Dynamic niche technology based hybrid breeding optimization algorithm for multimodal feature selection

As a primary approach to address feature selection problems, evolutionary algorithms have been widely proposed to deal with the problem. Most of these methods are designed to find a single feature subset. However, the optimal feature subset within a dataset is often not unique, indicating that feature selection exhibits multimodal characteristics. Representing data information with a single feature subset will be biased. Nevertheless, most existing evolutionary algorithms suffered from a lack of diversity, making them insufficiently effective in finding multiple optimal solutions. To address this issue, this paper investigates a new evolutionary algorithm derived from the Heterosis theory, the hybrid breeding optimization algorithm (HBO). Additionally, HBO is incorporated with dynamic niching technology and a double-stage multimodal hybrid breeding optimization (DSMHBO) is proposed. Further, to enhance the performance of the traditional HBO, neighborhood search and elite mutation strategies are introduced in the global search, and a neighborhood crossover strategy is applied to broaden the diversity of population. When the number of niches is set to 1, DSMHBO is equivalent to the double-stage hybrid breeding optimization (DSHBO). Finally, eight algorithms such as DSHBO, cuckoo search (CS), fruit fly algorithm (FA) are compared over 13 datasets. DSHBO achieves the best average classification accuracy (ACA) on 7 datasets and the best highest classification accuracy (HCA) on 10 datasets, significantly surpassing the comparison algorithms. In addition, the proposed DSMHBO is compared with newly proposed algorithms, such as whale optimization algorithm (WOA) and Harris hawk optimization algorithm(HHO) over 10 datasets. DSMHBO achieved average ACA and HCA values of 93.54% and 95.52%, much higher than the comparison models. It also can identify up to 187 feature subsets on the Lung Cancer dataset, which indicates its ability to locate multiple peaks. Moreover, even as the error level increases, the global search capability of DSMHBO remains superior to other algorithms, proving that DSMHBO is an effective method for multimodal feature selection.