Research On Multi-label Learning Algorithms With Ensemble Learning

In traditional machine learning,a sample is only labeled with a single class.However,many real-world objects have multiple semantic meanings simultaneously.In order to model these ambiguous objects,a large number of multi-label learning approaches have been proposed.In the multi-label learning paradigm,each object is represented by a single instance which de-scribes its characteristics,while associated with a set of labels simultaneously.As an important machine learning method,ensemble learning can effectively improve the generalization per-formance of learning algorithms.In this thesis,the problem of multi-label learning is studied through ensemble learning,mainly containing the following two works:On the one hand,in many real-world applications,a large amount of data can be ob-tained easily,while labeling these data is extremely time-consuming and expensive,especially in multi-label learning.Therefore,it is worth studying how to utilize the unlabeled data for performance improvement of the learning system.In this thesis,we extend co-training,a tra-ditional semi-supervised learning method,to multi-label learning scenarios,and a novel semi-supervised multi-label method named COINs(CO-training for INductive Semi-supervised multi-label learning)is proposed.Compared with existing transductive semi-supervised multi-label learning method,Caoins can realize inductive learning modeling,with better generalization per-formance.On the other hand,in multi-label learning,each label is supposed to possess specific char-acteristics of its own.Correspondingly,exploiting label-specific features to learn from multi-label data serves as one of the most important multi-label learning techniques.LiFft generates the label-specific features by clustering method to improve the generalization performance of the multi-label learning system,while it ignores the label-correlation which plays an important role in multi-label learning.In this thesis,a multi-label learning method named LiFTACE(multi-label learning with Label-speclfic FeaTures viA Clustering Emsemble)is proposed,which gener-ates label-specific features by considering label correlations via clustering ensemble techniques.Compared with LIFT,LIFTACE can obtain better generalization performance by effectively uti-lizing clustering ensemble.There are five chapters in this thesis.In Chapter I,we give a brief overview of multi-label learning research and the problems remaining to be studied.In Chapter 2,formal definition on multi-label learning is given and five representative multi-label learning algorithms are in-troduced.Chapter 3 and 4 respectively introduce two kinds of multi-label learning algorithm based on ensemble learning,including LIFTACE and COINS,together with their detailed exper-imental results.In Chapter 5,we summarize the whole thesis and present several future work directions.