Scholarlink

<p>Nonparallel support vector machine (NPSVM) and its extended models have been widely applied in recent years. Among them, the improved NPSVM (INPSVM) not only inherits the advantages of NPSVM but also enhances its noise insensitivity. However, existing solvers fail to effectively address the computational speed issue of INPSVM when dealing with large-scale problems. Inspired by the sparsity of INPSVM solutions, we propose a safe screening rule to accelerate INPSVM (SSR-INPSVM) and reduce its computational burden. SSR-INPSVM can identify and discard most non-support vectors (SVs) before solving, thus reducing the scale of the problems. At the same time, the classification accuracy obtained is identical to that of the original INPSVM, ensuring its safety. Furthermore, embedding SSR-INPSVM into the parameter tuning process can significantly improve overall computational efficiency. In addition, our method can be combined with other fast solvers since it is used prior to the solving phase. Numerical experiments on two synthetic datasets, twelve benchmark datasets, and a real-world dataset demonstrate the efficiency and feasibility of SSR-INPSVM.</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 safe screening rule for accelerating improved nonparallel support vector machine

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Kun Jiang,
  • Hongmei Wang,
  • Ping Li

摘要

Nonparallel support vector machine (NPSVM) and its extended models have been widely applied in recent years. Among them, the improved NPSVM (INPSVM) not only inherits the advantages of NPSVM but also enhances its noise insensitivity. However, existing solvers fail to effectively address the computational speed issue of INPSVM when dealing with large-scale problems. Inspired by the sparsity of INPSVM solutions, we propose a safe screening rule to accelerate INPSVM (SSR-INPSVM) and reduce its computational burden. SSR-INPSVM can identify and discard most non-support vectors (SVs) before solving, thus reducing the scale of the problems. At the same time, the classification accuracy obtained is identical to that of the original INPSVM, ensuring its safety. Furthermore, embedding SSR-INPSVM into the parameter tuning process can significantly improve overall computational efficiency. In addition, our method can be combined with other fast solvers since it is used prior to the solving phase. Numerical experiments on two synthetic datasets, twelve benchmark datasets, and a real-world dataset demonstrate the efficiency and feasibility of SSR-INPSVM.