Performance Analysis for SVM Combining with Metric Learning

Performance Analysis for SVM Combining with Metric Learning
复制标题

结合度量学习的SVM性能分析

DOI:
10.1007/s11063-017-9771-7
复制
发表时间:
2018-01
影响因子:
3.1
通讯作者:
Yaxin Peng
Yaxin Peng
中科院分区:
计算机科学4区
文献类型:
--
作者:
Lingfang Hu;Juan Hu;Zhen Ye;Chaomin Shen;Yaxin Peng

文献摘要

参考文献

相似文献

本文分析了支持向量机与度量学习相结合的性能,以评价度量学习对支持向量机的改进效果。首先,我们建立了支持向量机不能通过度量学习提高性能的充分条件。其次,为了验证充分条件是否成立,我们通过分别学习一个标准正交矩阵和一个对角矩阵,建立了一个两步度量学习策略。第三,我们分析了两步度量学习后充分条件成立的情况,从而证明了支持向量机提高准确率的实用性。最后,我们提供了一些实验,并将度量学习应用于支持向量机的三维目标分类和人脸识别。实验结果表明,利用度量学习提高SVM分类性能是有效的。
This paper analyses the performance of combining Support Vector Machines (SVMs) and metric learning, in order to evaluate the effect of metric learning on improving SVM. First, we establish the sufficient condition under which the performance of SVM cannot be improved by metric learning. Second, to verify whether the sufficient condition holds, we develop a two-step metric learning strategy by learning an orthonormal matrix and a diagonal matrix respectively. Third, we analyze the case when the sufficient condition holds after the two-step metric learning, and therefore demonstrate the practicability of improving the accuracy of SVM. Finally, we provide some experiments, and also apply metric learning into SVM for 3D object classification and face recognition. The experimental results demonstrate the effectiveness of improving the SVM classification performance by metric learning.
DOI: --
发表时间: 1936
期刊: --
影响因子: --
作者:
P. Mahalanobis
通讯作者: P. Mahalanobis
动物机器人小型光纤耦合激光控制器研究
DOI: 10.1108/aa-11-2015-097
发表时间: 2016-04
影响因子: 2.1
作者:
王海霞;王玉良;孙耀宗;蒲群;卢晓
通讯作者: 卢晓
仿生方法对织物组织图案和纱线颜色的不变识别和分类
DOI: 10.1108/aa-11-2015-100
发表时间: 2016-04
影响因子: 2.1
作者:
Babar Khan;Fang Han;Zhijie Wang;Rana J. Masood
通讯作者: Rana J. Masood
DOI: 10.1109/cvpr.2013.452
发表时间: 2013-06
期刊: 2013 IEEE Conference on Computer Vision and Pattern Recognition
影响因子: --
作者:
Lior Wolf;Noga Levy
通讯作者: Lior Wolf;Noga Levy
DOI: 10.1109/tip.2012.2202676
发表时间: 2013-01-01
影响因子: 10.6
作者:
Gao, Yue;Wang, Meng;Wu, Xindong
通讯作者: Wu, Xindong