Research On Label Distribution Learning Method With Label Restriction

Label distribution learning is a new machine learning paradigm that takes label distribution data as the research object.Because of its good performance in dealing with label ambiguity,label distribution learning has attracted more and more scholars’ attention.The traditional label distribution learning algorithm is mainly designed for the situation where the data is completely labeled.However,in practical applications,it is often faced with the problem of limited labeling.Due to the high cost of labeling,the labels of the collected data are often incomplete,inaccurate,or inexact.Especially in label distribution learning,it is more difficult to give an example of label distribution.Therefore,this paper conducts label distribution learning research in two scenarios with incomplete labeling and the emergence of new classes of labeling restrictions.The main work is as follows:(1)Fragmentary label distribution learning via graph regularized maximum entropy criteriaTraditional label distribution learning algorithms are mainly designed based on fully labeled data,and it is difficult to obtain satisfactory results in incompletely labeled scenarios.In addition,the existing incompletely labeled label distribution learning methods cannot effectively use the relationship between the data,and cannot obtain the problem of direct learning classifiers.Aiming at the limitations of existing label distribution learning algorithms,by constructing a data graph matrix,mining the internal connections between the data,and designing an inductive algorithm combined with a unique classifier.The convergence of the algorithm is further analyzed and verified,and the parameters of the algorithm are analyzed to determine the optimal parameters.Finally,the effectiveness of the algorithm is verified through experiments.(2)Label Distribution Changing Learning with Sample Space ExpandingIn view of the traditional label distribution learning algorithm that cannot deal with the label distribution learning problem that new labels appear and the label distribution changes,based on the idea of graphs and manifolds,a structural information matrix between data is constructed.And compress the existing mark distribution by the scaling factor to reconstruct the new mark distribution.An effective algorithm is designed for this problem,combined with theory to prove the feasibility of the algorithm,and two feasibility theoretical results are obtained,which provides a solid theoretical support for the feasibility of the algorithm.Furthermore,experiments verify the superiority of the proposed algorithm in facial expression recognition and age estimation.Carrying out the research of the above methods has further enriched and perfected the theory and methods of label distribution learning in the context of labeling restrictions,and provided some methods for practical applications.