Metric-Based Semi-Supervised Regression

Abstract

Regression problems are present in many industrial applications, and many supervised learning algorithms have been devised over decades. However, available labeled examples are limited in some application settings; meanwhile, enormous unlabeled examples are relatively easy to collect. Thus, this work proposes a simple but effective method to cope with semi-supervised regression problems. We propose to use deep neural networks to develop our proposed method as deep learning has shown promising results in recent years. Our proposed method is a metric-based approach, and the goal is to learn an embedding space by metric learning with few labeled examples and enormous unlabeled examples. The regression estimation of the target data point is performed on the new space. We generate an artificial dataset based on several criteria to investigate whether the proposed model could make accurate predictions on the data samples that have specific properties. The experimental results point that our proposed model could capture the trend of a non-linear function and normally predict well even though this dataset comprises extreme outliers. Moreover, we conduct experiments on four datasets and compare our proposed work with several alternatives. The experimental results indicate that our proposed method achieves promising results. Besides performance evaluation, detailed analysis about our proposed method is also provided in this work.