University of Pittsburgh

Causal Discovery with Arbitrary Mixtures of Observational and Experimental Data

graduate student
Friday, November 3, 2017 - 12:30pm - 1:00pm

We propose a general framework which enables Causal Discovery algorithms for learning causal networks to deal with multiple datasets, including arbitrary mixtures of observational and experimental data.  We represent the context from which the data originated as an additional variable.  This allows data across many datasets to be pooled into a singular dataset.  Experiments are incorporated as variables as well which allows the modeling of hard, soft, fat-handed, precise-handed, and unknown-effect interventions.  We validate the proposed method on both simulated and real-world examples and conclude our framework for handling multiple datasets is flexible, computationally efficient, and increases statistical power.  The method presented in this work will be available for use as part of Tetrad's suite of algorithms shortly.

Copyright 2009–2018 | Send feedback about this site