Research On Semi-supervised Label Distribution Learning And Label Enhancement Algorith

Learning with ambiguity has attracted more and more attention in machine learning area.Label distribution learning is a new machine learning framework proposed to solve the problem of label ambiguity.Compared with single-label learning and multi-label learning,Label distribution learning is more concerned with the relevance degree of each label to unknown instances,which is a more generalized machine learning paradigm.This learning paradigm gives different label description degree of the same instance in the form of label distribution.In practice,annotations are often given by human annotators,thus assigning each label a real value to indicate its association with a particular instance will result in a large cost in labor and time,especially when there is a large number of labels and instances.Therefore,how to obtain the Label distribution data set becomes very important.Label enhancement is proposed to transform multi-label data set into label distribution learning data set.In recent years,more and more label distribution and label enhancement algorithms have been proposed,but most of them are based on supervised learning.In order to reduce the high cost of data annotation process,this paper proposes a label distribution learning algorithm and a label enhancement algorithm from a semi-supervised perspective,respectively,and implements a sentiment analysis system.Firstly,a semi-supervised label distribution learning algorithm based on projection graph embedding is proposed in this paper.It assumes that the feature space of the dataset is complete,while only a few instances are with label distribution,and most instances of the label distribution do not have manual annotating.Based on this assumption,the complete label distribution space can be obtained.We seek a potential space by orthogonal neighborhood preserving projections,named capture space.This capture space is used to select more valuable features and construct a graph that contains more accurate data structure information.We utilize the sample correlation information contained between graph nodes to recover the unknown label distribution.In addition,compared with fixed graphs in traditional semi-supervised learning,we carry out projection and graph construction simultaneously to obtain a selfupdating projection graph,which is more helpful to learn label distribution.The experimental results validate the effectiveness of the proposed algorithm.Secondly,a semi-supervised label enhancement algorithm based on structured semantic extraction is proposed.It is assumed that the feature space of the dataset is complete,while the logical label space is partially missing,that is,there are a small number of instances with a label value of 0 or 1,most instances have no logical label value,and the label distribution space is completely missing.Based on this assumption,we propose a label enhancement algorithm that can directly recover the complete label distribution from a few logical labels.Firstly,we extract self-semantic of sample by expressing inherent ambiguity of each sample in the input space appropriately,and fill in the missing labels based on this kind of information.Secondly,we take advantage of low rank representation to extract the inter-semantics of between samples and between labels,respectively.Finally,we apply a simple but effective linear model to recover the complete label distribution,by utilizing the structured semantic information including intrasample,inter-sample and inter-label based information.Extensive comparative experiments validate the effectiveness of the proposed method.Thirdly,this paper implements an emotion analysis system.In facial emotion analysis,facial expressions are often the result of a mixture of basic emotions,such as happiness,sadness,surprise,anger,disgust and fear.And these basic emotions are often expressed in different intensities in a specific expression.Therefore,these two algorithms can be used for emotion analysis,and multiple emotion description degrees can be learned for each instance at the same time to comprehensively express the instance.