Yab, Li Yu (2022) A modified whale optimisation algorithm for efficient filter-based feature selection in high-dimensional datasets. Masters thesis, Universiti Tun Hussein Onn Malaysia.

Preview	Text 24p YAB LI YU.pdf Download (540kB) | Preview
	Text (Copyright Declaration) YAB LI YU COPYRIGHT DECLARATION.pdf Restricted to Repository staff only Download (321kB) | Request a copy
	Text (Full Text) YAB LI YU WATERMARK.pdf Restricted to Registered users only Download (21MB) | Request a copy

Abstract

In a high-dimensional dataset (HDD), reducing the “curse of dimensionality” via feature selection is crucial. Recently, the integration of metaheuristic algorithms in solving feature selection problems had yielded outstanding results, according to literature findings. Hence, this study focused on integrating the Whale Optimisation Algorithm (WOA) for filter-based feature selection in HDDs. However, the WOA is known to have a slow convergence speed issue caused by the control parameter, a, which influences the balancing of the exploration and exploitation phases, eventually affecting the searching strategy. Therefore, this research proposed a modified WOA (mWOA) by inversing the control parameter values to allow the mWOA more search spaces during the initial searching phase, which eventually would increase the convergence speed. The proposed mWOA was implemented as the filter-based feature selection in four benchmark medical HDDs, namely Colon, CNS, SMK_CAN_187, and GLI_85. The performance of the proposed mWOA was compared against those of two filter-based feature selection algorithms, namely the original WOA and the Grey Wolf Optimiser (GWO). It was proven that the proposed mWOA outperformed the WOA and the GWO in 3 out of 4 cases (75%) in both best and average execution times when selecting the most relevant features of the HDDs. In addition, the mWOA also outperformed the WOA and the GWO in 8 out of 12 test cases (67%) in classification accuracy when using Decision Tree, Support Vector Machine, and Naïve Bayes classifiers

Actions (login required)

View Item