Abstract
Maximum entropy discrimination (MED) is already shown to be effective for discriminative classification and regression, and can be applied to multitask learning (MTL) with some further assumptions. Self-training is a commonly used technique for semi-supervised learning. In order to integrate the merits offered by semi-supervised learning and MTL, this paper presents semi-supervised MTL via self-training and MED. We select the suitable measure metric and identify how to use unlabeled data. Experimental results on two UCI data sets demonstrate that our method yields better performance than semi-supervised single-task learning (STL) and supervised MTL.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have