Abstract: Understanding the behaviour of a system under the influence of interventions is the ultimate goal of many scientific studies. As a result, there is growing interest in the possibilities of identifying and estimating causal effects P(y|do(x)) without often unavailable randomized controlled trials. We present new theoretical results that appear recently at the NeurIPS conference. The first approach extends the identifiability of do-calculus, by taking into account context-specific independence relations. The second approach produces a Bayesian posterior distribution of linear causal effects in the most basic setting, where we only have a passively observed set of measurements over the variables of interest. This is based on a state-of-the-art MCMC posterior sampling approach for DAGs. The presentation includes the background necessary to understand these new results.

Speakers:  Antti Hyttinen & Jussi Viinikka

Bios: Dr. Hyttinen is a University Researcher in the Sums of Products research group at the University of Helsinki CS department. He received his Ph.D. in 2013 at the University of Helsinki. He has worked on various aspects of causal inference, such as structure discovery, causal effect estimation, and experimental design. 
M.Sc. Jussi Viinikka is a first-year doctoral student in the Sums of Products research group at the University of Helsinki CS department; he is supervised by Mikko Koivisto.

Affiliation: Department of Computer Science, Helsinki University

Place of Seminar:  Zoom (Now available on Youtube)