Temporal Maximum Margin Markov Network

Abstract

Typical structured learning models consist of a regression component of the explanatory variables (observations) and another regression component that accounts for the neighboring states. Such models, including Conditional Random Fields (CRFs) and Maximum Margin Markov Network (M3N), are essentially Markov random fields with the pairwise spatial dependence. They are effective tools for modeling spatial correlated responses; however, ignoring the temporal correlation often limits their performance to model the more complex scenarios. In this paper, we introduce a novel Temporal Maximum Margin Markov Network (TM3N) model to learn the spatial-temporal correlated hidden states, simultaneously. For learning, we estimate the model’s parameters by leveraging on loopy belief propagation (LBP); for predicting, we forecast hidden states use linear integer programming (LIP); for evaluation, we apply TM3N to the simulated datasets and the real world challenge for occupancy estimation. The results are compared with other state-of-the-art models and demonstrate superior performance.