Bayesian structure learning using dynamic programming and MCMC

1
Bayesian structure learning using dynamic
programming and MCMC
Daniel Eaton and Kevin Murphy
Department of Computer Science, University of
British Columbia

• Want to learn posterior over DAGs, or some
  marginal summary thereof
• Want causal interpretation
• Current computational approaches have caveats
  (Order MCMC and DP)
• Restrictive prior
• Lack of posterior (DAG) samples to make
  objective comparisons with other methods
• We do MCMC in DAG space, but with a powerful
  proposal based on the exact edge marginals gotten
  from the DP algorithm
• Solves above problems and performs better

We get a better solution sooner
2
Bayesian structure learning using dynamic
programming and MCMC
Daniel Eaton and Kevin Murphy
Department of Computer Science, University of
British Columbia

• We apply our method to a protein-signaling
  network dataset
• Inferred model shows agreement with commonly
  accepted ground truth
• Interventional data handled by standard trick of
  Cooper and Yoo

Ground truth cartoon
Learned model

• We explore a novel model for interventions
• Uncertainty in effect and target(s) of an
  intervention
• Handles fat hand interventions where the mad
  scientist is not sure of what his or her
  chemicals are affecting (or there are hidden
  variables!)

Learned model with interventions