Scholarlink

<p>In daily life, real data are often affected by various noises, including feature noise and labeling noise. The valley loss function inherits the merits of pinball loss function and ramp loss function, which is not only robust to feature noise and label noise, but also has good sparsity. The multiple birth support vector machine (MBSVM) has a faster training speed in solving multi-classification problems, which has an obvious advantage over other multi-classification algorithms. However, due to the use of hinge loss function in MBSVM, it is sensitive to outliers and unstable to re-sampling. To tackle these problems, we propose a novel multiple birth support vector machine with a valley loss function (VMBSVM). We further theoretically analyze the noise insensitivity and sparsity of VMBSVM and discuss the computational complexity of VMBSVM. Since the valley loss function is difficult to optimize due to its non-convex property, we employ a concave-convex procedure (CCCP) to solve VMBSVM. A host of experiments are conducted to verify the effectiveness of our proposed VMBSVM.</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

Valley-loss multiple birth support vector machine for multi-class classification

  • 期刊论文

  • 发表日期：

    2025年4月28日

查看全文
  • Xue Li,
  • Jiaqi Zhang,
  • Hu Yang

摘要

In daily life, real data are often affected by various noises, including feature noise and labeling noise. The valley loss function inherits the merits of pinball loss function and ramp loss function, which is not only robust to feature noise and label noise, but also has good sparsity. The multiple birth support vector machine (MBSVM) has a faster training speed in solving multi-classification problems, which has an obvious advantage over other multi-classification algorithms. However, due to the use of hinge loss function in MBSVM, it is sensitive to outliers and unstable to re-sampling. To tackle these problems, we propose a novel multiple birth support vector machine with a valley loss function (VMBSVM). We further theoretically analyze the noise insensitivity and sparsity of VMBSVM and discuss the computational complexity of VMBSVM. Since the valley loss function is difficult to optimize due to its non-convex property, we employ a concave-convex procedure (CCCP) to solve VMBSVM. A host of experiments are conducted to verify the effectiveness of our proposed VMBSVM.