一种SVM训练样本集寻优算法  

A SEARCH ALGORITHM FOR SVM TRAINING SAMPLE DATASET

作　　者：吴东洋[1] 业巧林[1] 业宁[1] 张训华[1] 武波[1] 

机构地区：[1]南京林业大学信息技术学院,江苏南京210037

出　　处：《计算机应用与软件》2010年第9期14-16,19,共4页Computer Applications and Software

基　　金：国家自然基金(30671639)资助;江苏省科技创新项目(CX08S-010Z)

摘　　要：首先运用Bagging算法解决样本数据变化带来的不稳定性,然后运用网格搜索法寻找合适的训练样本尺寸,再结合两者的特点,提出了一种自助网格搜索算法,从多个支持向量机(SVM)分类器中寻求一个最优的SVM分类器。实验结果表明,算法有效地提高了分类器的学习精度与学习性能,对大样本数据来说,可以用相对较少的样本进行训练后的性能来预测它对一个非常庞大的训练集的性能,大大减少了SVM训练的时间。Generally speaking, the training sample＇ s alteration will cause the classifier to change. Selecting appropriate training samples is especially significant to constructing the classifier with superior performances. In practical training of the classifier, the alterations of the sample data and sample size are to lead the variation of the SVM classifier. In this paper, the hagging algorithm is used first to solve the instability caused by the alteration of sample data, and then the grid search algorithm is used to find appropriate training sample size. Based on the char- acteristics of these two algorithms, the self-help grid search algorithm is proposed for searching an optimal SVM classifier among many of them. Experimental result shows that the SVM Classifier found with this algorithm increases its learning precision and performance efficiently. As to the large sample data, through the search with this algorithm,it can predict the performance of a huge training set by the performance of the trained sample which is relatively small, which remarkably decreases the training time of SVM so as to save the time cost.

关 键 词：SVM分类器 BAGGING算法 自助网格搜索算法 训练样本数量 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]