对利用超平面、超(椭)球等几何形状实现数据分类的基于边界的主流分类方法进行了研究, 在此基础上, 提出了一种将空间点作为分类依据的最大间隔模糊分类器(MFC)。该方法首先在模式空间中找到一个模糊分类点c, c点距离两类样本要尽可能近且类间夹角间隔尽可能大。然后, 测试样本通过c与训练样本间的最大化夹角间隔实现分类。最后, 利用MFC的核化对偶式与最小包含球(MEB)的等价性, 将MFC的应用范围从二类推广到单类。与主流分类方法的比较实验表明, MFC具有优良的分类性能和抗噪能力, 其分类最高精度可达98.8%。

Several kinds of current boundary classification methods based on hyperplane, hypersphere or ellipsoid were studied, and a novel classification method called Maximum-margin Fuzzy Classifier (MFC) was proposed by using a space point as the classification criterion. By the proposed method, a fuzzy classified point c was chosen in the pattern space firstly, which should be as close to two classes as possible. Moreover, the angle between the two classes should be also as large as possible. Then, the testing points could be classified in terms of the maximum angular margin between c and all the training points. Finally, the applications of the MFC were popularized from two-class classification to one-class classification according to the kernel dual of MFC equivalent to the Minimum Enclosed Ball (MEB). Comparative experiments with current classification methods verify that the MFC has good classification performance and noise resistance ability and its classification accuracy has been reached 98.9%.