Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (9): 1532-1544.DOI: 10.3778/j.issn.1673-9418.1906018

Previous Articles     Next Articles

Style Regularized Least Squares Support Vector Machine Based on Multiple Kernel Learning

SHEN Hao, WANG Shitong   

  1. 1. School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China
    2. Key Laboratory of Media Design and Software Technology of Jiangsu Province, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2020-09-01 Published:2020-09-07

基于多核学习的风格正则化最小二乘支持向量机

沈浩王士同   

  1. 1. 江南大学 数字媒体学院,江苏 无锡 214122
    2. 江南大学 江苏省媒体设计与软件技术重点实验室,江苏 无锡 214122

Abstract:

Though current multiple kernel learning algorithms integrate the abilities of different kernel functions on the representation of the physical features of data, they do not make full use of the style information existing in the stylistic dataset. Therefore, a style regularized least squares support vector machine based on multiple kernel learning (MK-SRLSSVM) is proposed. Its basic idea is to consider the style transformation matrices which represent the style information of samples and it is regularized in the objective function. The commonly-used alternating optimization technique is utilized to optimize the objective function. Style transformation matrices and classifier parameters are simultaneously updated during iteration. In order to use the trained style information in the prediction process, two new rules are considered in the traditional prediction model. The unknown patterns are normalized by the learned style transform matrix before being classified. The proposed classifier not only keeps the advantages of existing multi-kernel learning algorithms in representing the physical features of samples, but also exploits the style information contained in the stylistic dataset so as to improve the classification performance effectively. The experimental results on the stylistic datasets confirm the effectiveness of the proposed classifier.

Key words: least squares support vector machines (LSSVM), multiple kernel learning, stylistic data, style information

摘要:

当前的多核学习方法结合了不同核函数在对数据的物理特性表示上的能力,但在风格化数据集中不能充分利用样本中所隐含的风格信息。由此,提出应用于风格化数据的基于多核学习的风格正则化最小二乘支持向量机(MK-SRLSSVM)。算法利用风格转换矩阵表示包含在样本中的风格信息,并在目标函数中对其进行正则化处理,通过常用的交替优化方法对目标函数进行优化,在迭代过程中同步更新风格转换矩阵和分类器参数。为在预测过程中利用已学习的风格信息,在传统预测方法中增加了两种新的规则,在分类之前预先使用风格转换矩阵对样本风格进行标准化处理。所提出的分类器不仅利用了现有多核学习算法在表示样本的物理特征方面的优势,同时有效挖掘了数据集内包含的风格信息以提高分类性能,在风格化数据集中的实验结果证明了算法的有效性。

关键词: 最小二乘支持向量机(LSSVM), 多核学习, 风格化数据, 风格信息