An adaptive initialization and multitasking based evolutionary algorithm for bi-objective feature selection in classification

Evolutionary algorithms have become a widely-used approach for solving multi-objective optimization problems over the last decades, while feature selection in classification is also a discrete bi-objective optimization problem that aims at simultaneously minimizing the classification error and the number of selected features. However, traditional multi-objective evolutionary algorithms often encounter drawbacks when the total number of features grows to a large-scale level. Thus, in this paper, an adaptive initialization and multitasking based evolutionary algorithm, termed AIMEA, is proposed to tackle bi-objective feature selection in classification, especially for large-scale datasets. More specifically, an adaptive initialization mechanism based on a set of task-related subpopulations is set up to provide a promising start for evolution, while a dynamic multitask framework is also built up with a flexible multitask merging mechanism and an effective hybrid reproduction mechanism. In the experiments, 7 existing algorithms are used to compare with the proposed AIMEA on 20 classification datasets. The Wilcoxon's Test and the Friedman's Test are also adopted for more comprehensive analyses on the experiment results. As a result, the proposed AIMEA shows significantly better performances on most datasets in terms of 3 widely-used performance indicators, along with generally less computational time and better solution distributions.