The Research Of Machine Learning Methods Based On Label Distribution Learning

To deal with label ambiguity,people proposed a new learning paradigm——label dis-tribution learning.Compared to traditional multi-label learning,label distribution learning can deal with the problem that different labels have different description degrees,and thus,it can make better use of the relevance among labels.Label distribution learning has been applied to many applications,such as facial age estimation,facial expression recognition,and multi-label ranking.However,there still remains some problems.First,in the aspect of data,label distri-bution learning requires the format of the training labels must be a distribution,while current machine learning researches mainly focus on classification problems.There are very few ap-plications that can satisfy the demand.This limits the applicationl range of label distribution learning.Second,in some areas(for example,facial age estimation),the label distributions can be learned from the classification data via an adaption process with the prior knowledge.However,the adaption process needs large scale of data and it will be failed in the cases of limited training data.Finally,in the aspect of method,the number of label distribution learning algorithms is still little,and many effective machine learning algorithms(for example,support vector machine)have not been introduced into label distribution learning.The target of this paper is to study the above problems and propose the corresponding solutions.This thesis contributes on the following aspects:1.expand label distribution to the fistly proposed concept called "label manifold" with taking multi-label learning as an example,and propose the multi-label manifold learning(ML2)algorithm which reduces the high demand of label distribution learning for input data and can be applied on traditional classification data di-rectly;2.propose the semi-supervised adaptive label distribution learning(SALDL)algorithm with taking facial age estimation problem as an example,and improve the performance of label distribution learning in the cases of limited labeled training data via the utilization of unlabeled data;3.propose label distribution support vector regression(LDSVR)algorithm,which ex-pands the classical support vector machine algorithm and uses the kernel trick and improves the performance of label distribution learning;4.propose to apply label distribution learning to the rating problems with taking the movie rating distribution prediction problem as an example,and collect and public the dataset,which widens the range of applications of label distribution learning.