| With the increasing of data dimensions in various fields,problems such as high computing cost,high storage cost and over-fitting have brought great obstacles to multi-label learning.In order to deal with these problems,more and more multi-label feature selection methods are proposed to eliminate redundant and irrelevant features.At home and abroad in recent years,many excellent multi-label feature selection methods have been proposed,but there are still some difficult problems:（1）more labels tend to have a lot of noise in the data,the label information is not entirely accurate,how to improve the robustness of the algorithm,reduce the stray data points,wrong label information and redundant information on the impact of the training process is a difficulty.（2）The relationship between the features and labels for multi-label data is more complex than the single-label data,the correlation between features and labels is also more difficult to use,and the correlation of the data in the multi-label feature selection usually have great influence on the algorithm performance,as a result,while retaining the labels,features and the correlation between the labels and features is a key issue.（3）Redundant information inevitably appears between features and labels.How to reduce the impact of redundant information on feature selection is a key problem to be solved.Based on the above problems,this thesis proposes three multi-label feature selection algorithms for in-depth study and discussion,as shown below:（1）This thesis proposes a multi-label feature selection method（GMFS）combining graph and coupling matrix decomposition.By introducing coupling matrix decomposition and effect on the two original matrix,can extract their shared latent semantic matrix,and the latent semantic matrix has less redundant or irrelevant information,at the same time,the manifold regularization will be used to retain the local label manifold information embedded in the sample space,help the mapping matrix model training out more information.In order to further preserve the dependency information of label matrix and feature matrix,an improved mapping matrix is introduced in this paper.（2）This thesis proposes a robust multi-label feature selection method（RGFS）based on graph and feature label correlation.RGFS extracts the low-dimensional latent semantic matrix of labels through matrix factor decomposition,retains the important information of labels,and enables the algorithm to have certain performance on the dataset with particularly sparse labels.Moreover,RGFS preserves the local manifold structure of features and labels by adding feature-manifold regularization and label-manifold regularization.In order to deal with the impact of noise data on the learning process,2,1-norm is applied to the loss term,which makes RGFS have certain robustness.Finally,we add the correlation between features and labels to the weight matrix by improving the weight matrix.The improved weight matrix preserves the labels,features,and the correlation between labels and features,and greatly retains the information of the feature matrix and the label matrix.（3）This thesis proposes a multi-label feature selection method called LMFS that considers local correlations.Existing embedded multi-label feature selection methods are difficult to use local features and label information for model training.LMFS,based on the least squares regression model,extracts the local features and label information of the data by clustering,and guides the model performs multi-label feature selection.At the same time,in order to make the model have a certain robustness,LMFS uses the2,1-norm to constrain the least squares term.Finally,this paper uses an iterative gradient descent method to solve the objective function.Finally,the three algorithms are compared with several multi-label feature selection algorithms with excellent performance in recent years on multiple domain data sets and prove the effectiveness of the three proposed methods. |
