支持向量机的缺陷及改进算法

摘要

Traditionary support vector machines (SVMs) usually focus on edge patterns of data distribution, and support vectors (SVs) usually generates from these patterns. This paper proposes an alternative algorithm, which generates SVs from all training patterns. The sparsity of the algorithm is validated on most data sets far better than typical SVMs. The complexity of the algorithm in multi-class problems is merely equivalent to two class SVMs, which greatly solves the problems of too many variables or too many binary classifiers in multi-class SVMs.%传统支持向量机通常关注于数据分布的边缘样本,支持向量通常在这些边缘样本中产生.本文提出一个新的支持向量算法,该算法的支持向量从全局的数据分布中产生,其稀疏性能在大部分数据集上远远优于经典支持向量机算法.该算法在多类问题上的时间复杂度仅等价于原支持向量机算法的二值问题,解决了设计多类算法时变量数目庞大或者二值子分类器数目过多的问题.