Year of Graduation
On the Performance of Linear Predictors in the Semi-Supervised Framework
This paper investigates the performance of linear and semi-supervised estimators compared to standard supervised learning techniques. We present a simple semi-supervised algorithm, which can be easily applied to different practical tasks, and provide an exhaustive theoretical analysis of the method. We discuss optimization issues and present the coordinate-wise minimization algorithm which can be implemented to compute our estimator. Simulations were conducted to compare the efficiency of the algorithm with supervised methods and reveal conditions under which the proposed semi-supervised method performs better. In particular, we found that when the number of labeled observations is small, semi-supervised methods outperform significantly analog supervised learning algorithms.