Improving Meta-Heuristic Algorithms for Feature Selection in Multiclass Classification

Classification is a significant part and plays a big role in solving data problems, as many multiclass classification problems have appeared in various applications. Focusing on selecting the type of important features from the overall feature set has become an important area of application and research. Several binary meta-heuristic algorithms have been used as a measure and as an aid for feature selection. These algorithms require that an initial set be specified, and initialization fit values play an important role in obtaining the desired final result. In the initial stage of population initialization, situations are initialized randomly by means of a uniform distribution function which has a variance in the generation of values and thus has an effect on the results and values of classification some parametric and non-parametric methods have been proposed, such as Kruskal-Wallis test and one-way ANOVA, to avoid the problem of randomness in generating a population and to take into account the relationship between the features and the category variable, as these methods were chosen as an initial set for binary meta-algorithms. The performance significance of the proposed methods was evaluated by applying them to ten high-dimensional datasets. Most experimental findings and statistical measurements support the suggested approaches’ superior performance when compared to the industry standard methods in terms of classification performance accuracy, the number of features chosen, and runtime.