| To deal with label ambiguity,the researchers proposed a new machine learning paradigm——label distribution learning.Compared to traditional learning paradigm for solving label ambiguity,label distribution learning is no longer concerned with whether each label belongs to the instance when using labels to describe the instance.Instead,it gives each marker describing the instance a specific description degree,which can accurately represent the difference in the description degree of each marker in the instance and strengthen the ability of labeling information.The task of label distribution learning is to build models from known training data to predict the label distribution of unknown instances.In recent years,label distribution learning has attracted more and more researchers’ attention.Most of the previous label distribution learning methods either do not consider the correlation between labels,or exploit the label correlations in a global way.However,in real-world tasks,label correlations that can be shared by all instances are virtually impossible.In most cases,label correlations are shared only by locally similar instances.Besides,the previous label distribution learning methods directly use the original high-dimensional feature space for calculation,without giving the label specific selection features.The performance of the algorithm is affected by irrelevant and redundant features.Aiming at these problems,an algorithm named Label Distribution learning based on local label correlations(LDL-LLC)is proposed.The main research contents are as follows:The training data are grouped and the label correlations of each group of training data is constrained on the output of labels.And the Laplacian matrix in the local manifold regularizers is not generated by specifying any marker correlation matrix in advance,but is iteratively updated as a variable to explore and utilize the local label correlations.Besides,the weight matrix is constrained by regularization which is commonly used in feature selection to learn label-specific features and the common features for all the labels.Finally,the weight matrix is used to predict the label distribution of unknown instances.The experimental results on several real label distribution learning data sets validate that LDL-LLC produces good performance and stable effect. |
