CMHS:Feature selection based on harmony search algorithm with cosine and momentum control factor

Selecting the optimal subset from high-dimensional features can improve the algorithm’s efficiency and performance for data mining tasks while also avoiding the overfitting phenomenon. Some works have improved the harmony search(HS) algorithm to handle the challenge of finding the appropriate feature subset. These algorithms, however, are still suffering from an imbalance between global optimization and convergence accuracy. In this paper, we propose and evaluate a harmony search algorithm based on cosine and momentum control factors (CMHS) for feature selection. We incorporate the cosine similarity between harmonic vectors into the dynamic smoothing adjustment strategy of the HS’s control parameters and design a novel fitness function to overcome the mentioned problem. We compared the accuracy of seven UC datasets side by side, and the average accuracy of CMHS is 96.49%, higher than the 93.78% of NGHS and the 93.33% of IBGHS. Moreover, when compared to NGHS and IBGHS, the fitness value of CMHS decreases by 25% on average, indicating that our proposed improvement strategy strikes a reasonable balance between global optimization and convergence accuracy, allowing CMHS to display more steady performance in the feature selection task.