The label-specific features learning is a kind of framework for extracting the specific features of each label for classification. At present, the label-specific features algorithm is generally based on the original label space to find a particular feature set. This kind of extraction method for label-specific features has a general adaptation when the label density is balanced. However, in most multi-label data sets, the number of positive and negative labels varies greatly, and the label density is unbalanced. And there is an inherent correlation between each label. When extracting label-specific features, the method needs to consider issues such as imbalanced label density and label correlations. Based on this, a label-specific features learning method based on Joint Label-density-margin Space and Extreme Elastic Net (JLSE2N) is proposed in this paper. Firstly, to enlarge the margin between the positive and negative labels in the original label space, the positive and negative label density of the total example are calculated. And the information of unbalanced label density is integrated into the label space, thus forming a priori knowledge of the label-density-margin space. Subsequently, to quickly extract the label-specific features, the elastic net regularization extreme learning machine is used for the first time to obtain the label-specific features. The transformed label density space is used to calculate the cosine similarity and is added to the L2 regularization term to consider the correlation between pairs of labels. The L1 regularization term in the elastic net can generate the sparse weight matrix to extract the required label-specific features. We compare the proposed method with five well-known algorithms on 11 benchmark data sets. Experimental analysis shows that our proposed method is superior to the state-of-the-art algorithms for multi-label learning.
1 May 2020 in JOURNAL OF SHENZHEN UNIVERSITY SCIENCE AND ENGINEERING
In view of the fact that most of the existing feature selection algorithms do not consider the possible differences existing in the sample description by different labels, a multi-label feature selection algorithm with imbalance label otherness (MSIO) is proposed. The frequency distributions of positive and negative labels under different labels are added to the process of feature selection as the label weight, the traditional method of calculating information entropy is modified to get a more efficient feature sequence. Based on ML-kNN (multi-label k-nearest neighbor), the features are classified on 11 multi-label benchmark datasets of Mulan database, and the algorithms of multi-label dimensionality reduction via dependency maximization (MDDM), pairwise multivariate mutual information (PMU), feature selection for multi-label naive Bayes classification (MLNB), multi-label feature selection with label correlation (MUCO) and MSIO algorithm are evaluated. Experimental results and statistical hypothesis tests show that MSIO algorithm has good stability, high classification accuracy, and certain effectiveness and superiority.