2 Dec 14:15 Joris Mooij: Additive noise models for causal inference

HIIT seminar, Wednesday Dec 2, 14:15 (coffee from 02), Exactum

Dr. Joris Mooij
Max-Planck-Institut for Biological Cybernetics Dept. Schölkopf Tübingen, Germany

Additive noise models for causal inference

In this talk, I consider a special class of causal models with continuous random variables where noise acts in an additive way. Each variable is assumed to be a function of its parent variables, plus noise, where all the noise terms are jointly independent. For the special case of only two variables, this model class is asymmetric in the sense that an additive noise model X->Y cannot be written as an additive noise model Y->X in the generic case.

This asymmetry is exploited for causal inference by postulating that if the observational data is well described by an additive noise model, the DAG structure of the additive noise model likely coincides with the causal structure. I discuss how additive noise models can be learnt from data using regression by dependence minimization. I present preliminary but promising empirical results on real-world data sets consisting of several pairs of variables, showing that the additive noise assumption can indeed be useful for causal inference. I propose an efficient algorithm for learning additive noise models from data in the multivariate case. I also present some first results on the case where two observed variables have a hidden common cause.

Last updated on 4 Jan 2010 by WWW administrator - Page created on 2 Dec 2009 by Visa Noronen