Partial Label Learning Algorithms Based On Norm And Distribution Characteristics

Partial label learning is a kind of weak supervised learning,and its ground-truth label is hidden in the candidate label set.The difficulty of partial label learning is to find the ground-truth label from the candidate label set,which is called disambiguation.Based on the employed strategy,existing approaches can be roughly grouped into two categories,including the average-based strategy and the identification-based strategy.In order to realize disambiguation more effectively,this paper studies partial label learning from the following two aspects:(1)In order to make full use of the potential useful information of sample feature space to help partial label learning disambiguation,this paper proposes a partial label learning algorithm based on l,norm and manifold hypothesis.In this algorithm,manifold hypothesis is applied to partial label learning,and a unified framework is proposed to realize disambiguation and classifier learning at the same time.In addition,l,norm is introduced into partial label learning,which makes the label distribution of samples as sparse as possible,so that the ground-truth label and false positive labels in the candidate label set can be more effectively distinguished.(2)The existing partial label learning does not consider the statistical characteristics of label space.So we propose a partial label learning algorithm based on distribution characteristics.In this algorithm,the standard deviation is introduced into the partial label learning,and the infinite norm is added.The algorithm can make the maximum value in the label distribution is as large as possible,and the minimum value is as small as possible.In the label distribution of the sample,the label corresponding to the maximum value is the ground-truth label,and other labels with smaller values are false positive labels.As a consequence,the ground-truth label and the false positive labels can be more clearly distinguished.The implementation results based on several data sets show that the partial label learning algorithm based on l,norm and manifold hypothesis and the partial label learning algorithm based on distribution characteristics have better performance than other partial label learning algorithms.