SAMAG: Structure-Aware Multi-Agent Graph Generation with Large Language Models

Jingcheng Cen, Jiarui Ji, Zhen Wang, Zhewei Wei, Yaliang Li, Bolin Ding

Abstract: Graph generation is fundamental in network science and graph machine learning, yet existing rule-based and deep learning models either miss fine-grained structures or remain static and data-hungry, while recent LLM-based multi-agent methods suffer from structural myopia by ignoring graph topology. We propose a Structure-Aware Multi-Agent-based Graph generation framework (SAMAG), which integrates structureaware information retrieval and agent orchestration to combine semantic and topological context, enabling agents to form realistic communities and yield temporally coherent interaction patterns. Experiments across various domains demonstrate that SAMAG consistently achieves state-of-the-art graph-level fidelity and outperforms previous LLM-based methods in communitylevel fidelity, while in the inductive setting improves GNN node classification accuracy by $18.5\%$ on average over the bestperformed baseline. SAMAG establishes the first structure-aware LLM-based framework, advancing both the fidelity and transferability of synthesized graphs. Code for SAMAG is available at https://github.com/JasonCen-sweetdreams/SAMAG.

External IDs:dblp:conf/bigdataconf/CenJWWLD25