Qualitative chain graphs and their application

• Definition and properties of qualitative chain graphs (QCGs).
• Key differences with qualitative probabilistic networks (QPNs).
• Modelling and reasoning of qualitative representations of feedback systems.
• Approximate reasoning of qualitative information with 2nd order distributions.
• Application to common diseases of real patient data in general practice.

For many problem domains, such as medicine, chain graphs are more attractive than Bayesian networks as they support representing interactions between variables that have no natural direction. In particular, interactions between variables that result from certain feedback mechanisms can be represented by chain graphs. Using qualitative abstractions of probabilistic interactions is also of interest, as these allow focusing on patterns in the interactions rather than on the numerical detail. Such patterns are often known by experts and sufficient for making decisions. So far, qualitative abstractions of probabilistic interactions have only been developed for Bayesian networks in the form of qualitative probabilistic networks. In this paper, such qualitative abstractions are developed for chain graphs with the practical aim of using qualitative knowledge as constraints on the hyperspace of probability distributions. The usefulness of qualitative chain graphs is explored for modelling and reasoning about the interactions between diseases.