Label Space Dimensionality Reduction Algorithms Based On Label Local Correlation

In multi-label classification problems,a instance has multiple labels in multi-label data sets and there exists certain correlation between labels,which makes multi-label classification problems difficult to be handled by a simple classifier.Label space dimensionality reduction algorithms(LSDR)reduce the label dimensionality to eliminate the correlation between labels,thus solving multi-label problems.LSDR usually use a coding matrix to preserve the correlation between labels,and projects labels to independent label space with smaller dimension.Then,the features and corresponding low-dimensional labels are sent to a learner for training.The learner's prediction for a new unknown instance also needs to be decoded to recover the original label space in order to get the final prediction result.Based on the local correlation property of labels,this paper builds two models of label dimension reduction applied to multi-label classification.(1)Explicit dimensionality reduction model based on minimum label local variance(ML-mLV).(2)An implicit dimensionality reduction model based on the minimum label local distance reconstruction(ML-mDR).ML-mLV maximizes the global variance of labels,minimizes the local variance of labels and combines the feature information with the Hilbert-Schmidt independence criterion to obtain an orthogonal coding matrix.The coding matrix is used to achieve the latent label space.In the experiment,four comparative algorithms which are PLST,CPLST,MLC-BMa D and Fa IE respectively are selected in this paper.The comparative experiments with the ML-mLV algorithm proposed in this paper are performed on four multi-label data sets.The ML-mLV proposed in this paper achieved the best experimental results.ML-mDR uses the idea of Laplacian feature mapping to minimize the reconstruction error of the local distance of labels,while maximizing the global variance of labels,and directly obtains the corresponding label of the original label in the latent label space.In the experiment,this paper compares ML-mDR with four comparative algorithms which are PLST,CPLST,MLC-BMa D and Fa IE respectively on six multi-label data sets.Experimental results show that ML-mDR proposed in this paper has better performance.