面向多类学习问题的核最近表面分类方法

Kernel Nearest Surface Classification Methods for Multi-class Learning Problems

虽然最邻近决策规则能很好地解决数据集的非线性和非平衡性问题,但其没有学习过程.在此基础上,提出了一种利用聚类方法来浓缩训练样本,再根据最近邻准则进行决策的方法——核最近表面分类方法.通过实验将其与几种常用的统计分类方法进行对比,结果表明,核最近表面分类方法具有决策速度快、存储空间需求小等优点,同时也能够很好地处理非平衡数据集的分类问题.

The nearest neighbor decision rules have a good classification on nonlinear and imbalanced data sets, but have no learning process. A method called as kernel nearest surface classifier is proposed to condense the training samples with clustering, and then make decision using the nearest neighhor rule. The method is compared with such as common-used statistical classification methods through the experiments. The results show that the proposed method has many merits, such as fast decision speed and small memory space requirement, and can also solve the classification problem of the imbalance data set effectively.