User-Centric Synthetic Road Network Generation Using Graph Neural Networks and Graph Autoencoders

This paper introduces a novel method for generating realistic synthetic road networks crucial for urban planning, traffic management, and intelligent transportation systems. The approach uses Graph Neural Networks (GNNs) to learn embeddings from real-world road data, capturing structural characteristics without needing explicit node features. Clustering analysis is applied to these embeddings for quality enhancement and customization. Then, using Graph Autoencoders (GAEs), synthetic road networks are generated, guided by user-provided keywords like "Istanbul-like" or "dense road network". The keywords allow for drawing samples from clusters identified previously, yielding synthetic networks closely resembling real-world networks. The approach, thus, effectively generates diverse road networks without needing node feature prediction models, significantly contributing to the field of synthetic road network generation.