Iterative Labeling for Semi-Supervised Learning
Steve Hanneke, Dan Roth · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2004
We propose a unified perspective of a large family of semi-supervised learning algorithms, which select and label unlabeled data in an iterative process. We discuss existing approaches that label examples based on the confidence of the current hypothesis, and propose an alternative approach that labels examples based on empirical risk. This new approach is shown to be statistically reasonable, allows for worst-case performance guarantees and, as we show, significantly outperforms confidence-based approaches in experiments. 1.