Cooperative Semi-supervised Regression Algorithm based on Belief Functions Theory

Abstract

Semi-supervised learning (SSL), which can exploit both labeled and unlabeled samples, has attracted a lot of research attention. Semi-supervised regression is an important content in semi-supervised learning. The traditional semi-supervised regression methods may encounter uncertainty problems in the learning process. In this paper, a cooperative semi-supervised regression method based on belief functions theory is proposed. The proposed method uses belief functions to address the uncertainty in the semi-supervised regression. The algorithm uses two belief functions based regressors and labels the unlabeled samples based on the combined results of the two regressors. The labeling confidence of an unlabeled sample is estimated through the reduction in mean squared error over the labeled neighborhood of the given sample. Experimental results show that the proposed method can effectively exploit unlabeled samples to obtain better regression performance.