|
Abstract:
|
we develop a learning algorithm for directed acyclic graphical (DAG) or Bayesian network models using a combination of observational and (randomized) experimental data. Prior work has mainly focused on algorithms involving first using observational data to learn the Markov equivalence class (MEC) for the DAG and then using do-calculus rules based on interventions to learn the additional directions. However when the number of nodes $p$ is large, errors are often made in learning the MEC. In addition existing approaches that rely on accurate recovery of the MEC do not scale well to large graphs although greedy search methods are applied. Our learning strategies we introduce address this challenge by using a combination of moralized graphs and do-calculus rules based on intervention graphs. Since learning the moralized graph is substantially more reliable and faster than learning the MEC as the number of nodes p becomes large, we show that our algorithm makes significantly fewer errors in terms of recovering the large-scale DAG model compared to state-of-the-art algorithms that rely on accurate recovery of the MEC.
|
