Experimental Learning of Causal Models with Latent Variables
Sam Maes, Stijn Meganck, Philippe Leray and Bernard Manderick
In: NIPS 2006 Workshop on Causality and Feature Selection, 8 Dec 2006, Vancouver.
Semi-Markovian causal models (SMCMs) are a recent formalism proposed for performing causal inference in Bayesian networks with latent variables. At this time, they have only been studied from a theoretical point of view.
However, if we want to use these models in practice, some additional questions have to be answered, i.e. how to actually perform causal inference, how to efficiently perform classical probabilistic inference, and how to determine the structure and the parameters of these models from data ?
In this paper, we propose some solutions for all these problems.
First of all, we deal with SMCM structure learning by combining recent work on Ancestral Graph Model learning from observational data with
learning from experiments in order to obtain a completely directed causal graph.
At this time, only the structure of SMCMs is well defined and no efficient parametrisation was provided. We will propose an alternative representation for SMCMs that can easily be parametrised and where the parameters can be learned with classical Bayesian methods.
Finally, we use this parametrisation to develop methods to efficiently perform both probabilistic and causal inference.