Semi-supervised network regression with Gaussian process

Abstract

In recent years, there has been a rapid growth in interest of using network-based machine learning. They offer the capacity to handle data that exist on irregular and complex structures with interactions between data points. In this paper, we present a semi-supervised regression model utilizing network-based Gaussian process. The proposed method constructs a Gaussian process prior using information from a given network. However, it incurs high computational costs from the required inversions to produce the predictive output and model selection. To overcome the difficulty, we further propose an approximated version that avoids matrix inversion. The proposed method was applied to several regression problems to validate the empirical performance and effectiveness in situations with limited amount of labeled data.