Seminar in applied mathematics and statistics

SPEAKER: Daniel Malinsky (Carnegie Mellon University)

TITLE: Learning ancestral graph Markov models from multivariate time series

ABSTRACT: Algorithms which estimate causal graphical models from observational data can be used to infer causal relations and inform decisions about interventions. While most research in causal structure learning has focused on the i.i.d. domain, it is in many cases straightforward to adapt existent algorithms to time series data. I discuss modeling dynamic systems with ancestral graph Markov models, which are well-suited to domains with possible unmeasured confounders (i.e., latent variables) since ancestral graphs are closed under marginalization. I give an overview of ancestral graphical models for time series with an emphasis on their relationship to structural vector autoregressions (SVARs), and present constraint-based and score-based procedures for learning equivalence classes of (dynamic) ancestral graphs.

-----

Tea and chocolate will be served on the 4th floor (4-4-19) after the seminar.

-----

Upcoming seminars in Applied Mathematics and Statistics:

Friday, October 27, 13:15 : Martin Jönsson (Oxford)

Friday, November 3, 13:15 : Lars Nørvang Andersen (Aarhus)

Thursday, January 11, 15:15 : Ryan Tibshirani (Carnegie Mellon University)