Consistent and Tractable Algorithm for Markov Network Learning

Vojtech Franc, Daniel Prusa, Andrii Yermakov

Abstract: Markov network (MN) structured output classifiers provide a transparent and powerful way to model dependencies between output labels. The MN classifiers can be learned using the M3N algorithm, which, however, is not statistically consistent and requires expensive fully annotated examples. We propose an algorithm to learn MN classifiers that is based on Fisher-consistent adversarial loss minimization. Learning is transformed into a tractable convex optimization that is amenable to standard gradient methods. We also extend the algorithm to learn from examples with missing labels. We show that the extended algorithm remains convex, tractable, and statistically consistent.