Learning causal network structure from multiple (in)dependence models

We tackle the problem of how to use information from multiple (in)dependence models, representing results from different experiments, including background knowledge, in causal discovery. We introduce the framework of a causal system in an external context to derive a connection between strict conditional independencies and causal relations between variables. Constraint-based causal discovery is shown to be decomposable into a candidate pair identification and a subsequent elimination step that can be applied separately from different models. The result is the first principled, provably sound method that is able to infer valid causal relations from different experiments in the large sample limit. We present a possible implementation that shows what results can be achieved and how it might be extended to other application areas.