Research On Several Issues Of Multi-Label Feature Representation

In traditional multi-label learning,each object is represented by a single feature vector.The goal of the learning system is to learn a mapping function from the feature space the power set of label space.Generally speaking,the feature representation of objects has an important influence on the generalization performance of learning systems.For multi-label feature representation,it is worthwhile to investigate how to adapt the traditional single feature vector representation to achieve stronger generalization performance.This paper focus on the research on multi-label feature representation,where the following two main contributions have been made:On the one hand,”label-specific features” aims to construct an appropriate feature vector for each label,which is an effective way to improve the generalization performance of multilabel classification system.However,existing label-specific features techniques usually ignore the correlation among class labels.In light of this,a novel multi-label learning approach named Reel is proposed which generates label-specific features by exploiting label correlations.Firstly,the centering points for label-specific features generation on each label are identified via positive and negative examples partition as well as k-means clustering.The feature values for label-specific features are determined based on the Euclidean distance between the instance and each centering point.After that,useful label correlation information from other class labels is integrated by imposing L1 regularization.Experimental studies show that Reel achieves superior performance against state-of-the-art multi-label learning approaches based on label-specific features.On the other hand,in many real-world applications,multi-label objects often have multiview feature representation and contain label noise.In this paper,a novel multi-view multi-label learning approach named Gradis is proposed.Firstly,the affinity graph over training examples is constructed by fusing the multi-view feature representation,which is then utilized to estimate labeling confidences based on label propagation.Then,the embedded representation of multiview features is obtained by spectral clustering analysis,based on which the resulting multilabel predictive model is induced.Compared with existing multi-view multi-label learning approaches,the Gradis approach shows better generalization performance on artificial datasets and real world datasets.This paper is divided into five chapters.The first chapter introduces the research background,related work and problems to be solved.In the second chapter,we introduce the techniques of label-specific feature in multi label learning,and propose the label-specific features learning approach by exploiting label correlations.In the third chapter,we introduce the problem of multi-view multi-label learning,and propose the multi-label learning approach by fusing multi-view representation.The fourth chapter summarizes the whole thesis.