特征提取算法KPCA的改进与设计  

The improvement and design of KPCA feature extraction algorithm

作　　者：何新[1] 李大辉[1] 付军[1] 

机构地区：[1]齐齐哈尔大学计算机与控制工程学院,黑龙江齐齐哈尔161006

出　　处：《高师理科学刊》2015年第6期34-37,共4页Journal of Science of Teachers＇College and University

基　　金：黑龙江省自然科学基金项目(F201218)

摘　　要：特征提取的本质就是变换.针对传统核主成分分析(KPCA)在分类问题中所提取出的特征组合的不足,提出了一种基于信息度量改进的KPCA算法.数据集使用广泛应用的KDDCUP99安全审计数据集,用训练样本各特征向量的类内聚集程度和类间离散程度所组成的信息度量来代替传统KPCA中的累积贡献率,选取有利于分类的特征组合.实验结果表明,改进的KPCA方法在较低的维数下就具有较明显的分类效果.Transformation is the essence of the feature extraction. In view of the traditional kernel principal component analysis（ KPCA ） lack of extracted feature combination in classification problems, proposes an improved KPCA based on information measure. Data set uses the normative KDDCUP99 security audit data set. The aggregation degree within class and dispersion degree between class comprise information measure of each feature vector in the training sample. It is used to replace the cumulative contribution rate of the traditional KPCA. The selected feature combination is advantageous to classification. A large amount of experimental results show that the improved KPCA method at a lower dimension will have a more pronounced effect of classification.

关 键 词：KPCA 特征提取 分类 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]