Mariangeles Serrano Moral
University of Barcelona
December 02
7:30 p.m.
CM 1.26
Renormalization of Complex Networks: From Multiscale Self-Similarity to Zero-Shot GNN Transfer
The renormalization group in physics provides a rigorous framework for understanding systems across length scales. Extending these ideas to complex networks, however, is challenging because their small-world structure induces correlations between coexisting scales and obscures the notion of locality. To address this challenge, we have developed geometric renormalization (GR) methods for both unweighted and weighted networks. Building on their mapping into hyperbolic space, GR reveals multiscale self-similarity as a pervasive organizing principle of real-world networks, including human brain connectomes. This self-similarity also emerges in the growth of certain real networks, whose evolution can be interpreted as a reverse renormalization process.
Beyond providing a theoretical framework for network organization and evolution, GR enables practical applications. It allows us to generate scaled-down and scaled-up network replicas that preserve structural properties, providing controlled substrates for studying size-dependent processes and large-scale simulations. As a recent application, we have shown that Graph Neural Networks (GNNs) trained on GR scaled-down replicas can be transferred zero-shot to the original networks, preserving much of their predictive performance while substantially reducing training cost. These results suggest that the same multiscale structure that underlies renormalization may also enable learning to generalize across scales.