Abstract

Chain Event Graphs encode a new class of finite discrete models that strictly contains discrete Bayesian Network models and their context specific generalizations as a very special case. They provide a particularly powerful graphical framework for eliciting, querying, encoding, performing inferences and estimating highly asymmetric models in an efficient and transparent way. Such model classes arise naturally in both in the social sciences and biology. The class exhibits many of the advantages of Bayesian Networks. There are direct analogues of graphical conditional independence querying techniques. The framework supports conjugate inference with complete data and hence efficient exact search algorithms over the model class. Furthermore, like the Bayesian Network, the class encodes algebraic constraints on a class of polynomials and so it can be mapped into its own associated an albeit typically inhomogeneous algebraic parametrization. Finally, being constructed from an event tree, Chain Event Graphs admit an excellent framework for expressing causal extensions of this model class. The talk will demonstrate these properties using a number of examples.