Research On Augmentation Methods For Partial Label Learning

Partial label learning is an emerging weakly supervised learning framework.In partial label learning,each training example is associated with multiple candidate labels,among which only one label corresponds to its ground-truth label.In many real-world scenarios,the labeled examples may be limited,or the feature representation of partial label examples may be less informative.To solve the problems,this paper investigates augmentation methods for partial label learning from two perspectives:On one hand,for the problem of partial label learning with unlabeled data,a novel semisupervised partial label learning approach named PARM is proposed.Firstly,label propagation is adopted to disambiguate partial label examples.After that,we esimate the labeling confidence of training examples and update model parameters based on the maximum margin formulation and alternating optimization technique.Experimental results on synthetic and real-world data sets indicate that PARM achieves superior performance against state-of-the-art semi-supervised partial label learning approaches.On the other hand,for the problem of weak discrimination of feature representation for partial label training examples,a novel feature augmentation method named PLDA is proposed.Firstly,we construct an optimization problem for learning labeling confidence and class prototype based on both global consistency in the feature space and local consistency in the label space,and thus derive the alternating optimization solution.After that,the original feature space is enriched with confidence-rated class prototype features.Experimental results on synthetic and real-world data sets show that PLDA can improve the generalization performance of the model significantly.This paper consists of four chapters.The first chapter introduces basic concept and research background of partial label learning.The second chapter introduces the PARM approach for semi-supervised partial label learning.The third chapter introduces the PLDA approach for partial label feature augmentation.The fourth chapter concludes the paper.