关键词: 粗糙集,一类支持向量机,加权核函数,主成分分析,超平面,过拟合

Abstract: Rough one-class support vector machine(ROCSVM) is a single class SVM.It defines upper approximation and lower approximation hyperplanes by a kernel function mapping,which makes the training samples have an impact on the decision hyperplane adaptively according to the position within the rough margin.Since the ROCSVM only has positive samples,to fully exploit and use the features of the classified training samples have important significance for improving the classification performance of ROCSVM.Thus,we presented a weighted feature-contribution-degree(WFCD) based Gaussian kernel(λ-RBF).First,principal component analysis(PCA) is done to the training set to get vector set sorted by eigenvalues,and then kernel function is constructed based on the vector set,which makes a larger eigenvalue have better effect in the kernel function.Experimental results on UCI standard data sets and simulation data show that compared with the general RBF-based ROCSVM,the λ-RBF based ROCSVM has better generalization and higher re-cognition rate.

Key words: Rough set,One-class SVM,Kernel function,PCA,Hyperplane,Over-fitting