Perfect simulation and inference for Exponential Random Graph Models

Wednesday 10.04, USI Lugano Campus, room A-13, Red building, 12:30-13:30

Joris Bierkens, Technical University Delft, Netherlands

Abstract:
A useful tool in understanding social networks from a statistical perspective is the flexible family of Exponential Random Graph Models (ERGMs). Although this is a highly versatile class of statistical models, it has the difficulty that it can be hard to compute expectations with respect to these models, or even to simulate a random instance of the model. In order to apply ERGMs in practice we require techniques to choose the parameters of the model in a statistically correct way. The above mentioned difficulty in performing computations with respect to the model then yields to a situation which is often called 'doubly intractable likelihood'. In this talk Dr. Bierkens will discuss the use of a new Monte Carlo simulation technique relying on piecewise deterministic Markov processes such as the Zig-Zag process. This methodology yield substantial promise with respect to the estimation difficulty, thanks to the possibility of using unbiased gradient estimates without loss of asymptotic correctness. In order to apply this method, it is necessary to obtain perfect samples from an ERGM, a topic which will also be discussed.

Biography:
Dr. Bierkens is assistant professor at the TU Delft. His interests lie in the computational challenges arising in Bayesian statistics and statistical physics, involving the development and analysis of novel Markov Chain Monte Carlo (MCMC) algorithms. At the moment his research has a strong focus on the use of Piecewise Deterministic Markov Processes for Monte Carlo purposes. This discovery has opened up an entirely new family of elegant algorithms which can be extremely efficient in challenging settings. Dr. Bierkens was funded by the prestigious Dutch VIDI award in 2018.

Host: Prof. Ernst Wit