Binary Decomposition Methods For Multi-Label Learning

In multi-label learning, a single instance could be associated with multiple class labels. Multi-label learning has been attracting increasing attention due to the capacity of handling multiple semantic annotations for real-world objects. In this thesis, we focus on a subfield of multi-label learning, the binary decomposition methods, which transform multi-label tasks into multiple binary classification tasks. Accordingly, three novel approaches have been proposed with resepct to the binary decomposition strategy:Firstly, label correlations exploitation is one of the most important issues in multi-label learning research. Considering that existing binary decomposition methods handle label correlations with a full-order or random mechanism, we propose a binary decomposition method CTRL, which employs a two-stage filter to search those labels which are much more relevant than others for each class label.Secondly, class-imbalance is a significant issue in machine learning, especially in multi-label learning when the label space is large. Therefore, we propose the COCOA method, an effective binary decomposition approach, which transforms multi-label problems into multiple tri-class problems, to solve the class-imbalance problems and handle label correlations simultaneously.At last, in real-world scenarios, the associated labels, which are regarded to be relevant for the objects, could share different labeling-importance. However, the labeling-importance information in multi-label training set are not accessible to the algorithms. Thus, the LIABLE method is proposed, which fidentifies the implicit diverse labeling-importance information by a label propagation process in order to achieve a better performance of classification.There are five chapters in this thesis. Firstly, we give a brief overview of multi-label learning research and the definition of binary decomposition methods in Chapter 1. Then we introduce the CTRL method and the results of experiments in Chapter 2. Thirdly, we introduce the COCOA method and the results of experiments in Chapter 3. Fourthly, we introduce the LIABLE method and the results of experiments in Chapter 4. In last chapter, we summarize the whole thesis and present several future work directions.