Supplementary material for Paper "Dissecting the Landscape of Multiobjective Feature Selection"

Content Supplementary materials (detailed results and figures) Reference Anonymous. Dissecting the Landscape of Multiobjective Feature Selection (under review) Abstract Feature selection is essential for improving the predictive performance and interpretability of machine learning (ML) models across a wide range of prediction tasks. Wrapper feature selection can be formulated as a multiobjective pseudo-Boolean optimization problem that seeks to maximize the predictive performance of an ML model while minimizing the number of selected features. Although various evolutionary multiobjective optimization algorithms have been proposed for feature selection, the fitness landscape of this problem remains poorly understood. To address this gap, this paper analyzes the landscape of multiobjective feature selection across fourteen classification datasets, six ML models, and six combinations of objectives. Specifically, we investigate the exact landscape characteristics of the resulting 504 problem instances by exhaustively enumerating all feasible feature subsets. We further examine how algorithm performance (with Pareto local search) relates to multiobjective landscape characteristics, thereby shedding light on the main challenges of wrapper multiobjective feature selection problems.