Following a discussion on the general form of regularization for semi-supervised learning, we propose a semi-supervised regression algorithm. It is based on the assumption that we have certain order preferences on unlabeled ...