Semi-Supervised Multi-Task Regression

Labeled data are needed for many machine learning applications but the amount available in some applications is scarce. Semi-supervised learning and multi-task learning are two of the approaches that have been proposed to alleviate this problem. In this paper, we seek to integrate these two approaches for regression applications. We first propose a new supervised multi-task regression method called SMTR, which is based on Gaussian processes (GP) with the assumption that the kernel parameters for all tasks share a common prior. We then incorporate unlabeled data into SMTR by changing the kernel function of the GP prior to a data-dependent kernel function, resulting in a semi-supervised extension of SMTR, called SSMTR. Moreover, we incorporate pairwise information into SSMTR to further boost the learning performance for applications in which such information is available. Experiments conducted on two commonly used data sets for multi-task regression demonstrate the effectiveness of our methods.