摘要
W.J.Hu提出的主分量分类器(PCC)通过最大化两类样本在分类面法方向上的投影代数和,实现样本分类.PCC是基于样本的统计平均特性,所以少量的野值对分类面方向的确定影响较小,而SVM对野值较为敏感.PCC与支持向量机相比具有较好的鲁棒性.但是PCC对野值的处理等同于其他样本,尽管有效果,但仍会影响分类面的求取,同时也缺乏直观上(或物理上)的解释,而且没有考虑随机噪声对分类面的影响.鉴于此,在PCC的基础上进行改进,引入模糊思想,设计了一组模糊型的主分量分类器,进一步弱化野值和随机噪声对分类面的影响.人工数据集和Beachmark数据集上的实验证明了新分类器的有效性.
Principal Component Classifier (PCC), a new linear classifier, was proposed by W. J. Hu. It separates the two-class samples through maximizing the algebraic sum of projection on the normal vector of justification plane. Based on statistically average character, PCC is immune to outliers, while support vector machine (SVM) is sensitive to outlier. On this point, PCC has better robustness compared to SVM. However, it is unreasonable for PCC to treats outlier as other normal samples. This also has no geometrical (or physical) interpretation.
关键词
主分量分类器
支持向量机
野值
核化
principal component classifier
support vector aachine
outlier
kernelized