<p>Cancer gene expression data is characterized by high dimensionality, multiple samples and multi-class classification. Feature selection (FS) serves as a crucial preprocessing step in addressing these data mining challenges. To tackle the issue of FS in cancer gene expression data, a novel transfer function structure known as the dynamic time-varying transfer function was proposed. By employing the concept of dynamic adjustment, an adaptive adjustment strategy based on population diversity was introduced, which is then integrated with the time-varying transfer function to yield the dynamic time-varying transfer function. Initially, in order to demonstrate the superiority of the dynamic time-varying transfer function a comparative analysis was carried out by involving three distinct design methodologies: dynamic transfer function, time-varying transfer function and dynamic time-varying transfer function. S-shaped, V-shaped, U-shaped and M-shaped transfer functions were selected as experimental carriers for our research. In the first phase of experimentation, these three types of design methods are combined with each of the aforementioned four different shapes of transfer functions. Performance tests are conducted on nine standard UCI datasets to validate the efficacy of the proposed dynamic time-varying transfer functions. Subsequently, to assess the generalizability of our proposed approach across 12 cancer gene expression datasets for testing purposes five algorithms (AOA, COA, PSO, WOA and ZOA) are employed. The Friedman test and Wilcoxon rank-sum test are utilized for result analysis. The simulation outcomes confirm that applying the dynamic time-varying transfer function effectively simplifies feature subsets while enhancing classification accuracy and achieving lower fitness values when applied to these datasets.</p>