Skip to content

What a generated graph should preserve

How can generated graphs preserve the structures that distinguish real network domains, and how can generation be steered towards a target class?

A noisy random graph branches into two generated graphs: a sparse spatial graph beside a transmission tower, for infrastructure and transport networks, and a modular graph beside a howling wolf, for social networks.
Generation steered towards a target class: the same noisy start resolves into a graph with the structure of an infrastructure network or of a social network. The graphs are drawn by simple rules to illustrate the idea, not sampled from a trained generator.
highlights

Generating graphs with plausible summary statistics does not guarantee that they retain the structural characteristics of a real network domain. Our work examines both how to evaluate this correspondence and how to guide generation towards a specified class of networks.

Beyond MMD introduces representation-aware evaluation using geometric deep learning. Its comparison of graph generators shows that agreement on selected topological statistics can coexist with a loss of the structural features that distinguish graph domains.

ProtoGuide uses a learned graph representation and target-class prototypes to guide a frozen discrete graph generator. The guidance can be retargeted by changing the prototype without retraining the generator. Experiments with two generative backbones show improved target-class fidelity, with effects on coverage and diversity that vary across classes and settings.

Both selected papers are preprints. Together they address a common question: what should a generated network preserve beyond a small set of summary statistics?

Selected papers

Salvatore Romano, Marco Grassia, Pietro Liò, Giuseppe Mangioni (2026).
ProtoGuide: Prototype-Driven Guidance for Class-Conditional Graph Generation.
Preprint DOI
Salvatore Romano, Marco Grassia, Giuseppe Mangioni (2025).
Beyond MMD: Evaluating Graph Generative Models with Geometric Deep Learning.
Preprint DOI

Research topic: Learning on Graphs