Empowering Simultaneous Feature and Instance Selection in Classification Problems through the Adaptation of Two Selection Algorithms

Abstract

This paper proposes a new approach to data selection, a key issue in classification problems. This approach, which is based on a feature selection algorithm and one instance selection algorithm, reduces the original dataset in two dimensions, selecting relevant features and retaining important instances simultaneously. The search processes for the best feature and instance subsets occur separately yet, due to the influence of features in the importance of instances and vice versa, they bias one another. The experiments validate the proposed approach showing that this existing relation between features and instances can be reproduced when constructing data selection algorithms and that it leads to a quality improval comparing to the sequential execution of both algorithms.