Agenda

Speaker:
Giacomo Rosin, Dipartimento Scienze Ambientali, Informatica e Statistica - Università Ca' Foscari Venezia

Abstract:
Graph generation has traditionally focused on static structures, despite the inherently dynamic nature of many real-world systems such as molecular interactions and evolving networks. In this talk, we begin by revisiting static graph generative modeling and its limitations when applied to dynamic settings. We then introduce TGGM, a Temporal Graph Generative Model that formulates temporal graph generation as next-token prediction over sequences of timestamped edge events using a decoder-only transformer. Unlike prior approaches that train on single graph instances, TGGM is designed for multi-sample datasets, enabling learning over graph distributions rather than individual examples. Building on this formulation, the framework further supports multi-dataset training, which improves robustness across related graph families and enables zero-shot generation on unseen graph and flexible conditioning. Four temporal modeling strategies are examined, topology-only generation, timestamp discretization, autoregressive digit modeling, and continuous-time regression, and their trade-offs. Evaluation employs temporal MMD-based metrics to assess both structural and temporal fidelity.

Bio Sketch:
Giacomo Rosin is a Research Assistant at Ca' Foscari University of Venice. He earned both his Bachelor's and Master's degrees in Computer Science from Ca’ Foscari University, graduating with full marks (110/110 cum laude). His research focuses on machine learning, with a particular interest in generative models. He works on generative approaches for both static and temporal graphs, exploring their theoretical and practical applications. His broader research interests include computer vision and natural language processing, reflecting a strong interdisciplinary orientation toward data-driven modeling and prediction.