LAUSR

Feature Selection (F.S.) reduces the number of features by removing unnecessary, redundant, and noisy information while keeping a relatively decent classification accuracy. F.S. can be considered an optimization problem. As the problem is challenging and there are many local solutions, stochastic optimization algorithms may be beneficial. This paper proposes a novel approach to dimension reduction in feature selection. As a seminal attempt, this work uses binary variants of the recent Marine Predators Algorithm (MPA) to select the optimal feature subset to improve classification accuracy. MPA is a new and novel nature-inspired metaheuristic. This research proposes an algorithm that is a hybridization between MPA and k-Nearest Neighbors (k-NN) called MPA-KNN. K-Nearest Neighbors (k-NN) is used to evaluate the selected features on medical datasets with feature sizes ranging from tiny to massive. The proposed methods are evaluated on 18 well-known UCI medical dataset benchmarks and compared with eight well-regarded metaheuristic wrapper-based approaches. The core exploratory and exploitative processes are adapted in MPA to select the optimal and meaningful features for achieving the most accurate classification. The results show that the proposed MPA-KNN approach had a remarkable capability to select the optimal and significant features. It performed better than the well-established metaheuristic algorithms we tested. The algorithms we used for comparison are Grey Wolf Optimizer (GWO), MothFlame Optimization Algorithm (MFO), Sine Cosine Algorithm (SCA), Whale Optimization Algorithm (WOA), Slap Swarm Algorithm (SSA), Butterfly Optimization Algorithm (BFO), and Harris Hawks Optimization (HHO). This paper is the first work that implements MPA for Feature Selection problems. The results ensure that the proposed MPA-KNN approach has a remarkable capability to select the optimal and significant features and performed better than several metaheuristic algorithms. MPA-KNN achieves the best averages accuracy, Sensitivity, and Specificity rates of all datasets.