A Duplication Analysis-Based Evolutionary Algorithm for Biobjective Feature Selection

Feature selection is a complex optimization problem with important real-world applications. Normally, its main target is to reduce the dimensionality of the dataset and increase the effectiveness of the classification. Owing to the population-inspired characteristics, different evolutionary algorithms (EAs) have been proposed to solve feature selection problems over the past decades. However, the majority of them only consider single-objective optimization while many real-world problems have multiple objectives, which creates a genuine demand for designing more suitable and effective EAs to handle multiobjective feature selection. A multiobjective feature selection problem usually consists of two objectives: one is to minimize the number of selected features and the other is to minimize the error of classification. In this article, we propose a duplication analysis-based EA (DAEA) for biobjective feature selection in classification. In the proposed algorithm, we make improvements on the basic dominance-based EA framework in three aspects: first, the reproduction process is modified to improve the quality of offspring; second, a duplication analysis method is proposed to filter out the redundant solutions; and third, a diversity-based selection method is adopted to further select the reserved solutions. In the experiments, we have compared the proposed algorithm with five state-of-the-art multiobjective EAs (MOEAs) and tested them on 20 classification datasets, using two widely used performance metrics. According to the empirical results, DAEA performs the best on most datasets, indicating that DAEA not only gains outstanding optimization performance but also obtains good classification and generalization results.

Automated summary

This paper proposes a duplication analysis-based EA (DAEA) for biobjective feature selection in classification and compares its performance with five state-of-the-art multiobjective EAs (MOEAs) on 20 classification datasets. The empirical results show that DAEA performs the best on most datasets, achieving outstanding optimization performance as well as good classification and generalization results.