一种新型的增式SVM训练算法

针对传统的增式支持向量机算法在计算时间和分类效率上的不足,提出了一种新型的增式SVM训练算法。该算法不是简单地保留上一步训练的支持向量,而是通过增加KKT(Karush-Kuhn-Tucke)限制条件并对决策函数的输出设定一个阈值,使得保留下来的样本都是最有效的样本,从而可减少训练样本的数目。在仿真实验中,选择了一组UCI数据,并选用RBF核函数作为核函数。实验结果表明:与传统增式算法相比,新算法在保证传统SVM性能的同时,在迭代速度和分类放率上分别提高了14%和4.39%。

An Improved Incremental Training Algorithm of Support Vector Machine

For the insufficiencies of original incremental algorithm of support vector machine in operation time and the efficiency of classification,this paper proposed a new incremental algorithm: it is not only reserves the support vectors of the former training step simply,but also adds KKT(Karush-Kuhn-Tucke)condition as the restrict condition,and sets a threshold to the decision function's output.It makes the reserved samples the most effective.The experiment data are chosen from UCI database,and choose RBF kernel function as the kernel function.According to the original incremental algorithm,The experiment result indicated that the new incremental algorithm not only can conserve the capacity of the original SVM,but also can improve 14% and 4.35% on operation speed and classification efficieng.