Large Structures Seminar: Elizabeth Gross

When using statistical models for network data, we would like to know the goodness-of-fit of the model (i.e., how well the model fits the data). This question has proved particularly challenging even for relatively simple classes of network models, as it currently requires sampling graphs with the same sufficient statistics (e.g., number of edges, number of triangles, degree sequence, etc) as the observed network. In this talk, we will present a method for goodness-of-fit testing for log-linear network models that is rooted in computational algebraic geometry. We will demonstrate the approach on the Holland-Leinhardt p_1 model for random directed graphs. This is joint work with Sonja Petrovic and Despina Stasi.