基于局部子空间的多标记特征选择算法  

Multi-label Feature Selection Algorithm Based on Local Subspace

作　　者：刘景华[1] 林梦雷[1] 王晨曦[2] 林耀进[1] 

机构地区：[1]闽南师范大学计算机学院,漳州363000 [2]漳州职业技术学院计算机工程系,漳州363000

出　　处：《模式识别与人工智能》2016年第3期240-251,共12页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61303131;61379021);福建省自然科学基金项目(No.2013J01028);福建省教育厅科技项目(No.JA14192)资助~~

摘　　要：在已有的特征选择算法中,常用策略是通过相关准则选择与标记集合相关性较强的特征,然而该策略不一定是最优选择,因为与标记集合相关性较弱的特征可能是决定某些类别标记的关键特征.基于这一假设,文中提出基于局部子空间的多标记特征选择算法.该算法首先利用特征与标记集合之间的互信息得到一个重要度由高到低的特征序列,然后将新的特征排序空间划分为几个局部子空间,并在每个子空间设置采样比例以选择冗余性较小的特征,最后融合各子空间的特征子集,得到一组合理的特征子集.在6个数据集和4个评价指标上的实验表明,文中算法优于一些通用的多标记特征选择算法.In the existing multi-label feature selection algorithms, the features with stronger relevance to label set are usually selected according to some related criteria. However, this strategy may not be the optimal option. As some features may be the key features for a few labels, but they are weakly related to the whole label set. Based on this assumption, a multi-label feature selection algorithm based on local subspace is proposed. Firstly, the mutual information between feature and label set is employed to measurethe importance degree of each feature, and original feature sequences are ranked by their importance degree from high to low to obtain a new feature space. Then, the new feature space is partitioned into several subspaces, and the less redundant features are selected in each subspace by setting a sampling ratio. Finally, the final feature subset is obtained by merging all feature subsets indifferent subspaces. Experiment is conducted on six datasets and four evaluation criteria are used to measure the effectiveness. Experimental results show that the proposed algorithm is superior to the state-of-the-art multi-label feature selection algorithms.

关 键 词：特征选择 多标记分类 局部子空间 互信息 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]