Study On Multilabel Ranking With Multiple Labeling Source

From the multiclass problem to multilabel classification problems, then to the label ranking problem, and finally to the multilabel ranking problem, Ambiguity is growing, but also more in line with the uncertainties and complexity of real-world. There exist two major tasks in supervised learning from multi-label data:multilabel classification (MLC) and label ranking (LR).Multilabel ranking (MLR) is a combination of MLC and LR. The research on multilabel ranking poses important significance for modeling the ambiguity of complicated object. Recently, it has attracted much attention in the maching learning community, motivated from an increasing number of new applications, such as semantic annotation of multimedia objects (images, audio/video), functional genomics, directed marketings, recommendation systems and information retrieval. However, most existing work paid much attention to multi-label ranking under a single reliable annotator setting, usually assumes the availability of a single objective label ranking for each training sample. However, in practice, this assumption does not necessarily hold true. This study deals with a more common case where subjective inconsistent rankings from multiple annotators are associated with each instance.The thesis consists of five chapters. In chapter 1, we introduce the background, the current progress, and the problems encountered in multilabel raning; In chapter 2, formal definition (including the learning framework and evaluation metrics) is given and we also introduce four representative multi-label algorithms with necessary analyses and discussions; In chapter 3,we first introduce the inconsistency problem in multilabel ranking under multiple annotator setting, and then proposing two multilabel ranking method based on LDL:instance-oriented IDL and annotator-oriented ADL, which effectively incorporates the inconsistency information given by multiple annotators; In chapter 4, reports the performance of different multilabel ranking algorithms on natural scene dataset, and draw some interesting conclusions; finally, chapter 5 is a summary of the whole thesis, some existing problem and research directions are discussed.This thesis contributes on the following three aspects:1. proposing the inconsistency problem in multilable ranking under multiple inconsistent labeling sources setting; 2. put forward the modeling methods based LDL which effectively incorporates the inconsistency between multiple annotators; 3. Colletecting a real multilable ranking dataset from multiple labeling sources.