Scholarlink

<p>The dynamic granularity search method is effective and efficient in feature selection (FS), grouping multiple features into a single bit to reduce the search space. However, existing algorithms suffer from unreasonable granularity initialization and incorrect splitting of feature groups due to neglecting feature complementarity and the performance differences among feature groups. To address these issues, an improved dynamic granularity particle swarm optimization algorithm is proposed for high-dimensional FS (IDGPSO-FS). Specifically, a new granularity initialization strategy is proposed to optimize initial granularity settings by considering feature complementarity. Secondly, a differentiated granularity update strategy is proposed to adjust the granularity structure based on feature group performance and merge groups with strong synergy. In addition, a new particle update strategy is proposed to reduce reliance on the global optimum, enhancing local search and avoiding early convergence. The experimental results show that the feature subsets obtained by IDGPSO-FS could achieve better classification accuracy with lower computational overhead compared to state-of-the-art FS algorithms on 10 high-dimensional datasets.</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

An improved feature selection method based on dynamic granularity particle swarm optimization

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Fei Han,
  • Xi Cheng,
  • Qinghua Ling,
  • Henry Han,
  • Huan Liu,
  • Hui Sun,
  • Sizhe Xu,
  • Yuhui Han

摘要

The dynamic granularity search method is effective and efficient in feature selection (FS), grouping multiple features into a single bit to reduce the search space. However, existing algorithms suffer from unreasonable granularity initialization and incorrect splitting of feature groups due to neglecting feature complementarity and the performance differences among feature groups. To address these issues, an improved dynamic granularity particle swarm optimization algorithm is proposed for high-dimensional FS (IDGPSO-FS). Specifically, a new granularity initialization strategy is proposed to optimize initial granularity settings by considering feature complementarity. Secondly, a differentiated granularity update strategy is proposed to adjust the granularity structure based on feature group performance and merge groups with strong synergy. In addition, a new particle update strategy is proposed to reduce reliance on the global optimum, enhancing local search and avoiding early convergence. The experimental results show that the feature subsets obtained by IDGPSO-FS could achieve better classification accuracy with lower computational overhead compared to state-of-the-art FS algorithms on 10 high-dimensional datasets.