Efficient Prediction for Tree Markov Random Fields in a Streaming Model
Mark Herbster, Stephen Pasteris and Fabio Vitale
In: NIPS Workshop on Discrete Optimization in Machine Learning (DISCML) 2011: Uncertainty, Generalization and Feedback, 17 December 2011, Sierra Nevada, Spain.
We consider streaming prediction model for tree Markov Random fields. Given the random field, at any point in time we may perform one of three actions: i) predict a label at a vertex on the tree ii) update by associating a label with a ver- tex or iii) delete the label at a vertex. Using the standard methodology of belief propagation each such action requires time linear in the size of the tree. We give a method based on an optimal decomposition tree that even in the worst case is an exponential speed-up over belief propagation.