用LDL^T并行分解优化大规模SVM的训练效率

Improving SVM’s learning efficiency by using matrix LDL^T parallel decomposition

支持向量机在大规模训练集上学习时,存在学习时间长、泛化能力下降的问题。研究使用路径跟踪内点法构建面向大规模训练集的SVM学习算法,找到影响算法学习效率的关键是求解大型线性修正方程,首先使用降维法降低修正方程的维数,再使用矩阵LDLT并行分解高效地求解子修正方程,达到优化大规模SVM学习效率的目的,实验结果说明SVM训练效率提升的同时不影响SVM模型的泛化能力。

If the support vector machine is trained on large-scale datasets, the training time will be longer and generalization capability will be descended.Path following interior point method is proposed to design the SVM＇s learning algorithm on large-scale datasets, and the key point being negative impact on SVM＇s learning efficiency on large-scale datasets is to solve the large-scale iterative direction equations efficiently.To improve the SVM＇s learning efficiency, the dimensions of direction equations are degraded,then LDLr parallel decomposition method is used to solve the direction sub-equations efficiently. The experimental results show that the new SVM＇s training algorithm is efficient for large-scale datasets and the generalization capacity of SVM is not affected.