Research On Multi-label Classification Under Labeling Noise And Feature Construction

In multi-label classification,each instance(represented by feature vector)is associated with multiple class labels.A classification model mapping from instance to a set of class labels can be induced based on multi-label training examples,which is suitable for modeling objects with rich semantics.In conventional multi-label classification tasks,it is usually assumed that the label set associated with each instance is accurate.Nonetheless,under many complicated scenarios(such as image annotations by web users),the label set associated with each instance would be noisy which necessitate the need on investigating multi-label classification with labeling noise.In this thesis,the problems of multi-label classification under labeling noise and feature construction are studied with the following main contributions:For multi-label classification under labeling noise,this thesis proposes a two-stage classification method Particle based on credible label elicitation,which elicits labels with high labeling confidence from the candidate label set for subsequent model construction to mitigate the negative impact of false positive labels.In the first stage,the labeling confidence of candidate label for each noisy training example is estimated via iterative label propagation;in the second stage,by utilizing credible labels with high labeling confidence,multi-label predictor is induced via pairwise label ranking with virtual label splitting or maximum a posteriori(MAP)reasoning.For multi-label classification under feature construction,this thesis extending the existing label-specific features strategy from uni-label mode to bi-label mode and proposes a simple yet effective Bi Label-specific features construction method BiLAS.Specifically,a group of tailored features are generated for a pair of class labels with heuristic prototype selection and embedding.Thereafter,predictions of classifiers induced by Bi Label-specific features are ensembled to determine the relevancy of each class label for unseen instance.This thesis consists of four chapters.The first chapter introduces the research background and pending issues on multi label classification;the second chapter introduces multi-label classification approach Particle which works under labeling noise;the third chapter introduces the multi-label classification approach BiLAS based on feature construction.The fourth chapter summarizes the whole thesis.