Research On Partial Label Loss Function

Partial label learning is an important weakly-supervised learning framework,in which labelling information is no more unique and unambiguous:each instance is associated with a set of candidate labels,among which only one is ground-truth label.As the ground-truth label of each training example is not directly accessible,the common supervised learning approaches cannot be used directly for solving the partial label learning problem.Generally,the design of loss function reflects how the learning algorithm characterizes properties of the investigated learning problems.Existing loss functions used by most partial label learning algorithms have two major drawbacks.Firstly,existing loss functions for partial label learning just focus on the mapping relation between instance and label,while the relation among instances is ignored.Secondly,existing loss functions for partial label learning assume equal importance weight for each label in candidate label set,which ignores the fact that the importance of the modeling outputs from ground-truth label and false positive label should be different.In view of the above two drawbacks,this paper studies the design of partial label loss functions from the following two aspects:Generally,machine learning algorithm takes the consistency assumption that instances with high similarity in the feature space will also share high similarity in the label space.Based on this,we propose a novel partial label learning algorithm COPAL by employing the consistency assumption.To disambiguate the candidate label set,the loss function of COPAL not only considers the modeling outputs on candidate labels,but also the similarity of neighboring instances' modeling outputs.Experimental results show that,by introducing consistency assumption into the partial label loss function for instance similarity exploitation,disambiguation over the training examples can be better achieved.In partial label learning,the modelling output of ground-truth label is likely to be overwhelmed by the false positive label.Thus,the confidence of different labels' modelling output should be considered during designing partial label loss function.Based on this,proposed novel confidence-rated partial label learning algorithm named CORD is proposed.To disambiguate the candidate label set,the loss function of CORD considers the modeling output as well as the labeling confidence of each candidate label,where the identification of ground-truth label and the confidence update of candidate label are performed via iterative optimization.Experimental results show that,by introducing labeling confidence into the partial label loss function,disambiguation over the training examples can be better achieved.This thesis contains five chapters.In Chapter 1,definition,state-of-the-art,and open research issues of partial label learning are introduced.In Chapter2,existing partial label learning algorithms are briefly reviewed.In Chapter 3,the partial label loss function based on consistency assumption is introduced.In Chapter 4,the partial label loss function based on labeling confidence is introduced.In the end,we conclude this thesis in Chapter 5.