This talk has captions. You can remove these by pressing CC on the video toolbar.
Name: Jonas Peters
Talk Title: Can causal discovery benefit from extreme values?
Abstract: In causal discovery, the goal is to learn causal structure from observational data. In this talk, we show two examples of methods that can potentially benefit from extreme values. (1) In linear models, the graph structure becomes identifiable if the noise variables have a non-Gaussian distribution. (2) For heterogeneous data sets, in which the causal mechanism for a target variable remains the same for all observations, it has been suggested to learn causal structures by searching for invariant models. We briefly introduce such methods and provide intuition for why it may help if the distributions are heavy-tailed. No prior knowledge about causality is required.
This talk is an invited talk at EVA 2021. View the programme here.