矩阵低秩近似在支持向量机中的研究-应用数学专业毕业论文.docxVIP

Abstract The Support Vector Machine is an efficient machine learning methods, and applies to classification, forecasting and many other areas. However, the kernel which formed by large-scale data in the training process is often difficult to store or participate in the calculation, which makes the algorithm itself not easy to achieve or the complex is too high. Reduced Support Vector Machine is such a method: pre-select a subset of the training data as support vectors, and then solve a smaller scale problem. It is an effective solution compared with SVM which use nonlinear kernel processing large-scale data difficulties. The reduced set selection method is the focus of this study. The original reduced set selection method is selecting 1% ~ 10% of training samples completely randomly; IRSVM method increases the training samples by solving a least squares problem gradually; Systematic sampling RSVM (SSRSVM) starts from a small reduced set, then perform the RSVM to get a classifier, and sample those points that are misclassed to increase more useful support vectors. This paper, based on the facts above, synthesizes the advantages of IRSVM and SSRSVM, namely select those points that are misclassed by solving a least squares problem instead of the original one, and show that the improved algorithm can achieve better classification effect without any increase in complexity through the experimental result. In addition, from a large number of experiments, we find an interesting result that some key points of the training set have great influence in the classification effect. We will illustrate the rationality of these key points by comparing eigenvalue and eigenvector of the approximate kernel with those of the full kernel. Keywords：SVM Low-rank Approximation Reduced Set Kernel 目录 第一章 绪论1 1.1 研究背景1 1.2 支持向量机1 1.2.1 传统支持向量机(SVM)模型介绍 1 1.2.2 传统支持向量机(SVM)的问题 5 1.3 低秩近似方法在支持向量机中的研究6 1.4 本文主要内容6 第二章 约简支持向量机9 2.1 约简支持向量机模型9 2.2 光滑支持向量机10 2.3 最小二乘支持向量机12