一种改进的SVM算法

An Improved SVM Algorithm

当训练集的规模很大特别是支持向量很多时,支持向量机的学习过程需要占用大量的内存,寻优速度非常缓慢,这给实际应用带来了很大的麻烦。文献[4]提出了一种针对大规模样本集的学习策略,该方法虽大幅降低了学习的代价,但存在着一个致命的弱点:如果初始样本集选择不当,SVM的分类精度将得不到保障。基于此,本文引入了“最远邻”,对文献[4]中算法进行了改进。实验表明,采用这种改进的算法不仅保留了文献[4]方法的优点,而且这样获得的分类器的分类精度完全可以与直接通过大规模样本集训练得到的分类器的分类精度相媲美,甚至更优。

When training set is very large,espetially,SVM are many,the process of learning requires a great deal of EMS memory,and the speed of count is very slow,this bings big trouble in practice applications.The literature[4]bings forward a learning stategy aimed at cosmically training set.Although this method largly reduces the cost of learning,it exist a deadiness weakness:if the choice of original training set is improper,the classify precision of SVM will not be guaranteed.So this article improves on the literature [4] through introduce the method of 'the farthest neighbors'.Experiment shows that this improving strategy not only holds the excellence of literature [4],but also obtains a classifier that has the same accuracy as (even better than) classifier by literature[4].