Semi-supervised regression using diffusion on graphs
Mohan Timilsina, Alejandro Figueroa, Mathieu d’Aquin, Haixuan Yang
Abstract
In real-world machine learning applications, unlabeled training data are readily available, but labeled data are expensive and hard to obtain. Therefore, semi-supervised learning algorithms have gathered much attention. Previous studies in this area mainly focused on a semi-supervised classification problem, whereas semi-supervised regression has received less attention. In this paper, we proposed a novel semi-supervised regression algorithm using heat diffusion with a boundary-condition that guarantees a closed-form solution. Experiments from artificial and real datasets from business, biomedical, physical, and social domain show that the boundary-based heat diffusion method can effectively outperform the top state of the art methods.