A feature selection approach combining neural networks with genetic algorithms

Article type: Research Article

Authors: Huang, Zhi^{; *}

Affiliations: School of Information Engineering, Mianyang Teachers’ College, Sichuan Province, China. E-mail: [email protected]

Correspondence: [*] Corresponding author. E-mail: [email protected].

Abstract: Value Feature selection is an effective method to solve the curse of dimensionality, which widely employs Evolutionary Computation (EC), such as Genetic Algorithms (GA), by regarding feature subsets as individuals. However, it is impossible for EC based feature selection approaches to possess big population sizes because of very long and infeasible computational time. We have proposed a method screening individuals by estimating their classification performances rapidly instead of deriving theirs with a certain classifier dilatorily. Consequently, aiming at improving classification accuracies, we propose an approach named as FS-NN-GA (Feature Selection approach based on Neural Networks and Genetic Algorithms) in this work. The proposed approach employs the neural networks trained with some randomly generated individuals, and their actual classification accuracies to estimate individuals’ classification accuracies and screens them in each round of GA. The individuals with low estimated accuracies are directly eliminated. Only a small number of individuals with high estimated accuracies are reserved, evaluated by deriving their accuracies with a certain classifier, and participate GA operations to be explored emphatically. As a result, big population sizes become feasible, and a huge number of individuals can be considered by GA in acceptable and feasible time, which improves performances of GA and derives high accuracies. We perform the experiments with 10 data sets in comparison with 11 available approaches. The experimental results show that FS-NN-GA outperforms other approaches on most data sets.

Keywords: Feature selection, classification accuracy, Genetic Algorithms, neural networks, population size, computational time

DOI: 10.3233/AIC-190626

Journal: AI Communications, vol. 32, no. 5-6, pp. 361-372, 2019