基于加权信息粒化的多标记数据特征选择算法

Feature selection algorithm of multi-labeled data based on weighted information granulation

特征选择能去除不相关和冗余的特征,是解决多标记数据维度灾难的有效工具。现有的多标记特征选择算法没有考虑标记空间存在的相关性,认为每个样本的相关标记的重要性相同,并且忽略了特征空间可能是标记重要性差异形成的内在因素,使得选择的特征不能精确全面地刻画样本且计算过程复杂。为此,本文利用标记间的相关性对标记空间进行划分以简化计算,并定义标记重要性度量和特征权重,在此基础上提出了一种基于加权信息粒化的多标记特征选择算法。通过在真实多标记数据集上的实验对比分析,本文提出的算法在各项评价指标上均优于其他对比算法,验证了算法的有效性和可行性。

Feature selection can remove irrelevant and redundant features.It is an efficient tool to solve the disaster of multi-labeled data dimensions.Existing multi-labeled feature selection algorithms did not take the correlation of label space into account,and considered that the relevant labels of each sample have the same importance,and ignored that the feature space may be the internal factor caused by the difference of label importance,so that the selected features can not accurately and comprehensively describe the samples and the calculation process is very complex.In this paper,the correlation between labels is used to divide the label space to simplify the calculation.Then,the label importance measure and feature weight are defined.And further,a feature selection algorithm of multi-label data based on weighted information granulation is proposed.The comparison and analysis on real multi-labeled data set of experiment show that the proposed algorithm is superior to other comparison algorithms in all evaluation indicators,which verifies effectiveness and feasibility of the algorithm.