Abstract

There is a difference between “seeing” and “doing”. Consider the following example: while one might observe that people in Florida live longer, this does not mean one should consider moving there to achieve longevity. It is actually the case that many retired Americans move to Florida, which explains its high life expectancy. This illustrates that we can have a feature that is both useful for predicting life expectancy, but at the same time worthless when considered as a treatment.

To predict effects of interventions such as medical treatments, public policies, or even the behaviour of genetically engineered cells, one needs a causal model. While there exists a well-established machinery designed to estimate such models using experimental data, quite often one cannot perform experiments for reasons such as high costs or ethical issues. Observational data (i.e., non-experimental), however, can be easily collected in many cases: data on the association between smoking and lung cancer is the classical example.

In this talk we will explore several modern techniques of learning causal effects from observational data. Such techniques rely on important assumptions linking statistical distributions to causal connections and can be explored in many exciting ways by machine learning algorithms. Inferring causality from observational data is certainly among the hardest learning tasks of all, but it is also a task with high pay-offs.

The outline of the talk is as follows:

- Motivation and definitions / observational studies / the problems of directionality and confounding

- Languages for causal modeling: graphical models and potential outcomes

- Identification of effects using graphical models

- Notions of learning causal structure from data

- Applications