Feature Selection for Classification Problems by Binary Bare Bones Fireworks Algorithm

Classification represents a rather simple concept but the number of real-world applications that have it as the crucial part is enormous. Several factors determine the quality of classification such as the algorithm used for classification, tuning hyperparameters of the used algorithm and features used to describe instances that should be classified. Choosing an optimal feature set is an important part of the classification. On one hand, more features can capture more differences and similarities between instances which will lead to better classification. On the other hand, a too large set of features can lead to a poor classification model due to many unnecessary features that do not have an effect on the output. Feature selection is a combinatorial optimization problem and metaheuristics like swarm intelligence algorithms (SI) have been used for tackling such problems. In this paper, we proposed an adjusted binary bare bones fireworks algorithm for feature selection. The fitness function was modeled as a weighted sum of two objectives: reducing the number of features and minimizing classification error. The proposed algorithm is tested on the benchmark datasets and compared to other SI approaches from the literature. The proposed method successfully kept or improved classification accuracy while reducing the number of used features.