems.