Research On Label Distribution Learning By Exploiting Label Correlations

Label distribution learning is a novel learning paradigm for solving label ambiguity.It is a further generalization of single-label learning and multi-label learning.It has been successfully applied to facial emotion analysis,head pose recognition and face age estimation.In the label set of label distribution learning,label correlations exist widely,which helps to improve the performance of label distribution learning.Therefore,from the perspective of the label correlations,this thesis makes an in-depth study of label distribution learning.First,local label correlations is studied.The existing algorithms exploiting local label correlations all assume that the smaller the Euclidean distance between samples,the more likely the samples are to share the same label correlations.Specifically,they all use K-means to divide samples into several clusters,and then mine local correlations in each cluster based on this assumption.However,this assumption is incorrect in some cases and may lead to inappropriate clustering results and biased correlations.In this thesis,we propose a novel LDL algorithm based on adaptive local label correlations.In particular,unlike K-means only based on Euclidean distance between samples,this algorithm introduces the clustering with learnable parameters into the model to achieve adaptive clustering which can be optimized jointly with the objective function.This adaptive approach gets rid of the above assumptions and enables the algorithm to obtain more accurate local label correlations,thus greatly improving the performance.Experimental results on 15 real-world datasets demonstrate the effectiveness of the proposed algorithm.Furthermore,incomplete label distribution learning is studied.In practical tasks,the supervised information of samples is often incomplete,and the lack of supervised information will seriously affect the performance of algorithms which are designed for complete supervised information scenario.The current works of incomplete label distribution learning are based on the nuclear norm relaxation.However,this method requires complex matrix singular value decomposition in each iteration,which limits the scalability and efficiency of the algorithms.In this thesis,we propose a fast incomplete label distribution learning algorithm via augmented label distribution based on label correlations.In particular,the label correlation matrix is used to enhance the observed label distribution matrix to fill the missing label values in the observed matrix.This not only enables us to recover the incomplete observed matrix to a certain extent,but also improves the performance of incomplete label distribution learning while avoiding the high complexity of singular value decomposition in each iteration.Experimental results on 15 real-world datasets with different missing ratios demonstrate the effectiveness of the proposed algorithm.