ANC / Data Science CDT Seminar Seminar: Ricardo Silva from UCL, Chair: Charles Sutton

Causal Inference in Machine Learning: From Structure to Predictions via Observational Data

Any intelligent agent that wants to achieve a particular goal has to predict the causal effects of its actions. For instance, in reinforcement learning one can try multiple actions, see what happens, fit a model, then profit from it. In many real applications, however, this scenario is almost a fantasy: performing actions in a controlled environment is costly, sometimes impossible, as in epidemiological studies. Yet there might be sources of cheap data where actions were chosen by unknown mechanisms. This is known as observational data, and the problem in this case is that the choice of action might be confounded with the outcome you want to achieve, popularly known as the "correlation is not causation" slogan.

 In machine learning, there is a rich literature on structure learning of graphical models. Under assumptions, this allows for an understanding of which causal pathways exist in a system. Less common is how to exploit that to make predictions: given treatment X on  an outcome Y, we want an estimate of what Y will be when we control X. We introduce two new approaches to solve this problem: one for continuous data, where linear causal effects can be found by looking for constraints implied by "surrogate experiments" known as instrumental variables, which control X indirectly by a mechanism that we can discover in principle. The other approach, for discrete data, finds robust bounds on causal effects by allowing for a whole continuum of violations of the structural assumptions used by common structure learning algorithms.

Joint work with Robin Evans (Oxford) and Shohei Shimizu (Osaka).