Inferring causal direction from relational data
David T. Arbour, Katerina Marazopoulou, David D. Jensen · Uncertainty in Artificial Intelligence · 2016
Inferring the direction of causal dependence from observational data is a fundamental problem in many scientific fields. Significant progress has been made in inferring causal direction from data that are independent and identically distributed (i.i.d.), but little is understood about this problem in the more general relational setting with multiple types of interacting entities. This work examines the task of inferring the causal direction of peer dependence in relational data. We show that, in contrast to the i.i.d. setting, the direction of peer dependence can be inferred using simple procedures, regardless of the form of the underlying distribution, and we provide a theoretical characterization on the identifiability of direction. We then examine the conditions under which the presence of confounding can be detected. Finally, we demonstrate the efficacy of the proposed methods with synthetic experiments, and we provide an application on real-world data.