Research On Label Propagation Of Semi-superyised Based On Clustering

Semi-supervised learning is an important branch of machine learning, with theexplosion of internet in recent years, there is a great amount of data analysis demand.Semi-supervised learning can utilize limited labeled data and a large number ofunlabeled data to solve data classification problem. Therefore researchers of machinelearning pay intense attention to this field of research. As a popular method ofsemi-supervised learning, graph-based semi-supervised learning has been a new hotspot of research.In this paper, we concentrate on graph-based semi-supervised learning, analyzingand optimizing label propagation algorithm. In order to solve the accuracy declineproblem caused by the labeled data close to classification bound, we proposedclustering based graph construction method. We add the weight of edges on both sideof which the data point belong to the same cluster center, as a result, the denselydistributed data points could be less probably to be separated. Furthermore, weintroduce the self-training framework to solve the problem of slow propagation speedand classification ambiguity of data far from labeled data. In the process of labelpropagation, we mark the data with high belief as labeled data and restart the labelpropagation process. The classification accuracy and speed is improved by thismethod.We make abundant experience on the UCI dataset, prove that clustering basedlabel propagation algorithm obtain slightly better result than typical label propagationalgorithm on common data, when the labeled data is close to classification boundary,the proposed method performs obviously better than typical method. Then we testedself-training based label propagation algorithm on the same dataset, and the proposedmethod costs less time and gain higher accuracy comparing with classic labelpropagation.