BSSFS: binary sparrow search algorithm for feature selection

Swarm intelligence algorithms can efficiently solve feature selection optimization problems for classification, and their classification performance is also excellent. The Sparrow Search Algorithm (SSA) has recently become a novel optimization method, which has the advantages of fewer parameters, a simple structure, and ease of implementation. Unfortunately, the traditional SSA has certain difficulties, trapping into a local optimum easily, and has a weakness in convergence. In this work, a novel SSA is studied to further develop two types of binary SSA for feature selection (BSSFS) to address optimization problems. First, to make the initial population of sparrows distributed as evenly as possible in the search space and augment the diversity of sparrows, a cubic chaotic mapping scheme needs to be introduced to initialize the sparrow population, and a reverse learning scheme is employed to enhance the global search ability of SSA. Second, the nonlinear adaptive inertia weight and the improved control parameters of step size are used to modify the position update formula of the sparrows and avert trapping into a local optimum. By integrating these strategies, a novel SSA (NSSA in short) is designed in this work. Third, NSSA is combined with the S- and V-shaped transfer functions, and the fitness function, which is raised on account of the set size of selected features and the classification error rate, are employed to design two types of BSSFS with the S- and V-shaped transfer functions, shortened as FSBSS and FSBSV, respectively. Finally, the optimization results for 18 classical benchmark functions illustrate that the optimization effectiveness of the NSSA is superior to that of other methods, and the experimental results on 10 low-dimensional and 10 high-dimensional datasets demonstrate that the FSBSV can outdo other comparative algorithms in terms of the classification effectiveness and robustness.