Abstract
In this work, we formalize the problem of causal inference over graph-basedrelational time-series data where each node in the graph has one or moretime-series associated to it. We propose causal inference models for thisproblem that leverage both the graph topology and time-series to accuratelyestimate local causal effects of nodes. Furthermore, the relational time-seriescausal inference models are able to estimate local effects for individual nodesby exploiting local node-centric temporal dependencies andtopological/structural dependencies. We show that simpler causal models that donot consider the graph topology are recovered as special cases of the proposedrelational time-series causal inference model. We describe the conditions underwhich the resulting estimate can be used to estimate a causal effect, anddescribe how the Durbin-Wu-Hausman test of specification can be used to testfor the consistency of the proposed estimator from data. Empirically, wedemonstrate the effectiveness of the causal inference models on both syntheticdata with known ground-truth and a large-scale observational relationaltime-series data set collected from Wikipedia.