Research map / Graph learning / Graph self-supervised learning
Towards Addressing Frontiers in Graph Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A doctoral dissertation summary that maps frontiers in graph generation via a social-network GRAN case study, then proposes HiGGs for hierarchical large-graph generation and ToP for topology-only multi-domain graph pre-training.
Problem
Graph generative models are limited in feature diversity, structure variety, graph size, and rely on aggregate or domain-locked quality metrics, and are locked to single domains.
Approach
Hierarchical top-down generation of graphs via graph partitioning and per-level generative models (HiGGs), plus topology-only pre-training across multiple graph domains (ToP) that transfers to downstream tasks without features during pre-training.
Key results
- HiGGs-DGD generates graphs of over twenty thousand nodes and hundreds of thousands of edges, a quadratic increase in possible scale over DiGress alone.
- ToP transfer is significantly positive on 75% of experiments and never significantly negative where not positive.
- A model pre-trained only on molecules performs significantly worse on most datasets than one pre-trained only on non-molecules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train with ToP before downstream fine-tuning | ToP-pretrained graph models · 75% of experiments | downstream performance: 75% | Helped |
| Use diverse pre-training data rather than only in-domain data | ToP · Multiple graph datasets | downstream performance | Helped |
| Pre-train only on molecules instead of only on non-molecules | ToP · Majority of datasets | downstream performance | Hurt |
| Apply feature and structure noise during pre-training | ToP · Downstream fine-tuning | reliance on structural information | Helped |
Surprising findings
- Diversity of pre-training domains matters more than in-domain pre-training when node and edge features are absent.
Limitations
- Metric work based on ToP is ongoing and not yet established as a graph generator metric.
- Current GNN generators are limited in feature diversity, structure variety, and size.
- Quality metrics for graph generation are aggregate-based and unexpressive.
Experimental setup
Datasets: social network data molecular graph datasets
Metrics: aggregate graph statistics transfer performance on downstream fine-tuning
Baselines: GRAN rule-based graph generators
Builds on: GRAN DiGress Message Passing Neural Networks
Closest papers by meaning
- HiGen: Hierarchical Graph Generative Networks (ICLR 2024 · similarity 0.90)
- Generative Graph Pattern Machine (NeurIPS 2025 · similarity 0.89)
- A Non-Asymptotic Convergent Analysis for Scored-Based Graph Generative Model via a System of Stochastic Differential Equations (ICML 2025 · similarity 0.88)
- On the Role of Edge Dependency in Graph Generative Models (ICML 2024 · similarity 0.88)
- Efficient and Scalable Graph Generation through Iterative Local Expansion (ICLR 2024 · similarity 0.87)
- What Is Missing For Graph Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.