非均衡随机梯度下降SVM在线算法
发布时间：2018-09-14 17:30
【摘要】：随机梯度下降方法(SGD)已被应用于大规模支持向量机训练,随机梯度下降方法在训练时采取随机选点的方式,对于非均衡分类问题,必然导致多数类点被抽取到的概率要远远大于少数类点,这就造成了种种计算上的不平衡。为了处理大规模非均衡数据分类问题,提出了加权随机梯度下降的SVM在线算法,对于多数类中的样例被赋予较小的权值,而少数类中的样例被赋予较大的权值,然后利用加权随机梯度下降算法对SVM原问题进行求解,减少了超平面向少数类的偏移,较好地解决了大规模学习中非均衡数据的分类问题。同时,通过采用二分对数损失函数、不同的加权方式以及采用小批量梯度下降算法,从而在精度以及稳定性上使算法对非均衡数据的分类效果得到提升。
[Abstract]:Random gradient descent (SGD) method has been applied to large-scale support vector machine training. The random gradient descent (SGD) method uses random selection of points in training. For unbalanced classification problems, the probability of most class points being extracted is much higher than that of a few class points, which results in various computational imbalances. In this paper, a weighted stochastic gradient descent SVM on-line algorithm is proposed to classify data with unbalanced scale. For most classes, the samples are given smaller weights, while the samples in a few classes are given larger weights. Then the original SVM problem is solved by weighted stochastic gradient descent algorithm, which reduces the offset from hyperplane to minority classes. At the same time, by using the Binary Logarithmic loss function, different weighting methods and the small batch gradient descent algorithm, the accuracy and stability of the algorithm are improved.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
本文编号：2243391
[Abstract]:Random gradient descent (SGD) method has been applied to large-scale support vector machine training. The random gradient descent (SGD) method uses random selection of points in training. For unbalanced classification problems, the probability of most class points being extracted is much higher than that of a few class points, which results in various computational imbalances. In this paper, a weighted stochastic gradient descent SVM on-line algorithm is proposed to classify data with unbalanced scale. For most classes, the samples are given smaller weights, while the samples in a few classes are given larger weights. Then the original SVM problem is solved by weighted stochastic gradient descent algorithm, which reduces the offset from hyperplane to minority classes. At the same time, by using the Binary Logarithmic loss function, different weighting methods and the small batch gradient descent algorithm, the accuracy and stability of the algorithm are improved.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
【参考文献】
相关期刊论文 前4条
1 缪志敏;潘志松;袁伟伟;赵陆文;;一种新的基于SVDD的多类分类算法[J];计算机科学;2009年03期
2 王和勇;樊泓坤;姚正安;;SMOTE和Biased-SVM相结合的不平衡数据分类方法[J];计算机科学;2008年05期
3 李昆仑,黄厚宽,田盛丰,刘振鹏,刘志强;模糊多类支持向量机及其在入侵检测中的应用[J];计算机学报;2005年02期
4 王华忠,张雪申,俞金寿;基于支持向量机的故障诊断方法[J];华东理工大学学报;2004年02期
相关博士学位论文 前1条
1 杨智明;面向不平衡数据的支持向量机分类方法研究[D];哈尔滨工业大学;2009年
，本文编号：2243391
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2243391.html
