UCLA

We propose a penalized log-likelihood method for estimating causal Bayesian networks using a mix of intervention data and observation data, based on Aragam and Zhou's previous algorithm [AZ15]. A causal Bayesian network is represented as a directed acyclic graph (DAG). With intervention data, we can distinguish the true graph from DAG equivalence classes and learn about special structures in the graph. Tests on simulated data show that this method is consistent and is superior to other methods in multiple aspects such as true positive rate and timing. A two-stage approach to estimate big networks is proposed as an application of our method.