Mutable composite firefly algorithm for gene selection in microarray based cancer classification

• 主要作者: Fajila, Mohamed Nisper Fathima
• 格式: Thesis
• 語言: 英语
• 出版: 2022
• 主題: RC0254 Neoplasms. Tumors. Oncology (including Cancer)
• 在線閱讀: https://etd.uum.edu.my/10171/1/s826643_01.pdf

Cancer classification is critical due to the strenuous effort required in cancer treatment and the rising cancer mortality rate. Recent trends with high throughput technologies have led to discoveries in terms of biomarkers that successfully contributed to cancerrelated issues. A computational approach for gene selection based on microarray data analysis has been applied in many cancer classification problems. However, the existing hybrid approaches with metaheuristic optimization algorithms in feature selection (specifically in gene selection) are not generalized enough to efficiently classify most cancer microarray data while maintaining a small set of genes. This leads to the classification accuracy and genes subset size problem. Hence, this study proposed to modify the Firefly Algorithm (FA) along with the Correlation-based Feature Selection (CFS) filter for the gene selection task. An improved FA was proposed to overcome FA slow convergence by generating mutable size solutions for the firefly population. In addition, a composite position update strategy was designed for the mutable size solutions. The proposed strategy was to balance FA exploration and exploitation in order to address the local optima problem. The proposed hybrid algorithm known as CFS-Mutable Composite Firefly Algorithm (CFS-MCFA) was evaluated on cancer microarray data for biomarker selection along with the deployment of Support Vector Machine (SVM) as the classifier. Evaluation was performed based on two metrics: classification accuracy and size of feature set. The results showed that the CFS-MCFA-SVM algorithm outperforms benchmark methods in terms of classification accuracy and genes subset size. In particular, 100 percent accuracy was achieved on all four datasets and with only a few biomarkers (between one and four). This result indicates that the proposed algorithm is one of the competitive alternatives in feature selection, which later contributes to the analysis of microarray data.