Domain Adaptation by Using Causal Inference to Predict Invariant Conditional Distributions

  • 2018-09-17 15:37:54
  • Sara Magliacane, Thijs van Ommen, Tom Claassen, Stephan Bongers, Philip Versteeg, Joris M. Mooij
  • 0

Abstract

An important goal common to domain adaptation and causal reasoning is to makeaccurate predictions when the distributions for the target domain(s) and thesource domain(s) differ. We consider the case in which the domains correspondto different contexts in which a system has been measured, for example, apurely observational context and several interventional contexts in which thesystem has been perturbed by external interventions. We consider a class ofsuch causal domain adaptation problems, where data for multiple source domainsare given, and the task is to predict the distribution of a certain targetvariable from measurements of other variables in one or more target domains. Wepropose an approach for solving these problems that exploits causal inferenceand does not rely on prior knowledge of the causal graph, nor of the type ofthe interventions or the intervention targets. We propose a practicalimplementation of the approach and evaluate it on simulated and real worlddata.

 

Quick Read (beta)

loading the full paper ...