Parallel Dynamic Adaptive Transfer Function Based on Population Diversity to Solve High‐Dimensional Cancer Gene Expression Data Feature Selection Problem

Abstract

In cancer genomics research, feature selection (FS) of high‐dimensional gene expression data is of great significance to improve classification accuracy and reduce feature number. Aiming at the limitations of traditional transfer functions, such as nonadaptability and being prone to fall into local optimum when dealing with high‐dimensional data, a parallel dynamic adaptive transfer function based on population diversity is proposed. Firstly, a universal mirror symmetric inversion strategy is proposed based on five different types of transfer function (TF) families. Compared with the basic TFs, the proposed strategy considers more possibilities for particles in both positive and negative directions, improving the algorithm's ability to escape local optimum. Then, time‐varying factors were taken into account, enabling adaptive adjustment of various TFs. Finally, considering the influence of algorithm population diversity, the population evolution degree (PED) was defined. A dynamic adaptive TF based on population diversity was proposed. Through PED, the TF was dynamically and adaptively adjusted to achieve better binary conversion, enhancing the exploitation and exploration capabilities of the algorithm. In the wrapper FS method, SHO was adopted as the optimizer. The proposed method enables the binary mapping of the algorithm to dynamically adaptively change during the iterative process. Eventually, it adjusts the TF mapping capability for the next iteration based on its own fitness value. In the experimental part, the performance of the proposed strategy was first tested and verified through nine UCI datasets. The results showed that ITV‐VrV1 could achieve better fitness, improve accuracy and effectively reduce the number of features. Then, ITV‐VrV1 was extended to the cancer gene expression datasets. By comparing it with other binary algorithms, it was found that ITV‐VrV1 achieved the lowest average fitness and the highest average classification accuracy on all datasets. According to the Friedman test and Wilcoxon test, it was proved that ITV‐VrV1 ranked first in terms of average fitness and average classification accuracy, ranked second in terms of average number of selected features and showed significant differences from other comparison methods. It can effectively solve the problem of FS for high‐dimensional cancer data.