🚀 SCOUT: Structure-Aware Aspect and Anchor-Count Selection for Node Attribute Augmentation via Positional Information

SCOUT is a model-agnostic augmentation framework that enhances graph neural networks (GNNs) when node attributes are missing, sparse, or uninformative, by leveraging multi-aspect positional information and a graph-aware anchor selection mechanism.

---

📖 Abstract

When node attributes are absent or limited, GNNs often fail to distinguish structurally similar nodes, leading to degraded downstream performance.
SCOUT addresses this by:

• Selecting positional aspects (centrality–similarity pairs) via a graph-level attention mechanism.
• Determining the anchor-count per graph using a principled elbow method grounded in power-law centrality distributions.
• Augmenting node features with positional information that complements original attributes (when present).

This results in significant gains across link prediction and node classification tasks, both with and without node attributes.

📄 Paper: SCOUT: Structure-Aware Aspect and Anchor-Count Selection for Node Attribute Augmentation via Positional Information
🌐 Conference: The Web Conference (WWW), 2026
📁 Code: https://github.com/seinkim01/SCOUT

---

📚 Table of Contents

• Installation
• Usage
• Project Structure
• Features
• Configuration
• Dependencies
• Examples
• Experimental Results
• Troubleshooting
• Contributors
• License

---

⚙️ Installation

We recommend using Conda to manage dependencies:

# Recommended
conda env create -f requirements_conda.yaml
conda activate scout

Or, use pip:

pip install -r requirements.txt

---

🚀 Usage

To run link prediction on Cora with SCOUT:

bash scripts/run_linkpred.sh

For node classification:

python src/core/train_nodeclf.py

All major configurations can be modified via the config files or script arguments.

---

🧱 Project Structure

SCOUT/
├── attrs/                      # Precomputed centrality & similarity scores
├── datasets/                  # Graph datasets (e.g., Cora, Citeseer)
├── logs/                      # Training logs
├── results/                   # Output metrics and predictions
├── scripts/                   # Automation scripts for experiments
├── src/
│   ├── core/                  # Training pipelines & preprocessing
│   ├── models/                # GNN modules, decoder, attribute gates
│   └── utils/                 # Data loading, helper functions
├── requirements.txt
├── requirements_conda.yaml
└── README.md

---

✨ Key Features

• ✅ Model-agnostic augmentation: Integrates with any GNN backbone.
• 🧠 Graph-level positional aspect selection: Learns which structural features matter.
• 🎯 Elbow-based anchor-count detection: Automatically selects anchor nodes per graph.
• 📈 Performance improvements on OGB & citation benchmarks, both with/without features.
• 🔧 Supports downstream tasks: link prediction & node classification.

---

⚙️ Configuration

• Datasets: Place in datasets/
• Attributes: Precompute and store in attrs/
• Custom aspects: Modify generate_attributes.py
• Logging: Enabled via logs/ directory
• Models: Can be swapped or extended under src/models/

---

📦 Dependencies

• Python ≥ 3.8
• PyTorch
• DGL or PyG
• NumPy, SciPy, tqdm
• cuGraph (optional for acceleration)

Check requirements_conda.yaml for exact versions.

---

🔍 Examples

# Link prediction with SCOUT augmentation
bash scripts/run_linkpred.sh

# Node classification on augmented features
python src/core/train_nodeclf.py

---

📊 Experimental Results

SCOUT achieves:

• +26.88% Hits@20 on ogbl-ddi (w/o original attributes)
• +11.69% accuracy on ogbn-mag (w/ original attributes)
• Outperforms HPLC, P-GNN, SEAL and others across tasks

Refer to the paper or results/ for detailed tables and plots.

---

🛠️ Troubleshooting

• Issue: Attribute files not found?
  Ensure correct folder structure under attrs/.

• Issue: CUDA memory overflow?
  Reduce batch size or number of anchors (K).

• Using PyG or DGL?
  Modify model imports in src/models/.

---

👥 Contributors

• Dong-Hyuk Seo — Hanyang University
• Sein Kim — Hanyang University
• Taeri Kim — Hanyang University
• Won-Yong Shin — Yonsei University
• Sang-Wook Kim — Hanyang University

---

📄 License

📌 This code is currently under review for publication at WWW 2026.
The license will be updated upon acceptance. For academic use only.

---

📣 For citation, please refer to the paper once it's published officially.