一种提高SVM训练速度的新方法

SVM解决两分类问题时,在大规模数据上训练速度很慢,利用数据提取的方法可以减少训练样本数目,加快训练速度。本文利用马氏距离和"aσ-方法"提出新的数据提取方法,根据样本点到训练集的马氏距离来确定样本点与样本集的位置关系,只提取对于建立超平面有作用的样本点,避免了以往数据提取方法的随机性;并考虑提取的数据占原来总样本集数目的比例,通过调整a的值,控制数据提取的数量,避免提取后训练样本集的数据太多或太少,从而加快SVM的训练速度。

A New Method to Increase Training Speed in SVM

In binary classification,the training speed of SVM on large dataset is very slow.So the method of choosing some useful data can decrease the number of training data and accelerate the training speed. This paper proposes a new method using Mahalanobis distance and aσmethod to choose useful data. During the course of choosing data,we determine the position relation between the data and the dataset according to Mahalanobis distance between the data and the dataset.Then we can choose the data useful to construct the hyperplane.Our method avoids the shortcoming of strong randomicity in other methods before.We also take into account of the proportion of data chosen when choosing data.By adjusting the value of a,we can control the number of data chosen,in case of the dataset being too small or too large. So this method can accelerate the training speed greatly.