| Label propagation is a graph-based semi-supervised learning framework.Because of its good performance and easy implementation,it has attracted extensive attention.However,the low-quality distribution of labels will cause a large performance degradation of the algorithm under the same label rate,compare to the balance label distribution.This dissertation focuses on how to alleviate the impact of low-quality label distribution on label propagation algorithms,and studies label propagation algorithms that address label imbalance and missing labels.The main research contents of this dissertation are as follows:(1)To get rid of the deficiency about imbalanced labels,this dissertation proposes an incremental label propagation framework.It comprises two key strategies: an incremental propagation strategy and an integration method of multiple propagation results.The incremental balance strategy can gradually add trusted pseudo labels to balance the label distribution so as to reduce the uncertainty caused by pseudo labels.Then,the integration method can evaluate and weighted integrate them to learn a robust result.In the experimental analysis,the framework is combined and compared with five label propagation algorithms on ten benchmark data sets.The experimental results show that the framework can effectively alleviate the performance degradation caused by the imbalanced distribution of labels.(2)To get rid of the label missing problem in some classes,this dissertation proposes a label propagation based on bipartite graph algorithm.This algorithm tries to learn a bipartite graph as exemplar constraints that reflect the relations between objects and exemplars representing all the classes to guide the learning process,instead of label constraints in original label propagation.The algorithm provides a method to produce highquality exemplars from two channels representing the known classes(where some objects have prior labels)and the missing classes(where all the objects have no prior labels),respectively.Based on the generated exemplars,the exemplar constraints can be learned by using relations in the known classes to evaluate that in the missing classes.This dissertation compares five methods on six data sets with different missing conditions and verifies that the proposed method can effectively solve the problem of performance degradation caused by label missing distribution.(3)This dissertation designs and develops a label propagation system for low-quality label distribution.The system includes functions such as algorithmic introduction,the effect of the low-quality label distribution and the exhibition of proposed algorithms et al.It can help researchers better understand the impact of low-quality label distribution and the operation principle of the algorithm.The system can realize the result comparison of multiple algorithms on different data sets,and has good operability.This dissertation investigates methods for improving label propagation algorithms by addressing the challenges posed by low-quality label distributions.Specifically,it proposes solutions for handling imbalanced and missing label distributions,which serve as effective technical support for advancing label propagation research. |
