Research On Feature Representation And Fusion For Partial Label Learning

Under the standard partial label learning framework,each object is characterized by a single feature vector and associated with multiple candidate labels,of which there is only one unknown ground-truth label.On the other hand,in real-world problems,the nature of objects is often more complicated,where each object might have multi-source feature representations and unknown ground-truth labels are not unique.Generally speaking,effective feature representation can significantly improve the generalization ability of learning system,while partial label feature representation task is more challenging as the ground-truth labeling information is unknown.This paper focuses on the feature representation and fusion for partial label learning,which consists of the following two main studies:Firstly,for single-view partial label training data,this paper studies the problem of partial label dimensionality reduction and proposes a new method named Delin.This method improves the generalization ability of partial label classification systems by endowing the popular linear discriminant analysis(LDA)techniques with the ability of dealing with partial label training examples.Specifically,Delin alternates between LDA dimensionality reduction and candidate label disambiguation based on estimated labeling confidences over candidate labels.On one hand,the projection matrix of LDA is optimized by utilizing disambiguation-guided labeling confidences.On the other hand,the labeling confidences are disambiguated by resorting to k NN aggregation in the LDA-induced feature space.Extensive experiments on synthetic as well as real-world partial label data sets clearly validate the effectiveness of Delin in improving the generalization ability of state-of-the-art partial label learning algorithms.Secondly,for partial label training data with multiple views and complicated noise,a new partial label learning algorithm named Fiman is proposed based on multi-view feature fusion.Firstly,an aggregate manifold structure over training examples is generated by adaptively fusing affinity information conveyed by feature vectors of different views.Then,candidate labels of each training example are disambiguated by preserving the feature-induced manifold structure in label space.Finally,the resulting predictive models are learned by fitting modeling outputs with the disambiguated labels.Extensive experiments show that Fiman achieves highly competitive performance against state-of-the-art approaches in solving the corresponding problem.This paper consists of five chapters.The first chapter introduces the background,research status and pending research problems of partial label learning.The second chapter gives the formal definition of single-view / multi-view partial label learning and the existing algorithms,and briefly introduces classic dimensionality reduction methods.The third chapter introduces the partial label dimensionality reduction algorithm Delin.Chapter 4 introduces the multiview partial multi-label feature fusion method Fiman.Chapter 5 summarizes the content of the thesis.