Giangiacomo Mercatali "Controllable Generative Modeling for Graphs and Dynamical Systems"

This talk can be viewed via zoom. (Note: this talk will be recorded)

Title: Controllable Generative Modeling for Graphs and Dynamical Systems

Abstract:
Recent progress in generative modeling has made it possible to learn from increasingly structured scientific data, but important challenges remain in controllability, efficiency, and interpretability. In this talk, I will present two recent works addressing these issues in graph-structured and dynamical settings. First, I will introduce Diffusion Twigs, a conditional graph generation framework based on multiple interacting diffusion processes, where a trunk process models graph structure and stem processes guide desired graph properties. This design enables accurate and flexible conditional generation for molecular and network graphs. Second, I will present Graph Neural Flows, a continuous-time framework for irregularly sampled multivariate time series that jointly learns system dynamics and a directed interaction graph, enabling prediction together with interpretable structure discovery. Together, these works illustrate a common research direction: building generative models for scientific data that are both controllable and structurally informative.