Feature selection of high-dimensional gene expression data is faced with redundancy and noise, which makes it difficult to identify truly biologically relevant features, thus affecting the predictive power and reliability of the model. Therefore, we propose a new hybrid feature selection algorithm, named as "Dynamic Feature-Maximum Spearman Maximum Variance Improved Binary Bat Algorithm (DF-MSMVIBBA)". Firstly, Maximum Spearman Maximum Variance filter algorithm is used to select the top \(K_{select}\) features with high scores to form a candidate feature subset. Features are selected based on the dynamic adjustment of the rows and columns of the dataset. Then the subset is initialized using spiral initialization method. Finally, the initialized candidate features are used in an improved wrapper algorithm, called Binary Bat Algorithm with Enhanced Local Search. By combining filter algorithm with wrapper algorithm through this linear transfer, the resulting DF-MSMVIBBA algorithm ensures computational efficiency while improving the performance of the model. Experimental results show that the algorithm outperforms other hybrid algorithms on most high-dimensional datasets, and BBA-ELS also outperforms the original Binary Bat Algorithm. Moreover, on all datasets, the average classification accuracy of the DF-MSMVIBBA exceeds \(90\%\) with a dimension reduction effect of at least \(99.8\%\) , demonstrating its effectiveness in improving performance.
