A Maximum Entropy Approach to Semi-supervised Learning

Various supervised inference methods can be analyzed as convex duals of a generalized maximum entropy framework, where the goal is to find a distribution with maximum entropy subject to the moment matching constraints on the data. We extend this framework to semi-supervised learning using two approaches: 1) by incorporating unlabeled data into the data constraints and 2) by imposing similarity constraints based on the geometry of the data. The proposed approach leads to a family of discriminative semi-supervised algorithms, that are convex, scalable, inherently multiclass, easy to implement, and that can be kernelized naturally. Experimental evaluation of special cases shows the competitiveness of our methodology.