Scholarlink

<p>Discriminative feature selection (DFS) methods derived from row-sparse linear discriminant analysis (LDA) are effective for handling high-dimensional data. However, similar to LDA, existing DFS methods often suffer from scatter matrix singularity and high computational complexity. While many methods address the former, practical solutions for computational efficiency remain scarce, which severely restricts their application to large-scale problems and hinders efficient model selection. To address this challenge, this paper first proposes a new optimization model for discriminative feature selection (DFS), termed EDFS-B. This model minimizes a row sparsity measure of the transformation matrix while constraining the between-class distances after transformation. Compared with existing DFS models, the core innovations of this model are as follows: (1) Within an iteratively reweighted (IRW) framework, the model’s solution can be obtained by applying QR decomposition to a small-scale matrix composed of class centers, thereby significantly reducing computational cost; (2) The model does not require a sparse regularization parameter, thus reducing the number of parameters that need to be tuned. To recover within-class information potentially lost in EDFS-B, this paper further develops a new DFS framework, termed EDFS, which is a two-stage algorithm that integrates EDFS-B with an existing ratio-trace criterion-based DFS (RTC-DFS) method. Since RTC-DFS operates on the dimensionally reduced space provided by EDFS-B, the overall framework also maintains exceptional efficiency. Extensive experiments on eleven benchmark datasets demonstrate that both EDFS-B and EDFS exhibit superior in terms of classification accuracy and running time compared to existing DFS methods. Notably, EDFS-B achieves more than a 10-fold speedup over other algorithms of the same category among those compared.</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

Efficient discriminative feature selection for high-dimensional data

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Xiaobin Zhi,
  • Xue Wang,
  • Yangyi Jing,
  • Shaoru Wu

摘要

Discriminative feature selection (DFS) methods derived from row-sparse linear discriminant analysis (LDA) are effective for handling high-dimensional data. However, similar to LDA, existing DFS methods often suffer from scatter matrix singularity and high computational complexity. While many methods address the former, practical solutions for computational efficiency remain scarce, which severely restricts their application to large-scale problems and hinders efficient model selection. To address this challenge, this paper first proposes a new optimization model for discriminative feature selection (DFS), termed EDFS-B. This model minimizes a row sparsity measure of the transformation matrix while constraining the between-class distances after transformation. Compared with existing DFS models, the core innovations of this model are as follows: (1) Within an iteratively reweighted (IRW) framework, the model’s solution can be obtained by applying QR decomposition to a small-scale matrix composed of class centers, thereby significantly reducing computational cost; (2) The model does not require a sparse regularization parameter, thus reducing the number of parameters that need to be tuned. To recover within-class information potentially lost in EDFS-B, this paper further develops a new DFS framework, termed EDFS, which is a two-stage algorithm that integrates EDFS-B with an existing ratio-trace criterion-based DFS (RTC-DFS) method. Since RTC-DFS operates on the dimensionally reduced space provided by EDFS-B, the overall framework also maintains exceptional efficiency. Extensive experiments on eleven benchmark datasets demonstrate that both EDFS-B and EDFS exhibit superior in terms of classification accuracy and running time compared to existing DFS methods. Notably, EDFS-B achieves more than a 10-fold speedup over other algorithms of the same category among those compared.