局部加权单类支持向量机研究

发布时间：2018-09-08 08:19

【摘要】：单类分类是介于监督学习和无监督学习之间的机器学习任务,它能够有效地解决仅有一类样本训练分类器的问题和类别极端不平衡问题。迄今为止,涌现了大量的单类分类方法,其中单类支持向量机(one-class support vector machine,OCSVM)是最为常用的方法之一。然而,传统的单类支持向量机存在一些不足,如:没有考虑到训练样本的几何分布对其分类器性能的影响。基于此,本文利用训练样本的局部几何信息,对单类支持向量机从间隔改进和错分样本加权两方面开展了研究。1.提出了基于局部相关保留的单类支持向量机(locality correlation preserving based one-class support vector machine,LCP-OCSVM)。所提方法将局部相关保留(locality correlation preserving,LCP)与单类支持向量机相结合,继承了LCP与OCSVM的优点,可在保留样本的局部相关性的同时最大化特征空间中样本的像与原点之间的间隔。在人工数据集及标准数据集上验证了所提方法的可行性。2.提出了局部保留加权的单类支持向量机(locality preserving weighted one-class support vector machine,LPWOCSVM)。为了降低错分样本对单类支持向量机分类边界的影响,所提方法根据训练样本的局部几何信息构造局部保留加权向量,为错分样本分配较小权重,使得单类支持向量机的分类边界更为紧致。在人工数据集以及标准数据集上的实验表明,所提方法具有更强的抗噪声能力和更优的泛化性能。
[Abstract]:Single class classification is a machine learning task between supervised learning and unsupervised learning. It can effectively solve the problem of only one kind of sample training classifier and extreme class imbalance problem. Up to now, a large number of single class classification methods have emerged, among which single class support vector machine (one-class support vector machine,OCSVM) is one of the most commonly used methods. However, there are some shortcomings in traditional single-class SVM, such as not considering the influence of geometric distribution of training samples on the performance of classifier. Based on this, this paper makes use of the local geometric information of the training samples to study the single-class support vector machine from the two aspects of interval improvement and misdivision sample weighting. A single class support vector machine (locality correlation preserving based one-class support vector machine,LCP-OCSVM) based on local correlation reservation is proposed. The proposed method combines local correlation preserving (locality correlation preserving,LCP) with single class support vector machine, inherits the advantages of LCP and OCSVM, and maximizes the interval between the image and origin of the sample in the feature space while preserving the local correlation of the sample. The feasibility of the proposed method is verified on the human data set and the standard data set. 2. 2. A locally reserved weighted single class support vector machine (locality preserving weighted one-class support vector machine,LPWOCSVM) is proposed. In order to reduce the influence of misdivided samples on the classification boundary of single class support vector machines, the proposed method constructs locally reserved weighted vectors according to the local geometric information of the training samples, and assigns a smaller weight to the misdivided samples. The classification boundary of single class support vector machine is more compact. Experiments on artificial data sets and standard datasets show that the proposed method has better anti-noise capability and better generalization performance.
【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】