Granular Ball Fuzzy Neighborhood Rough Sets-Based Feature Selection via Multiobjective Mayfly Optimization

Most feature selection models via swarm intelligence optimization have difficulty achieving an optimal global subset of features and are not ideal for classifying high-dimensional data. We study a granular ball fuzzy neighborhood rough sets-based feature selection approach via multiobjective mayfly optimization on high-dimensional datasets. First, to enhance the ability to search for samples in granular balls, the granular ball radius is defined by the standard deviation coefficient. To measure sparse samples with noise in the granular ball, a new fuzzy neighborhood is constructed inside the granular ball, and upper and lower approximations are presented to develop the granular ball fuzzy neighborhood sets model. Second, to estimate the uncertainty of features in granular balls, fuzzy neighborhood entropy is provided. In the process of searching for features in fuzzy neighborhood decision systems, a feature-partitioning strategy based on the average fuzzy neighborhood entropy is studied. A subset of the preselected features is subsequently formed in the first stage. Third, to enhance the diversity in nondominated solutions, the feature vector is decoded into the mayfly, which is optimized through the mesh model. The mayfly ranking strategy updates the mayfly velocity and position to avoid local optima. Thus, in the second stage, the improved multiobjective mayfly optimization strategy can be utilized in selecting the optimal subset of features. Finally, a feature selection scheme is proposed for high-dimensional data with noise. Experimental findings prove that the developed methodology is viable and has excellent classification efficiency on 12 high-dimensional datasets.