基于组标签的多标签流特征选择算法

Multi-label streaming feature selection based on group labels

流标签是当前多标签学习领域中一个较新颖的挑战性问题,存在标签空间未定、标签数量不断增加甚至趋于无穷等问题.在多标签学习的特征选择中,每当有新的标签到达时标签空间都将发生改变,传统的多标签特征选择算法需重新进行特征选择,所以不适用.针对此问题,采用将流标签进行分组批量处理的方式,并考虑标签之间的相关性,提出一种新的流式多标签特征选择方法,考虑分组后每组标签内部潜在的关联结构和不同标签组之间的标签差异性,赋予每组标签不同的权重来计算每个特征与标签空间的模糊互信息.同时,结合mRMR(Max-Relevance and Min-Redundancy)的特征选择策略进行冗余特征的剔除,从而挑选最优的特征子集.该方法同时适用于固定标签空间和流式标签空间中的特征选择问题.最后,选取八个多标签基准数据集,采用四种评价指标与已有相关的多标签特征选择方法进行对比实验,实验结果证明了提出方法的有效性和高效性.

Streaming labels are currently a relatively new challenge for multi-label learning,which means that the label space is uncertain,that is,the number of labels is constantly increasing,and even tends to be infinity.In the problem of multi-label feature selection,because the label space changes whenever a new label arrives,traditional multi-label feature selection algorithms need to re-select the feature,so it is not applicable.Aiming at this problem,this paper proposes a novel streaming multi-label feature selection method by grouping streaming label and considering the correlation between labels.This method considers the potential association structure within each group of labels after grouping and the label differences between different groups,and assigns different weights to each group of labels to calculate the fuzzy mutual information between each feature and the label space.At the same time,the feature selection strategy of mRMR(Max-Relevance and Min-Redundancy)is combined to consider the redundancy of the features to select the optimal subset of features.And this method applies both the traditional fixed label space and the novel streaming label space.Finally,we selected eight multi-label reference datasets and 4 kinds of indicators for simulation experiments,the experimental results proved the effectiveness of the proposed method by comparing with the existed multi-label feature selection methods.