Causal Discovery in Non-Ideal Frameworks

A Causal Bayesian network (CBN) is a well-known technique to represent causal
relations between variables. To learn the causal relations, experiments are needed in
general because from observational data alone we can only learn up to Markov
equivalence. The elicitation of this causal information from data provides useful insight in
the system under study. Most existing algorithms for causal discovery implicitly assume a
setting where all variables of the system under study are known and sufficient amount of
data is present to perform reliable statistical tests. These assumptions provide an ideal
framework, which facilitates the discovery task significantly. In this article we use existing
techniques from this ideal setting as a starting point for new algorithms in non-ideal
settings. We present two algorithms, UnCaDo and MyCaDo++, which can be used when
insufficient data is present to perform statistical tests reliably and when there are latent
variables respectively. Both techniques provide experimentation plans to recover the
causal structure.