Self-paced Safe Co-training for Regression

In semi-supervised learning, co-training is successfully in augmenting the training data with predicted pseudo-labels. With two independently trained regressors, a co-trainer iteratively exchanges their selected instances coupled with pseudo-labels. However, some low-quality pseudo-labels may significantly decrease the prediction accuracy. In this paper, we propose a self-paced safe co-training for regression (SPOR) algorithm to enrich the training data with unlabeled instances and their pseudo-labels. First, a safe mechanism is designed to enhance the quality of pseudo-labels without side effects. Second, a self-paced learning technique is designed to select "easy" instances in the current situation. Third, a "qualifier-based" treatment is designed to remove "weak" instances selected in previous rounds. Experiments were undertaken on nine benchmark datasets. The results show that SPOR is superior to both popular co-training regression methods and state-of-the-art semi-supervised regressors.