Paper No. 11-15
PA Thwaites
Using Chain Event Graphs to refine model selection
Abstract: Chain Event Graphs (CEGs) are specifically designed to embody the conditional independence structure of problems whose state spaces are asymmetric and do not admit a natural product structure. The learning of CEGs is closely related to the learning of BNs, and if we use (for example) MAP model selection then where a model can be represented as both a BN and a CEG, the two methods assign this model the same score. If we suspect that a problem incorporates significant context-specific conditional independence structure we can use standard BN-based learning methods to select a good approximate model, and then use the CEG-based learning methods described here to further refine this model.