Title: Bayesian extensions of neural network-based graphon approximations
Authors: Creighton Heaukulani - Goldman Sachs (Hong Kong) [presenting]
Abstract: The perspective of directly modeling the graphon of an exchangeable graph has, naturally, been approached recently with neural network function approximations. Desirable properties in many statistical network applications, such as block structure (i.e., community detection) and dynamic extensions are straightforwardly accomplished using tools from deep learning. We will discuss these two extensions, in particular, and see that inference for our suggested models demands a Bayesian approach. Several applications in the context of finance will be demonstrated.