A Discrete Sine Cosine Algorithm with Enhanced Comprehensive Learning for Feature Selection

Feature selection (FS) plays a crucial role in enhancing the accuracy and response time of machine learning techniques. Due to the NP-hardness of the FS problem, an Enhanced Comprehensive Learning Discrete Sine Cosine Algorithm (ECLDSCA) incorporates several key components to effectively remove redundant features and identify the optimal feature subset. In the ECLDSCA, a tournament selection is first used to replace all individual update modes and maintain population diversity. Three enhanced comprehensive learning operators are then designed, including the global best learning operator, which learns the best optimization information in the search area; the population information learning operator, which learns in the local area; and the feature correlation learning operator, which refines the search space using feature information. In addition, the algorithm incorporates an enhanced greedy local search and random removal perturbation based on feature correlation. These mechanisms enhance the algorithm’s exploration capability, allowing for a more thorough search of the solution space. The tests were carried out on 10 benchmark datasets and compared with recent similar algorithms, the ECLDSCA demonstrated superior performance in terms of accuracy and the number of selected features.