核化正交平衡类鉴别分析

摘要

In the real world,the distribution of the data is usually nonlinear and uneven. Under this circumstance,it is difficult to extract valid discriminant information by using those traditional discriminant approaches. To solve this problem,propose an approach named Ker-nel Orthogonal Class Balanced Discrimination ( KOCBD) . KOCBD maps the class-imbalanced sample set into kernel space. Then,for a specific class which has fewer samples,KOCBD establishes a nearest sample set in the remaining samples,and then redivides them into some even subsets. At last,KOCBD imposes orthogonal constraint on the extracted discriminant vectors among those high correlated clas-ses to remove redundant information. The experimental results on the Coil 20 and USPS databases demonstrate that the KOCBD approach can effectively solve the class-imbalance problem in nonlinear subspace,and achieve better recognition performance.%现实生活中数据的分布往往是非线性且不平衡的,传统的线性鉴别方法已经很难提取有效的鉴别信息,于是文中将算法扩展到核空间,提出了基于欠采样技术的核化正交平衡类鉴别分析( KOCBD)的方法。该方法在非线性空间中使用核映射,令少样本类为特定类,在剩余样本中构建其近邻样本集,并重新进行平衡类划分,然后提取鉴别特征。为了得到更具鉴别力的特征,进一步去除特征间的冗余信息,文中为相关性大的类之间所获得的鉴别向量加上正交约束。在Coil 20和USPS数据库上的实验结果表明,KOCBD方法能够有效地解决非线性空间的类不平衡问题,识别效果有一定程度的提高。