Abstract
Unsupervised region representation learning aims to extract dense and effective features from unlabeled urban data. While some efforts have been made for solving this problem based on multiple views, existing methods are still insufficient in extracting representations in a view and/or incorporating representations from different views. Motivated by the success of contrastive learning for representation learning, we propose to leverage it for multi-view region representation learning and design a model called ReMVC (Region Embedding with Multi-View Contrastive Learning) by following two guidelines: i) comparing a region with others within each view for effective representation extraction and ii) comparing a region with itself across different views for cross-view information sharing. We design the intra-view contrastive learning module which helps to learn distinguished region embeddings and the inter-view contrastive learning module which serves as a soft co-regularizer to constrain the embedding parameters and transfer knowledge across multi-views. We exploit the learned region embeddings in two downstream tasks named land usage clustering and region popularity prediction. Extensive experiments demonstrate that our model achieves impressive improvements compared with seven state-of-the-art baseline methods, and the margins are over 30% in the land usage clustering task.
Introduction
In urban computing, "region embedding" involves transforming physical city blocks into dense numerical vectors that capture their functional identity—such as identifying a business district versus a residential area. This is essential for applications like real estate valuation, traffic flow prediction, and commercial site selection.
Existing approaches typically face two major challenges. First, many rely on manually defined heuristic graphs to link regions, which fail to capture complex, learnable patterns. Second, methods focused on data reconstruction often miss high-level functional features by over-emphasizing raw data counts. Furthermore, effectively fusing disparate data sources—such as Point-of-Interest (POI) data and human mobility patterns—remains a difficult task, often resulting in simple concatenations that do not ensure a unified representation of a region's identity.
The ReMVC Framework
To address these limitations, we introduce ReMVC (Region Embedding with Multi-View Contrastive Learning). This framework utilizes contrastive learning to distinguish between similar "positive" pairs and dissimilar "negative" pairs across different data views.
Intra-View Contrastive Learning
Within each data view (e.g., POI or Mobility), the model is trained to recognize a region even after data augmentations, such as adding noise to mobility heatmaps or dropping POIs. This forces the model to learn robust, abstract features rather than memorizing raw statistics. The use of "hard negatives"—regions that are feature-wise similar but distinct—further improves the model's discriminative power.
Inter-View Contrastive Learning
This module acts as a cross-check by treating the POI view and the Mobility view of the same region as a positive pair. By aligning these embeddings in a shared space, the model ensures that different data signals describe the same underlying functional identity. This serves as a soft co-regularizer, preventing any single view from dominating and facilitating knowledge transfer between views.
Experimental Evaluation
We evaluated ReMVC using real-world data from Manhattan, New York City. The dataset included 270 census blocks, approximately 10 million taxi trip records for mobility heatmaps, and over 17,000 POIs across 200+ categories.
Quantitative Results
ReMVC was tested on two downstream tasks: Land Usage Clustering and Region Popularity Prediction. As shown in the results below, ReMVC significantly outperforms seven state-of-the-art baselines, including M2GRL and MVURE.
In the land usage clustering task, ReMVC achieved a breakthrough improvement of over 30% in the Adjusted Rand Index (ARI) compared to the previous best-performing method. It also demonstrated superior accuracy in predicting region popularity, as evidenced by the lowest MAE and RMSE scores and the highest R² value.
Conclusion
The success of ReMVC demonstrates that contrastive learning is more effective for region embedding than traditional reconstruction methods. By focusing on distinguishing properties rather than average densities, the model captures the unique functional character of urban areas. Future work will explore domain adaptation for different city structures, improved scalability for larger urban areas, and the integration of temporal dynamics to capture how urban functions evolve over time.
Summary updated