Scholarlink

<p>One of the most crucial steps for preparing datasets for machine learning is Feature Selection (FS). It aims to identify the smallest subset of features that best capture the essence of all features in a given dataset. However, it is time-consuming to solve the FS problem using exact methods due to the exponential complexity of the FS problem in high-dimensional datasets (HDDs). Therefore, approximation methods, such as meta-heuristic algorithms, can be used to solve the FS problem in a reasonable time. The Artificial Rabbit Optimization (ARO) algorithm is a new optimization search method that models the social activities of rabbits in the wild to solve complex continuous optimization problems. In its basic form, ARO cannot be applied directly to solve binary optimization problems, such as the FS problem in HDDs. Besides, ARO suffers from some performance shortcomings such as premature convergence and limited exploration ability. Therefore, this research study introduces the Binary Improved ARO Algorithm (BIARO) to provide more efficient solutions for the FS problem in HDDs. BIARO improves the search process of the ARO algorithm in four ways. First, it implements the Gaussian perturbation (GP) method at the beginning of the ARO algorithm to add stochastic dynamic behavior to the creation process of the initial binary combinations of features. Second, it utilizes the Adaptive β-hill climbing (ABHC) method in the evolution process of ARO to improve its global search and exploration abilities. Third, it uses the Mixed Opposition-based Learning (MOBL) approach in the evolution process of ARO to explore the opposite solutions of the best and worst rabbits in the population of rabbits. Finally, it employs binary conversion methods to compute binary subsets of features from the real numbers produced by the evolution processes of the ARO and ABHC algorithms. The performance of BIARO, six machine learning classifiers, and nine efficient, popular, nature-inspired FS algorithms was investigated using 21 real-life HDDs from different fields. According to experimental results, BIARO achieved exceptional results in real-life HDDs, with average accuracy values between 0.69 and 1, average precision values in the range of 0.67 to 1, average F1 scores in the range of 0.67 to 1, average Recall values in the range of 0.67 to 1, and a performance improvement rate between − 2 and 66. Furthermore, statistical verification methods were employed to evaluate the dependability of the experimental results. The statistical results show that BIARO scored the highest rank based on the fitness values for 14 HDDs, with a p-value of 0.001. Besides, it shows that it scored the highest rank based on the number of selected features for 18 HDDs, with a p-value of 0.001. All experimental and statistical results provide evidence that BIARO is an efficient FS method for HDDs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

BIARO: an improved artificial rabbit optimization algorithm for feature selection in high-dimensional data

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Esraa Al-Daom,
  • Bilal H. Abed-alguni

摘要

One of the most crucial steps for preparing datasets for machine learning is Feature Selection (FS). It aims to identify the smallest subset of features that best capture the essence of all features in a given dataset. However, it is time-consuming to solve the FS problem using exact methods due to the exponential complexity of the FS problem in high-dimensional datasets (HDDs). Therefore, approximation methods, such as meta-heuristic algorithms, can be used to solve the FS problem in a reasonable time. The Artificial Rabbit Optimization (ARO) algorithm is a new optimization search method that models the social activities of rabbits in the wild to solve complex continuous optimization problems. In its basic form, ARO cannot be applied directly to solve binary optimization problems, such as the FS problem in HDDs. Besides, ARO suffers from some performance shortcomings such as premature convergence and limited exploration ability. Therefore, this research study introduces the Binary Improved ARO Algorithm (BIARO) to provide more efficient solutions for the FS problem in HDDs. BIARO improves the search process of the ARO algorithm in four ways. First, it implements the Gaussian perturbation (GP) method at the beginning of the ARO algorithm to add stochastic dynamic behavior to the creation process of the initial binary combinations of features. Second, it utilizes the Adaptive β-hill climbing (ABHC) method in the evolution process of ARO to improve its global search and exploration abilities. Third, it uses the Mixed Opposition-based Learning (MOBL) approach in the evolution process of ARO to explore the opposite solutions of the best and worst rabbits in the population of rabbits. Finally, it employs binary conversion methods to compute binary subsets of features from the real numbers produced by the evolution processes of the ARO and ABHC algorithms. The performance of BIARO, six machine learning classifiers, and nine efficient, popular, nature-inspired FS algorithms was investigated using 21 real-life HDDs from different fields. According to experimental results, BIARO achieved exceptional results in real-life HDDs, with average accuracy values between 0.69 and 1, average precision values in the range of 0.67 to 1, average F1 scores in the range of 0.67 to 1, average Recall values in the range of 0.67 to 1, and a performance improvement rate between − 2 and 66. Furthermore, statistical verification methods were employed to evaluate the dependability of the experimental results. The statistical results show that BIARO scored the highest rank based on the fitness values for 14 HDDs, with a p-value of 0.001. Besides, it shows that it scored the highest rank based on the number of selected features for 18 HDDs, with a p-value of 0.001. All experimental and statistical results provide evidence that BIARO is an efficient FS method for HDDs.