Asynchronous accelerating multi-leader salp chains for feature selection

Feature selection is an imperative preprocessing step that can positively affect the performance of machine learning techniques. Searching for the optimal feature subset amongst an unabridged dataset is a challenging problem, especially for large-scale datasets. In this research, a binary Salp Swarm Algorithm (SSA) with asynchronous updating rules and a new leadership structure is proposed. To set the best leadership structure, several extensive experiments are performed to determine the most effective number of leaders in the social organization of the artificial salp chain. Inspired from the behavior of a termite colony (TC) in dividing the termites into four types, the salp chain is then divided into several sub-chains, where the salps in each sub-chain can follow a different strategy to adaptively update their locations. Three different updating strategies are employed in this paper. The proposed algorithm is tested and validated on 20 well-known datasets from the UCI repository. The results and comparisons verify that utilizing half of the salps as leaders of the chain can significantly improve the performance of SSA in terms of accuracy metric. Furthermore, dynamically tuning the single parameter of algorithm enable it to more effectively explore the search space in dealing with different feature selection datasets.