Synthetic Graph Data Augmentation for Generative Digital Twins
2026 IEEE International Conference on Communications Workshops, ICC Workshops 2026, Glasgow, İngiltere, 24 - 28 Mayıs 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/iccworkshops63917.2026.11586654
- Basıldığı Şehir: Glasgow
- Basıldığı Ülke: İngiltere
- Anahtar Kelimeler: digital twin (DT), generative AI, generative graph learning, graph learning, scalable network modeling, synthetic data augmentation
- Atatürk Üniversitesi Adresli: Evet
Özet
Digital twin systems rely heavily on high-fidelity data to simulate, analyze, and optimize real-world processes. In domains such as communication networks, power grids, transportation systems, and industrial automation, these systems are naturally modeled as graphs, where nodes represent entities (e.g., routers, vehicles, sensors) and edges represent relationships (e.g., data flows, road links, control dependencies). However, collecting large-scale, structurally diverse graph data for training and experimentation remains a major bottleneck, especially in networked systems where real deployments are expensive or impractical. To address this challenge, we propose a scalable, generative graph learning framework that synthesizes larger, structurally consistent graphs from smaller input samples, called GraphSynTwin. Our method captures latent structural patterns using permutation-invariant embeddings and generates synthetic node representations through probabilistic modeling. Connectivity is reconstructed via a geometric decoding scheme that preserves topological characteristics such as community structure and degree distribution. Beyond standalone generation, we demonstrate the effectiveness of this approach for graph-based data augmentation. Through experiments on benchmark datasets, including node classification tasks, we show that augmenting training sets with synthetically generated graph data consistently improves downstream model accuracy. These results highlight the potential of generative graph learning as a powerful tool for enhancing digital twin simulations and enabling robust, data-efficient learning in graph-centric domains.