The evolution of many real-world systems is best described by dynamic graphs, whose statistical properties reflect the constraints of the system. When forecasting their dynamics, the goal is to generate a time series of graphs respecting these underlying constraints. Existing scalable dynamic graph learning methods, however, are designed for local tasks such as link prediction or node classification, and their independent, local predictions are ill-suited for graph generation. This limitation is particularly relevant for discrete-time dynamic graphs, where coarse time resolution induces dependencies among edges within each snapshot. We propose using a generalized notion of degrees to model such dependencies directly, thereby shifting the focus from individual links to node dynamics. This approach bypasses the need to learn a sparse graph representation and yields an inductive representation that enables the generation of large-scale discrete-time dynamic graphs.

Please note that this is an internal talk. If you are interested in participating in person, please write an email to