Speaker: Tobias Boege

Abstract: The time-honored tool for modeling cause-effect relationships is a Bayesian network. This model postulates noisy functional dependencies among random variables according to a directed graph. Bayesian networks enjoy widespread use despite two shortcomings: (1) different causal specifications may define the same class of probability distributions, so cause-effect relationships cannot be learned reliably from observational data alone; and (2) if the directed graph contains cycles (“feedback loops”), the model loses desirable statistical properties such as global identifiability of parameters.

In this talk I want to introduce an emergent alternative paradigm in causal modeling which aims to address these issues. The interactions of random variables are modeled by stochastic diffusion processes in equilibrium. This temporal perspective easily accommodates feedback loops. I will cover the algebraic nature of these models, their independence structure and recent identifiability results.