Data driven semi-supervised learning

Neural Information Processing Systems 

We obtain low regret and efficient algorithms in the online setting, and generalization guarantees in the distributional setting.