Abstract

We introduce CGNN, a framework to learn functional causal models asgenerative neural networks. These networks are trained using backpropagation tominimize the maximum mean discrepancy to the observed data. Unlike previousapproaches, CGNN leverages both conditional independences and distributionalasymmetries to seamlessly discover bivariate and multivariate causalstructures, with or without hidden variables. CGNN does not only estimate thecausal structure, but a full and differentiable generative model of the data.Throughout an extensive variety of experiments, we illustrate the competitiveresults of CGNN w.r.t state-of-the-art alternatives in observational causaldiscovery on both simulated and real data, in the tasks of cause-effectinference, v-structure identification, and multivariate causal discovery.

 

Introduction (beta)

None

 

Conclusion (beta)

None