Scholarlink

<p>Feature selection in high-dimensional datasets is a critical challenge in machine learning, as it involves reducing the number of features while enhancing classification accuracy. This study addresses this challenge by comparing and analyzing three state-of-the-art binary multi-objective optimization algorithms, each from a different category, specifically designed to tackle large-scale multi-objective feature selection: Diverse NSGA-II, Compact NSGA-II (CNSGA-II), and Multi-objective Coordinate Search (MOCS). Diverse NSGA-II employs evolutionary strategies to enhance search space exploration by replacing the worst individuals in each generation with randomly generated ones, each with a limited number of features. CNSGA-II uses a memory-efficient, non-population-based approach, which represents the population as a probability distribution to improve memory efficiency and accelerate convergence. Finally, MOCS applies a non-evolutionary population-based coordinate-search strategy to efficiently generate new solutions. We extended these algorithms with genuine uniform initialization and evaluated their performance against the widely used NSGA-II across nine large-scale datasets from bioinformatics and image recognition domains, with dimensions ranging from 2400 to 11,340. Statistical analyses, including the t-test and Friedman test, were used to validate the results, which demonstrate that genuine initialization significantly enhances classification accuracy and feature reduction in most cases, with genuinely initialized Diverse NSGA-II achieving the highest hypervolume (HV) in most scenarios. MOCS also performs competitively, particularly in memory-constrained environments. These methods are applicable to a range of practical problems, including gene selection and biomarker discovery in bioinformatics, as well as combinatorial optimization problem and feature reduction in image recognition tasks. This research highlights the versatility and effectiveness of these algorithms, providing valuable insights for tackling binary multi-objective optimization challenges.</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

A comparative study on population-based multi-objective feature selection

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Sevil Zanjani Miyandoab,
  • Shahryar Rahnamayan,
  • Azam Asilian Bidgoli,
  • Masoud Makrehchi

摘要

Feature selection in high-dimensional datasets is a critical challenge in machine learning, as it involves reducing the number of features while enhancing classification accuracy. This study addresses this challenge by comparing and analyzing three state-of-the-art binary multi-objective optimization algorithms, each from a different category, specifically designed to tackle large-scale multi-objective feature selection: Diverse NSGA-II, Compact NSGA-II (CNSGA-II), and Multi-objective Coordinate Search (MOCS). Diverse NSGA-II employs evolutionary strategies to enhance search space exploration by replacing the worst individuals in each generation with randomly generated ones, each with a limited number of features. CNSGA-II uses a memory-efficient, non-population-based approach, which represents the population as a probability distribution to improve memory efficiency and accelerate convergence. Finally, MOCS applies a non-evolutionary population-based coordinate-search strategy to efficiently generate new solutions. We extended these algorithms with genuine uniform initialization and evaluated their performance against the widely used NSGA-II across nine large-scale datasets from bioinformatics and image recognition domains, with dimensions ranging from 2400 to 11,340. Statistical analyses, including the t-test and Friedman test, were used to validate the results, which demonstrate that genuine initialization significantly enhances classification accuracy and feature reduction in most cases, with genuinely initialized Diverse NSGA-II achieving the highest hypervolume (HV) in most scenarios. MOCS also performs competitively, particularly in memory-constrained environments. These methods are applicable to a range of practical problems, including gene selection and biomarker discovery in bioinformatics, as well as combinatorial optimization problem and feature reduction in image recognition tasks. This research highlights the versatility and effectiveness of these algorithms, providing valuable insights for tackling binary multi-objective optimization challenges.