UniRoad: Unifying Road Representations Across Cities and Tasks
Abstract
Learning generalizable road representations is essential for urban computing, yet remains challenging due to heterogeneous urban signals, city-specific road characteristics, and diverse downstream tasks. Existing methods often encode multi-source information jointly, causing functional, structural, and mobility semantics to be entangled in a single representation space. We propose UniRoad, a unified multi-view framework for cross-city road representation learning. UniRoad decomposes road semantics into three complementary views: an LLM-based profile view that captures functional semantics from road attributes and POI contexts, a prototype-based structure view that learns transferable structural roles from local topology, and a trajectory-informed mobility view that models travel patterns from road sequences. UniRoad further introduces confidence-guided fusion to adaptively integrate these views according to their reliability. Extensive experiments on four cities and four downstream tasks demonstrate that UniRoad consistently outperforms strong baselines and achieves robust cross-city transferability. Code and data are available at: https://github.com/csjiezhao/UniRoad.