Fitting feature-dependent Markov chains

We describe a method for fitting a Markov chain, with a state transition matrix that depends on a feature vector, to data that can include missing values. Our model consists of separate logistic regressions for each row of the transition matrix. We fit the parameters in the model by maximizing the log-likelihood of the data minus a regularizer. When there are missing values, the log-likelihood becomes intractable, and we resort to the expectation-maximization (EM) heuristic. We illustrate the method on several examples, and describe our efficient Python open-source implementation.