LIU, YONGHUI (Korea Advanced Institute of Science and Technology), Kim, Inhi (Korea Advanced Institute of Science and Technology)

A Generalizable Semantic Graph Learning Framework for Road Network Representation

Scheduled for presentation during the Regular Session "S09c-Optimization for Multimodal and On-Demand Urban Mobility Systems" (WE-LA-T9), Wednesday, November 19, 2025, 16:00−16:20, Coolangata 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

Road networks (RNs) are fundamental to transportation systems, serving as critical input for applications like traffic prediction and autonomous vehicle navigation. However, effectively representing large-scale RNs is challenging due to their complex topology, dynamic evolution, and heterogeneous associated data. Existing approaches often struggle with scalability, generalization to new cities, and integrating contextual semantic information. In this work, we propose HetSemNet, a heterogeneous graph framework that enriches road network graphs with semantic context nodes derived from point-of-interest (POI) descriptions encoded by a pre-trained large language model (LLM). On this graph, we introduce SaRNSAGE, a semantic-aware graph neural network that employs a sampling-and-aggregation mechanism to learn robust and transferable road segment representations. SaRNSAGE jointly captures road connectivity and surrounding semantic context, leveraging basic road attributes as initial features while avoiding complex feature engineering. Experiments on real-world urban road networks demonstrate that our method significantly outperforms conventional GNN baselines on a road type classification task and exhibits strong generalization in a cross-city transfer scenario. These results underscore the effectiveness of combining LLM-derived semantic information with heterogeneous graph modeling for generalizable road network representation learning.