Research On Learning Method Via Self-representation

In high-dimensional data,many samples have redundant or irrelevant features.The existence of redundant features will reduce the efficiency of the algorithm.The presence of uncorrelated features will adversely affect the performance of the learning algorithm.Feature selection can bring many benefits to the algorithm,such as reducing computational cost,improving efficiency and enhancing generalization ability.As data samples continue to grow,the complexity of annotating data structures is also on the rise.Although multi-label learning is now capable of handling a large number of labels ambiguity problems,there are many more in the real world that need to reflect the exact description of each label for an instance,ie markup distributional data.In view of the above data problems,this article applies the self-representation theory to the learning process and proposes two learning methods based on self-representation:?1?An unsupervised feature selection method?DMSR?based on self-describing dependency metrics is proposed for unlabeled data with a large number of redundant and irrelevant features.The algorithm firstly defines that the performance of the feature depends on the self-representation dependence metric of the original data,that is,the more the projected low-dimensional spatial data depends on the original data,the performance is better.Then,by dependency maximization,the data projected into the low-dimensional space can keep the original data's characteristic information as much as possible,and then reduce the original data.After obtaining the reliable low-dimensional data,the sparse representation technique is introduced for feature selection.Finally,experiments were performed on four open datasets and compared with the three existing unsupervised feature selection algorithms.The experimental results show that the proposed method of DMSR feature selection is effective.?2?Label distribution learning by regularized sample self-representation?RSSR-LDL?is proposed for the data which label is a distribution that needs to reflect the accurate description of the instances.First of all,starting from the idea of reconstructing the label distribution,each label distribution can be expressed as a linear combination of the relevant sample features by using the sample feature and the transformation matrix.Then,the model is established by using the residual function between the reconstructed label distribution and the original label distribution.The least square method is used to optimize the model.At the same time,L₂-norm and L_2,1-norm regularization are respectively introduced to solve the problem.Finally,experiments are performed on 12 open datasets and compared with the three existing labeled distribution learning algorithms on five evaluation indexes.The experimental results show that the proposed RSSR-LDL learning method is effective.