基于互信息的Fisher Score多标记特征选择

Multilabel feature selection based on Fisher Score with mutual information

目前,Fisher Score模型在处理多标记数据时没有考虑样本和整个特征空间之间以及特征和标记之间的关系.提出一种基于互信息的Fisher Score多标记特征选择方法.首先,在多标记决策系统中考虑整个样本空间对特征选择的影响,根据异类样本与同类样本之间的欧式距离定义权重公式,并在特征空间下对标记赋予权重衡量标记的重要程度.然后,基于互信息理论定义特征与每个标记之间的互信息来计算每个特征和每个标记之间的相关度,将特征与标记之间的相关度与该标记所占的权重相结合来定义特征和标记集之间的总相关度.将Fisher得分与总相关度结合,定义每个特征的新的Fisher得分,进而构建多标记Fisher Score模型.最后,设计了一种基于互信息的Fisher Score多标记特征选择算法.在六个多标记数据集上的实验证明,提出的算法与其他算法相比,其四种评价指标都表现良好,分类性能出色.

When processing multilabel data,the current Fisher Score models do not consider the relationship between samples and the entire feature space,and also ignore the relationship between features and labels.To address the issues,a multilabel feature selection method based on Fisher Score with mutual information is developed in this paper.Firstly,in multilabel decision systems,to consider the influence of the entire sample space on feature selection,a weight formula is defined based on the Euclidean distance between the heterogeneous samples and the similar samples,and the weights are assigned to labels in feature space to measure the importance of labels.Secondly,the mutual information theory is introduced to define the mutual information between the feature and each label to calculate the correlation between each feature and each label.When combining this correlation between feature and label with the token weight of the label,the total correlation between the feature and the label set is defined.This total correlation is employed to design a new Fisher Score for each feature,and then,a multilabel Fisher Score model is constructed.Finally,a multilabel feature selection algorithm based on Fisher Score with mutual information is designed.Experimental results applied to six multilabel datasets show that the proposed algorithm shows great classification performance in terms of four evaluation metrics when compared with the other related algorithms.