practical successes of semi-supervised learning?
What do people view as being the most compelling practical successes of semi-supervised learning? Of the papers in
Chapelle, O., Scholkopf, B., and Zien, A. Semi-Supervised Learning. MIT Press, 2006.
the protein ones (Weston/Leslie/Ie/Noble and Shin/Tsuda) look most compelling to me, but I have no idea if bioinformatics practitioners actually find unlabeled data useful. There's lots of proposed applications in language processing, but it's unclear to me how much the unlabeled data is really contributing, and whether one might better have just used an active learning approach. Thoughts?
Update (11-Sep-10): There's actually three versions of this question:
Semi-supervised learning is used during active selection of labels, but ignored afterwards?
Semi-supervised learning is used both during active selection of labels, and afterwards in fitting final model/making final predictions?
Semi-supervised learning is used in combination with randomly selected labels?
In all cases I'm interested in whether there's industrial applications where semi-supervised learning is actually used and provides real benefits. (For active learning this is unambiguously true.)
