Research On Multi-Label Learning Algorithms With Distance Metric Learning

Multi-label learning deals with training examples each represented by a single instance(feature vector)while associated with multiple class labels,and the task is to train a predictive model which can assign a set of proper labels for the unseen instance.The framework is suitable for describing real-world tasks with rich semantic information.However,due to the combinatorial nature of class labels,computational complexity stands as a major issue for practical use of multi-label learning techniques.This thesis mainly focuses on distance metric learning in multi-label scenarios,which includes the following two aspects:On one hand,based on the structural interaction of features and labels in multi-label data,this thesis proposes a multi-label compositional distance metric learning method Commu,which models the structural interaction information between features and label spaces through the compositional distance metrics,and implicitly considers the high-order correlations between labels.Specifically,Commu assumes that the multi-label distance metric is compositional,and the discriminant information of coding label space is used to construct the compositional base metrics;the semantic similarity matrix of multi-label data is defined based on the collaborative computation of feature space and label space,and the set of triple constraints in the distance metric is generated;based on this,the optimization objective is formalized into a quadratic programming problem w.r.t.the combinatorial coefficient w.Consequently,the iterative optimization algorithm FISTA is used to solve the formalized problem.On the other hand,unlabeled data are readily available in many real-world scenarios while it is usually time-consuming and costly to obtain their labels.In order to make full use of unlabeled data,this thesis proposes a compositional distance metric learning algorithm Celsil under the semi-supervised multi-label learning setting.Based on the framework of Mahalanobis distance,this method uses the compositional representation of M proposed in Commu to jointly optimize the multi-label learner and the distance metric.Specifically,Celsil constructs the discriminant base metrics of M with principal component analysis(PCA).The solution includes two modules: multi-label loss term and manifold preservation term.Based on the multi-label loss term,the ranking relationship between relevant labels and irrelevant labels is investigated to characterize the second-order correlation between labels.Furthermore,based on the manifold preservation term,pairwise constraints are implemented to maintain the intrinsic structure information of data.Thereafter,the optimization objective of the algorithm is formalized as a quadratic programming problem w.r.t.the distance metric combination coefficient vector v and the weight w of the multi-label learner,which is solved by the alternating iteration gradient descent method.This thesis consists of five chapters.Chapter 1 mainly introduces the fundamental concepts,research status and problems to be solved on multi-label distance metric learning;Chapter2 mainly introduces the related research of multi-label distance metric learning,including the current mainstream multi-label distance metric learning algorithms;Chapter 3 introduces the multi-label compositional distance metric learning algorithm Commu.Chapter 4 introduces the semi-supervised multi-label compositional distance metric learning algorithm Celsil.Finally,Chapter 5 summarizes and discusses future works.