PP-PLL:probability Propagation For Partial Label Learning

In recent years,partial label learning,as an important weak supervised learning framework,has been widely concerned in many fields,such as target detection and clinical medicine.In the framework of partial label learning,each example is associated with a set of candidate labels,among which only one is correct.Because training data with explicit label information is hard to be obtained directly,the classical supervised method cannot be directly used to solve partial label learning problem.Generally,In order to design an effective partial label learning algorithm,an intuitive idea is to disambiguate the labels in the candidate labels set.However,the existing classical partial label disambiguation methods are easily misled by false positive labels in each candidate label set.Although the corrected partial label learning reduces the negative effects of false labels by introducing the correlation of input space,there are still two main problems.Firstly,there is still some false information among the samples in the neighbor weight graph.In the process of information propagation,the model will expand the influence of false information.Secondly,the improved partial label disambiguation method is often too sensitive to abnormal data,resulting in very unstable model.In terms of the two aspects mentioned above,the main research work of this thesis is as follows:1.Due to the correlation between input spaces finally influencing the sharing information among different candidate label set of each sample,the negative information caused by the correlation of input spaces cannot be ignored easily.Based on this problem,candidate label information is added to filter out the neighbor samples with false information in the process of propagation.Thos way can improve the correctness of relevance,which improves the efficiency and confidence level of information spread.2.In order to further strengthen the mapping relationship from feature space to label space and effectively strengthen the exclusiveness among labels,this thesis proposed a biconvex objective function based on the manifold hypothesis,which utilize the maximum entropy model,and constructs a linear mapping relationship between the input feature and the ground-truth label.This thesis verifies the effectiveness and superiority of the algorithm in the two tasks including self correction and prediction of unseen samples by experiments on four controled UCI partial label datasets and five open and real-world partial label datasets.