Causality and Prediction
Speaker: Jonas Peters, Associate Professor, Dept. of Mathematical Sciences, University of Copenhagen.
Causal models are known to be stable with respect to distributional shifts such as interventions. In this talk, we discuss two ideas. (1) The method of Invariant Causal Prediction turns this principle around and learns causal structure by searching for invariant models. (2) We introduce CausalKinetiX, a framework that trades off invariance and predictability in dynamical systems. The two methods are applied to gene data (predicting the effect of gene deletions in yeast) and data from metabolic networks (learning ODE models from heterogeneous experiments), respectively. No prior knowledge about causality is required.
(1a) J. Peters, P. Bühlmann, N. Meinshausen: Causal inference using invariant prediction: identification and confidence intervals, Journal of the Royal Statistical Society, Series B (with discussion) 78(5):947-1012, 2016.
(1b) N. Meinshausen, A. Hauser, J. Mooij, P. Versteeg, J. Peters, P. Bühlmann: Causal inference from gene perturbation experiments: methods, software and validation, Proceedings of the National Academy of Sciences 113(27):7361-7368, 2016.
(2) N. Pfister, S. Bauer, J. Peters: Identifying Causal Structure in Large-Scale Kinetic Systems, arXiv 1810.11776.