Department Seminar Series: Rina Foygel Barber, Robust inference on edges in non-Gaussian graphical models with ROCKET
In many modern scientific applications, high-dimensional data is often represented with a graphical model, where edges represent two variables that are dependent even after conditioning on all other variables. There has been extensive work on sparse Gaussian graphical models, where the goal is often to identify the sparse graph structure over p many variables using n<p samples. However, in many fields, real data may be far from Gaussian - the data may be marginally heavy-tailed, or may exhibit heavy tail dependence between multiple variables. While the first problem can be addressed by a Gaussian copula model (also known as the nonparanormal model), the second issue cannot.