Learning Causal Models via Algebraic Constraints

Explore the intricacies of causal inference in this André Aisenstadt Prize Lecture delivered by Elina Robeva. Delve into the challenge of learning direct causal relationships among observed random variables, typically represented by directed graphs. Examine the problem of learning directed graphs for linear causal models, with a focus on cases involving directed cycles or hidden variables. Discover why causal graphs generally cannot be uniquely determined from observational data alone, and learn about the special case of linear non-Gaussian acyclic causal models where unique graph identification is possible. Investigate the characterization of equivalence classes for cyclic graphs and the proposed algorithms for causal discovery. Gain insights into the use of specific polynomial relationships among 2nd and higher order moments of random vectors to aid in graph identification.