Semi-Supervised Regression with Co-Training

In many practical machine learning and data min-ing applications, unlabeled training examples arereadily available but labeled ones are fairly expen-sive to obtain. Therefore, semi-supervised learn-ing algorithms such asco-traininghave attractedmuch attention. Previous research mainly focuseson semi-supervised classification. In this paper, aco-training style semi-supervised regression algo-rithm, i.e.COREG, is proposed. This algorithmuses twok-nearest neighbor regressors with differ-ent distance metrics, each of which labels the unla-beled data for the other regressor where the label-ing confidence is estimated through consulting theinfluence of the labeling of unlabeled examples onthe labeled ones. Experiments show thatCOREGcan effectively exploit unlabeled data to improveregression estimates.