论文部分内容阅读
大量数据下支持向量机(SVM)的训练算法是SVM研究的一个重要方向和广大研究者关注的焦点。该文回顾了近几年来这一领域的研究情况。该文从分析SVM训练问题的实质和难点出发,结合目前一些主要的SVM训练方法及它们之间的联系,重点阐述当前最有代表性的一种算法——序贯最小优化(SMO)算法及其改进算法。从中可以看到,包括SMO在内的分解算法通过求解一系列规模较小的子问题逐步逼近最优解,从而避免存储整个Hessian矩阵,是解决大规模SVM训练问题的主要方法。而工作集的选择对于分解算法的收敛与否和收敛速度至关重要。
SVM training algorithm under massive data is an important direction of SVM research and the focus of researchers. This paper reviews the research in this area in recent years. Based on the analysis of the essence and the difficulty of the SVM training problem, this paper combines some main SVM training methods and the relationship between them, and focuses on one of the most representative algorithms - sequential minimal optimization (SMO) algorithm and Its improved algorithm. It can be seen that the decomposition algorithm including SMO is the main method to solve the large-scale SVM training problem by solving a series of smaller sub-problems and gradually approximating the optimal solution so as to avoid storing the entire Hessian matrix. However, the choice of working set is crucial to the convergence of the decomposition algorithm and its convergence speed.