Causal learning methods are often evaluated in terms of their ability to discover a true underlying directed acyclic graph (DAG) structure. However, in general the true structure is unknown and may not be a DAG structure. We therefore consider evaluating causal learning methods in terms of predicting the effects of interventions on unseen test data. Given this task, we show that there exist a variety of approaches to modeling causality, generalizing DAG-based methods. Our experiments on synthetic and biological data indicate that some non-DAG models perform as well or better than DAG-based methods at causal prediction tasks.
David Duvenaud, Daniel Eaton, Kevin P. Murphy, Mar