针对特征空间中存在潜在相关特征的规律,分别利用谱聚类探索特征间的相关性及邻域互信息以寻求最大相关特征子集,提出联合谱聚类与邻域互信息的特征选择算法.首先利用邻域互信息移除与标记不相干的特征.然后采用谱聚类将特征进行分簇,使同一簇组中的特征强相关而不同簇组中的特征强相异.继而基于邻域互信息从每一特征簇组中选择与类标记强相关而与本组特征低冗余的特征子集.最后将所有选中特征子集组成最终的特征选择结果.在2个基分类器下的实验表明,文中算法能以较少的合理特征获得较高的分类性能.%Aiming at some potential correlation between features in feature space, spectral clustering and neighborhood mutual information are exploited to explore the correlation features and obtain maximal relevant feature subset, respectively. And a feature selection algorithm combining spectral clustering and neighborhood mutual information is proposed. In this paper, the neighborhood mutual information is firstly applied to remove uncorrelated features, and then the spectral clustering is utilized to group features. The features of the same group are strongly correlated and the features of different groups are strongly different. Then, the feature subset strongly associated with class label is selected from each feature group. Finally, all selected feature subsets are collected together to form the final selected features. Extensive experiment is conducted with two different classifiers. Experimental results show that the proposed model effectively improves the classification performance with less features.
展开▼
