A novel Tri-training algorithm based on Evidence Theory

Hongshun He, Zhipeng Zhang, Lei Yang, Yang Li, Jiakang Shi

Abstract: In many machine learning tasks, it is usually difficult to obtain enough labeled samples. Semi-supervised classification(SSC) trains a few labeled samples and a large amount of unlabeled samples to obtain a better learning model, which has attracted a lot of research attention. Traditional semi-supervised methods may encounter uncertainty problems and information loss when dealing with those samples having ambiguous class belongingness. In this paper, the uncertainties encountered in semi-supervised learning are addressed using the Evidence Theory, and semi-supervised learning methods based on belief functions are proposed. The proposed methods label the unlabeled data with belief modeling. They can effectively use limited supervised information to facilitate the classification process. Experimental results based on benchmark data sets show that the proposed approaches can effectively exploit unlabeled data and perform better compared with prevailing approaches.