Scholarlink

<p>Hybrid feature selection (FS) performs well by effectively balancing classification accuracy and computational efficiency. However, existing hybrid FS methods still suffer from the high computational cost during the clustering stage and often operate at a single granularity, making it difficult to simultaneously capture correlations at both the feature-group and individual feature levels. To address these challenges, a three-stage FS method based on Pareto-guided collaborative filtering and two-layer particle swarm optimization (PGCF-TLPSO-FS) is proposed. Firstly, inspired by multi-objective optimization, multiple metrics are used to collaboratively filter out irrelevant features, with relevant ones iteratively selected from the Pareto front. Secondly, relevant features are clustered into multiple feature groups through a new feature clustering strategy based on feature importance differences. Finally, a two-layer particle swarm optimization (TLPSO) algorithm is proposed to hierarchically model the feature selection search space, enabling the simultaneous capture of inter-group and intra-group feature correlations. Moreover, experimental results on 10 high-dimensional datasets show that, compared with the best-performing baseline, PGCF-TLPSO-FS improves classification accuracy by 3.77% and reduces running time by 33.35% on average.</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 three-stage feature selection method based on pareto-guided collaborative filtering and two-layer particle swarm optimization

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Fei Han,
  • Xi Cheng,
  • Qinghua Ling,
  • Henry Han,
  • Huan Liu,
  • Qing Liu

摘要

Hybrid feature selection (FS) performs well by effectively balancing classification accuracy and computational efficiency. However, existing hybrid FS methods still suffer from the high computational cost during the clustering stage and often operate at a single granularity, making it difficult to simultaneously capture correlations at both the feature-group and individual feature levels. To address these challenges, a three-stage FS method based on Pareto-guided collaborative filtering and two-layer particle swarm optimization (PGCF-TLPSO-FS) is proposed. Firstly, inspired by multi-objective optimization, multiple metrics are used to collaboratively filter out irrelevant features, with relevant ones iteratively selected from the Pareto front. Secondly, relevant features are clustered into multiple feature groups through a new feature clustering strategy based on feature importance differences. Finally, a two-layer particle swarm optimization (TLPSO) algorithm is proposed to hierarchically model the feature selection search space, enabling the simultaneous capture of inter-group and intra-group feature correlations. Moreover, experimental results on 10 high-dimensional datasets show that, compared with the best-performing baseline, PGCF-TLPSO-FS improves classification accuracy by 3.77% and reduces running time by 33.35% on average.