ANR-Lab Seminar "Cusna: Statistical Network Models on GPUs"

On September 21 at 3:00 PM, the International Laboratory for Applied Network Research will hold its first seminar of the new academic year. Artem Maltsev, a lecturer in the Department of Politics and Governance at the HSE University, will speak at the seminar.

Stochastic actor-based models (SAOM, RSiena family) and exponential random graph models (ERGMs) are traditional approaches in inferential network analysis. However, the practical use of these methods is often limited by computational complexity—simulation-based methods for estimating Markov chain-based models often take hours or even days to produce results for a single run. The cusna package for R and Python offers a solution to this problem using GPU parallelization: individual simulation chains are computed simultaneously on separate CUDA cores on the graphics card. Preliminary experiments show that computational speedups reach two orders of magnitude. At the seminar, the package's author will explain the design of the cusna engine and how it integrates with familiar estimators (siena07() and others). Special attention will be given to the new potential methodological capabilities of the new package: goodness-of-fit tests using dozens of auxiliary statistics, exhaustive searches for multiple alternative specifications of endogenous network parameters, the use of a long third phase of the algorithm to significantly reduce the noise of t-statistics, simulation studies with thousands of replications, including counterfactual simulations, and a controlled cross-comparison of SAOM and (S)TERGM. Finally, new approaches to specialized estimators (e.g., an amortized Bayesian estimator for SAOM) will be explored.  

Event language: Russian  

Registration is required to participate.