Lizzie Silver is a Research Fellow at the University of Melbourne, where she analyses data for The SWARM Project, an effort to improve reasoning in human teams. She did her PhD at Carnegie Mellon University, on multitask methods for learning causal graphical models.
A tutorial & overview of methods for learning causal graphical models. We need causal models in several situations: to figure out which variables are confounders in observational studies; to suggest new experiments; and to predict the results of interventions if experiments are infeasible or unethical. I'll show you how to learn a causal model from observational data, and I'll show you the strengths and weaknesses of the methods.
Graphical models have a structure and a parameterisation. The structure represents the qualitative causal relationships - "what causes what". The parameterisation represents the strength and functional form of those relationships. You need to know the structure before you can learn the parameterisation.
I'll cover various algorithms for learning the model structure from observational data:
1. The PC algorithm ("Peter and Clark"), and PC-Stable
2. GES (Greedy Equivalence Search)
3. FCI (Fast Causal Inference)
4. LiNGAM (Linear Non-Gaussian Additive Models)
5. If we have time, more fancy stuff: latent variable models, non-linear causal additive models, SAT-solver methods, etc.
And I'll describe some of the major difficulties with causal inference:
* Feature definition
* Measurement error