广义特征值多类分类算法  

Generalized eigenvalue multi-class classification algorithm

作　　者：阳红英[1] 杨志霞[1] 

机构地区：[1]新疆大学数学与系统科学学院,乌鲁木齐830046

出　　处：《计算机工程与应用》2014年第8期154-157,共4页Computer Engineering and Applications

基　　金：国家自然科学基金(No.11161045)

摘　　要：提出了一个新的多类分类算法,该算法的目标是寻找M个相互不平行的超平面,使得第m(m=12M)类的各点到第m个超平面的距离之和尽可能小,而其余类的所有点到该超平面的距离之和尽可能大。基于这个思想,寻求第m个超平面的优化模型最终可转化为一个广义特征值问题。该方法编程简单,易于实现。在数值试验部分,该算法与一些经典的基于支持向量机的多类分类算法进行比较,表明了该算法的优越性。In this paper, it proposes a new multi-class classification algorithm. The task is to find M non-parallel hyper- planes. The m - th（m = 1,2, …, M） hyperplane is constructed such that the inputs of the m - th class are close to the hyper- plane and the rest inputs are far away from it as far as possible. Based on this idea, for finding the m -th hyperplane, the optimization model can be transformed into a generalized eigenvalue problem. The method is easily implemented by a single Matlab command that solves the classical generalized eigenvalue problem. The preliminary numerical experiments show that the method is competitive with several classical multi-class classification algorithms based on support vector machine.

关 键 词：多类分类问题 广义特征值 支持向量机 

分 类 号：O235[理学—运筹学与控制论]