噪声消除与SMO算法收敛性

近年来,随着序列最小优化分类算法SMO等一系列快速算法的推出,支持向量机在自动文本分类研究领域取得了很大的成功。大多数文本分类问题是线性可分的,使用线性核函数的SMO算法能够取得非常好的分类效果。但是文本向量是一种非常稀疏的向量,采用线性核函数的SMO算法对噪声样本非常敏感,容易产生发散的问题。文章分析证明了噪声如何影响SMO算法收敛性。为了解决训练样本中噪声样本影响SMO算法收敛的问题,设计了一个消除噪声样本的算法,取得了非常好的效果。

Eliminating Noisy and SMO Algorithm Convergence

In recent years,accompany with the appearance of a series of rapid training algorithm as Sequential Minimal Optimization(SMO),support vector machines achieved great success in text categorization.Most text categorization problems are linearly separable,and SMO algorithm using linear kernel-induced can perform well for text categorization.However,text vectors are a kind of extremely sparse vector,and SMO algorithm with linear kernel or polynomial kernel is very sensitive to the extremely sparse noisy example,which is easy to bring on the problem that algorithm can not converge.It is been proved that the noisy example how to influence the convergence of SMO algorithm in the paper.To solve the problem that noisy sample in training samples affect the convergence of SMO algorithm,this paper designs the algorithm that can eliminate noisy samples,and good results is achieved.