GraphMaster: Automated Graph Synthesis via LLM Agents in Data-Limited Environments

GraphMaster is a novel multi-agent system for graph data enhancement, built upon the Retrieval-Augmented Generation (RAG) paradigm and powered by Large Language Models (LLMs). It is designed for few-shot or low-resource graph learning tasks, where both semantic diversity and structural quality are critical.

🚀 Key Features

• Multi-Agent Architecture simulating human-in-the-loop perception, enhancement, evaluation, and management.
• RAG-based Iterative Enhancement over graph data using LLMs.
• Semantic & Topological Modes for diversified and structure-aware node generation.
• Auto-Adaptive Objective Weights across semantic, structural, and label balance metrics.
• Plug-and-Play LLMs: Easily switch between Qwen, Deepseek, LLaMA, or any HF-supported model.
• Data-Limited Datasets: For more details, please refer Dataset_Creation README.

🧠 Architecture

+--------------------+ +--------------------+ +------------------------+
| Perception Agent | --> | Enhancement Agent | --> | Evaluation Agent |
+--------------------+ +--------------------+ +------------------------+
^ |
| v
+--------------------+ +------------------+
| Manager Agent |<------------------------| Enhanced Graph |
+--------------------+ +------------------+

📂 Project Structure

\src
├── main.py # Entry point
├── manager_agent.py # Agent that controls the full pipeline
├── perception_agent.py # Builds graph, samples subgraphs, computes stats
├── enhancement_agent.py # Generates new nodes (semantic/topological)
├── evaluation_agent.py # Evaluates generated nodes and detects convergence
├── data/
│ └── cora.json # Input graph (JSON format)
\data # data-limited datasets, and the corresponding generate data \log # logs while run the pipline
\tricks # Some preprocessing codes
\Vertification # GNN verification model, used for Bert&GNN to verify data effects

📦 Installation

conda create -n graphmaster python=3.11
conda activate graphmaster
pip install -r requirements.txt

Requirements include transformers, networkx, scikit-learn, community (for Louvain), matplotlib

The experiment is best run on either 8 A6000 GPUs with 48GB memory each or 4 A100 GPUs with 80GB memory each. However, based on our experiments, a single A100 GPU with 80GB memory can also run the experiment, albeit with a significant increase in runtime.

📄 Input Format

Each node is described in JSON:

{
"node_id": "123",
"label": 2,
"text": "A novel GNN model is proposed...",
"neighbors": ["45", "78"],
"mask": "Train"
}

🧪 Running the Pipeline

cd src
python main.py \
--data_file ./data/SubCora.json \
--llm_model QwQ \
--enhancement_mode semantic \
--max_iterations 10 \
--visualize_sampling

or

python3 main.py \
--llm_model path/to/Qwen3-VL-8B-Instruct/ \
--gpu 0,1,2,3,4,5,6,7 \
--data_file ../data/SubCora.json

Supported --llm_model:

• Qwen → Qwen1.5-32B
• Deepseek → DeepSeek-R1-Distill-Qwen-32B
• LLaMA → Samantha 1.1 (LLaMA 33B)
• QwQ → Qwen/QwQ-32B (preview model)
• Qwen3-VL-8B

Custom models also supported by providing HF path.

📈 Outputs

• Enhanced graph stored in cora_enhanced.json
• Adaptive weights saved per iteration
• Visualizations:
  • adaptive_weights_evolution.png
  • label_distribution_change.png

Verification

For Verification, please refer to Verification_README

🤖 Agent Highlights

PerceptionAgent

• Graph construction (using NetworkX)
• Louvain community detection with semantic similarity
• PPR-based sampling from high-variance community

EnhancementAgent

• Prompt-based LLM generation
• Supports both semantic and topological enhancements
• Edge construction via probabilistic model (sim + overlap + centrality)

EvaluationAgent

• Computes composite quality score (0-10 scale)
• Adaptive threshold & early stopping
• Convergence analysis using quality gradients + LLM summary

ManagerAgent

• Controls the full loop
• Auto-selects enhancement mode based on multi-objective utility
• Updates adaptive weights (λ₁, λ₂, λ₃)

Datasets

Full source datasets are open-source at https://huggingface.co/datasets/EnjunDu/GraphMaster.

📊 Citation-Style Motivation

"GraphMaster simulates a human-guided editing process on attributed graphs by iteratively improving data with structured perception, controlled generation, and critical evaluation — powered by LLMs."

📘 License

MIT License