POP-NETS, a DFF Project, aims to develop machine learning for the analysis of population of networks. A network, or graph, is a mathematical structure encoding relational phenomena in as different fields as medicine, sociology, transport etc.. Networks are an example of structured data, where machine learning has classically focused on "simple" predictive tasks such as classification or regressing a single real-valued variable. We will learn network-valued regression models, which will allow us to visualize trends within the domain of networks. Generative models for networks will be used both for uncertainty quantification and for model transparency. These models are far more challenging than graph classification: At their heart, solving these problems requires being able to interpolate between networks, and to define meaningful probability distributions over networks. POP-NETS will utilize the known
structure of network data to obtain structured output models for populations of
networks.