Segmented initialization and offspring modification in evolutionary algorithms for bi-objective feature selection
Hang Xu, Bing Xue, Mengjie Zhang
Abstract
In classification, feature selection mainly aims at reducing the dataset dimensionality and increasing the classification accuracy, which also results in higher computational efficiency than using the original full set of features. Population-based meta-heuristic, evolutionary algorithms have been widely used to solve the bi-objective feature selection problem, which minimizes the number of selected features and the error of classification model. However, most of them are not specifically designed for feature selection, and disregard many of its complex characteristics. In this paper, we propose a generic approach that focuses on improving the initialization effectiveness and offspring quality, in order to boost the performance of existing evolutionary algorithms for bi-objective feature selection. To be more specific, a segmented initialization mechanism is used to enhance the exploration width, while an offspring modification mechanism is proposed to ensure the exploitation depth. Combining them together will make a good trade-off between the diversity and convergence. In the experiments, we plug the proposed approach into three different types of multi-objective evolutionary algorithms, and test them on 18 classification datasets with two widely-used performance metrics. The empirical results prove the significant contribution of the proposed approach on the optimization and classification performance.