Urban Region Pre-training and Prompting: A Graph-based Approach

2024-08-12

arXiv.org

AI Summary

To address the challenges of acquiring generalizable knowledge and adapting to downstream tasks in urban region representation, this paper proposes a spatial-entity-augmented regional graph modeling framework. It pioneers the unification of graph pretraining and graph prompting in urban computing: (i) a subgraph-centered self-supervised pretraining paradigm is designed, integrating structure-aware masked subgraph reconstruction with subgraph-level contrastive learning; and (ii) a dual-path graph prompting mechanism is introduced, jointly injecting task-specific knowledge via explicit template-based and implicit soft prompts. Evaluated across multiple cities and diverse downstream tasks—including POI prediction, crowd flow forecasting, and functional zone identification—the method consistently outperforms state-of-the-art approaches, achieving average accuracy improvements of 5.2%–9.7%. The framework significantly enhances the generalizability of regional representations and enables effective cross-regional knowledge transfer.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Urban region representation is crucial for various urban downstream tasks. However, despite the proliferation of methods and their success, acquiring general urban region knowledge and adapting to different tasks remains challenging. Previous work often neglects the spatial structures and functional layouts between entities, limiting their ability to capture transferable knowledge across regions. Further, these methods struggle to adapt effectively to specific downstream tasks, as they do not adequately address the unique features and relationships required for different downstream tasks. In this paper, we propose a $ extbf{G}$raph-based $ extbf{U}$rban $ extbf{R}$egion $ extbf{P}$re-training and $ extbf{P}$rompting framework ($ extbf{GURPP}$) for region representation learning. Specifically, we first construct an urban region graph that integrates detailed spatial entity data for more effective urban region representation. Then, we develop a subgraph-centric urban region pre-training model to capture the heterogeneous and transferable patterns of interactions among entities. To further enhance the adaptability of these embeddings to different tasks, we design two graph-based prompting methods to incorporate explicit/hidden task knowledge. Extensive experiments on various urban region prediction tasks and different cities demonstrate the superior performance of our GURPP framework.

Problem

• Urban region representation lacks general knowledge adaptation
• Fine-grained functional layout semantics are often overlooked
• Inadequate handling of unique features hinders task adaptation

Innovation

• Graph-based urban region pre-training model
• Contrastive and multi-view learning methods
• Graph-based prompting for task adaptation