Research On Partial Multi-Label Learning

In multi-label learning,each example is associated with multiple class labels.The goal is to construct a mapping from the input space to the output space based on the multi-label training data,which is used to predict a label set for the unknown example.In the traditional multi-label learning,it is usually assumed that all the class label associated with example are valid.However,it is difficult to accurately assign labels to each example in the real world.Under many scenarios,the labeling quality of training examples would be less satisfactory with incorrect labels,missing labels or redundant labels,which lead to the problem of multi-label learning with noisy labeling information.Multi-label learning with noisy labeling information can be viewed as a weakly supervised learning problem.Specifically,this paper aims to study the Partial Multi-label Learning(PML)problem,where the training examples are annotated with redundant labeling information.In PML,each training example is associated with a set of candidate labels,and the goal is to learn a mapping from the input space to the output space based on the given weakly supervised information.In this paper,the problem of PML is studied in the following aspects:One maximum margin PML algorithm is proposed,which integrates the noisy labeling information into the traditional SVM framework,and then converts the PML problem into a convex optimization problem that minimizes empirical loss,hinged loss and confidence constraints.In order to maximize the margin between the weighted output of models in the candidate label set with the maximum output of models in the non-candidate label set,a classification model was learned for each label by alternate iterative optimization strategy.Furthermore,a two-stage classification method with label-specific features is proposed.In the first stage,the internal structural information of each label is used to disambiguate the training examples.In the second stage,the binary classification model is constructed with labelspecific features.Thereafter,the label set with unseen example is predicted by the classification model.This paper is divided into four chapters.The first chapter briefly introduces the background,related work and problems to be solved.In the second chapter,we introduce the proposed large margin PML algorithm.In the third chapter,we introduce the proposed PML algorithm with label-specific features.The fourth chapter summarizes the whole thesis.