A maximum entropy approach to learn Bayesian networks from incomplete data

This paper addresses the problem of estimating the parameters of a Bayesian network from incomplete data. This is a hard problem, which for computational reasons cannot be effectively tackled by a full Bayesian approach. The workaround is to search for the estimate with maximum posterior probability. This is usually done by selecting the highest posterior probability estimate among those found by multiple runs of Expectation-Maximization with distinct starting points. However, many local maxima characterize the posterior probability function, and several of them have similar high probability. We argue that high probability is necessary but not sufficient in order to obtain good estimates. We present an approach based on maximum entropy to address this problem and describe a simple and effective way to implement it. Experiments show that our approach produces significantly better estimates than the most commonly used method.