基于filterwrapper的两步特征变量提取方法*

作者：陈岩，来海锋，王清，王卫伟 日期：2010-06-28/span> 浏览：3123 查看PDF文档

基于filterwrapper的两步特征变量提取方法*
陈岩，来海锋，王清，王卫伟
（杭州电子科技大学 自动化学院，浙江 杭州 310018）

摘要：特征变量选择是高维数据分类问题的核心，主要有过滤法和缠绕法两种特征变量选择方法。针对“过滤法与分类算法相互独立，不利于对分类性能优化，而缠绕法依赖于分类算法，在高维高噪的数据中容易过拟合”这个问题，为了能有效提取特征变量，提出了一种新的特征提取方法，即filterwrapper两步法，先通过有监督奇异值分解方法降维去噪，粗选出一部分备选变量；再应用MonteCarlo决策树策略从备选变量中精选出重要的特征变量。通过以典型的高维高噪数据为例验证了该方法，实验结果表明了上述方法的可行性和有效性。
关键词：有监督奇异值分解；信息增益；决策树；特征提取；分类
中图分类号：TP391.4文献标识码：A文章编号：1001-4551(2010)04-0067-05

Twostep feature selection algorithm based on filter and wrapper

CHEN Yan, LAI Haifeng, WANG Qing, WANG Weiwei
（School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China）

Abstract: Feature selection is an essential step to classification with high dimensional data. Filter method is independent of classifiers, and hence the classification quality irrelevant, wrapper method is a classification supervised method, which is prone to overfitting in high dimensional and high noise data. Aiming at these problems of filter and wrapper methods, a new feature selection approach was proposed, which combined filter and wrapper methods, first supervised singular value decomposition was used to reduce variance dimension and extract domain relevant metafeatures, then MonteCarlo decision trees was used to select important variables. This method is applied to three gene expression profiles datasets, and a better gene subset is obtained in contrast with other typical methods, which shows the feasibility and effectiveness of the methods proposed.
Key words: supervised singular value decomposition(SSVD); information gain; decision tree; feature extraction; classification