Kun Zhang
(Max Planck Institute)
Wednesday 26th September 2012
Time: 4pm
B10 Basement Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Functional causal models: Beyond linear instantaneous relations
Functional causal models have served as one of the standard approaches for causal discovery from non-experimental data. I start with relating functional causal models to the invariance of the conditional distribution, and then discuss the conditions, benefits, and cautions of using this approach. After reviewing a fundamental functional causal model, namely, the linear, nongaussian, and acyclic model (LiNGAM), I focus on two more practical models: Granger causality with instantaneous effects and the post-nonlinear (PNL) causal model. The former is an extension of the traditional Granger causality to incorporate the linear instantaneous causal relations, or equivalently, it is a linear functional causal model with certain temporal constraints, and it can be efficiently estimated from data. The latter takes into account the nonlinear effect of the cause, the inner noise effect, and possible measurement distortion in the effect; I show its identifiability, its application in distinguishing cause from effect, and how to extend it to the case of many variables. Finally, I make an attempt to compare the functional causal model based approaches and the constraint-based ones for causal discovery.
http://www.kyb.mpg.de/nc/employee/details/kzhang.html