Multi adaptive semi-supervised learning based on Evidence Theory

Abstract

Semi-supervised classification(SSC) trains few labeled samples and a great many of unlabeled samples to obtain a good classification model. Graph semi-supervised learning is one of the main research directions of SSC. It uses unlabeled samples and the few limited labeled samples to build a more reliable graph model and spreads the label information on the graph model. However, the traditional graph semi-supervised learning algorithms are time-consuming and can not well reflect the underlying data distribution of high-dimensional data. To deal with these problems, we propose an Multi adaptive semi-supervised learning based on Evidence Theory (MASSL). MASSL uses some multiple feature subspaces of original data to train different adaptive semi-supervised learning models, and uses Fuzzy Cautious ordered weighted average reasoning method and Evidence theory to combine the different results to obtain a better recognition effect. Experimental results show that the proposed approach can effectively exploit unlabeled data and perform better compared with prevailing related approaches.