A Multimodal Multiobjective Genetic Algorithm for Feature Selection

When performing feature selection on most data sets, there is a general situation that some different feature subsets have the same number of selected features and classification error rate. This indicates that feature selection in some data sets is a multimodal multiobjective optimization (MMO) problem. Most of the current studies on feature selection ignore the MMO problems. Therefore, this paper proposes a feature selection method based on a multimodal multiobjective genetic algorithm (MMOGA) to solve the problem. This algorithm is mainly improved in three aspects. First, a special initialization strategy based on symmetric uncertainty is designed to improve the fitness of the initial population. Second, this paper adds a niche strategy to the genetic algorithm to search for multimodal solutions. Unlike traditional niche methods that has a central individual, this algorithm also considers the distances between individuals in the niche. Third, to effectively utilize excellent individuals for evolution, this algorithm uses a method based on the Pareto set of the niche to generate offspring. Finally, by comparing with other algorithms, the effectiveness of the MMOGA in feature selection is verified. This algorithm can successfully find equivalent feature subsets on different datasets.