一种基于多支持向量机的增量式并行训练算法

An Incremental Parallel Training Algorithm with Multiple Support Vector Machine Classifiers

为了改进传统算法,利用支持向量的特性,提出了一种基于多支持向量机的增量式并行训练算法(PMSVM)。选择对分类超平面有影响的样本点作为支持向量,以增加单个分类器的训练时间为代价换取整体训练和分类的精度。考虑到训练样本的分布对最终结果的影响,加入反馈向量进行适当的重复训练,以调整各分类器的学习性能。通过在测试数据集上进行的实验表明,该算法与批学习增量BSVM算法相比,在提高训练效率和分类精度的前提下,大大降低了训练时间。

To improve traditional algorithms and take advantage of characteristics of support vectors, an incremental parallel training algorithm with multiple SVM classifiers is presented in this paper. In this parallel algorithm, the samples point which affects the classified hyperplane is chosen as support vector to gain a sufficient accuracy of the whole training and classification process at the price of increasing the training time of single SVM classifier. Considering the relationship between training samples distribution and the form of hyperplane, some samples are taken as feedbacks for use in appropriate repeating training to adjust the learning performance of each classifier. The practical experiment results show that compared with the batch SVM, this parallel MSVM training algorithm is efficient and can significantly reduce training time with high training efficiency and classification accuracy.