Entropic Graph-based Posterior Regularization for Learning Probabilistic Models

The advent of high-throughput DNA sequencing methods has led to an explosion in the availability of genome-wide data and a corresponding opportunity to use computational methods to derive insights into cellular function. A predominant form of statistical model used for genomic data has been temporal, such as the hidden Markov model (HMM) (Rabiner, 1989) or dynamic Bayesian network (DBN) (Dean and Kanazawa, 1988). Temporal models consist of a chain of random variables exhibiting the temporal Markov property, which asserts that “future” and “past” variables are independent given some notion of the present. This property enables exact inference to be performed using temporal dynamic programming. Most often, inference takes the form of a forward and then a backward pass along the chain. In genomics, the “temporal” axis is generally position along the genome rather than actual time.