Region Embedding With Intra and Inter-View Contrastive Learning

Liang Zhang, Cheng Long, Gao Cong

TL;DR

ReMVC (Region Embedding with Multi-View Contrastive Learning) is a new framework that generates high-quality numerical representations of city regions by analyzing both Point-of-Interest (POI) data and human mobility patterns. By using contrastive learning to compare regions both within a single data "view" and across different views, it achieves a massive 30% performance boost over state-of-the-art methods in urban planning tasks like land-use clustering.

What Problem Does This Paper Address?

In urban computing, "region embedding" means turning a physical city block into a dense vector of numbers that captures its function (e.g., is this a business district or a park?). This is vital for predicting things like real estate prices, traffic flow, or commercial popularity.

Existing methods usually fall into two traps:

  1. Heuristic Graphs: They rely on manually defined rules to link regions, which can't capture complex, learnable patterns.
  2. Over-focus on Detail: Methods that try to "reconstruct" raw data (like counting every single residential building) often miss the "big picture" functional features—like a few specific bars or transit hubs that actually define a neighborhood's character.

Furthermore, most models struggle to effectively fuse different types of data (like POIs vs. taxi trips), often just "sticking them together" without ensuring they actually represent the same underlying reality.

Methods

The authors introduce ReMVC, which leverages Contrastive Learning—a technique where a model learns by distinguishing between similar "positive" pairs and dissimilar "negative" pairs. The framework has two main pillars:

  • Intra-View Contrastive Learning: Within a single data type (e.g., just POIs), the model is trained to recognize a region even if the data is slightly "jittered" or augmented (adding/removing POIs or adding noise to mobility heatmaps). This forces the model to learn robust, abstract features rather than memorizing raw counts.
  • Inter-View Contrastive Learning: This is the "cross-check." The model treats the POI view and the Mobility view of the same region as a positive pair. It forces the embeddings from both views to align in a shared space. If the POI data says "this is a shopping mall" and the mobility data says "people visit here at 2 PM on Saturdays," the model learns that these two signals describe the same functional identity.

The final region representation is a concatenation of the features learned from both views.

Datasets

The study uses real-world data from New York City (specifically the Manhattan borough):

  • Census Blocks: 270 distinct regions defined by the US Census Bureau.
  • Taxi Trips: Approximately 10 million taxi records from August 2013, used to build mobility heatmaps (pick-up/drop-off locations and timestamps).
  • POI Data: Over 17,000 Points-of-Interest categorized into 200+ fine-grained types (e.g., restaurants, theaters, schools).

Key Insights and Ideas

  • Contrastive > Reconstruction: The authors found that contrastive learning (asking "is this the same place?") is far more effective for region embedding than traditional reconstruction (asking "can I recreate the raw data?"). This is because urban functions are often defined by "distinguished properties" (like a specific landmark) rather than the average density of buildings.
  • Cross-View Regularization: The inter-view module acts as a "soft co-regularizer." It prevents one view from dominating the other and helps transfer knowledge between views—useful when one data source (like POIs) might be sparse in certain areas.
  • Hard Negatives: For the intra-view task, the model doesn't just pick random regions as "negatives." It uses a sampling strategy based on feature distance, forcing the model to work harder to distinguish between regions that look somewhat similar.

Key Results

The model was evaluated on two downstream tasks: Land Usage Clustering (grouping similar neighborhoods) and Region Popularity Prediction (predicting check-in counts).

  • Clustering Breakthrough: ReMVC outperformed seven state-of-the-art baselines (including M2GRL and MVURE). In the Adjusted Rand Index (ARI) metric, it showed a 30% improvement over the previous best method.
  • Clearer Boundaries: Qualitative visualizations of Manhattan showed that ReMVC produced much cleaner, more intuitive clusters of functional districts compared to competing models, which often produced "noisy" or mixed-up results.
  • Robustness: Ablation studies confirmed that removing either the POI view, the Mobility view, or the Inter-view alignment module significantly degraded performance, proving that all components are necessary for a holistic understanding of a city.

Limitations and Future Work

While highly effective, the paper notes a few areas for growth:

  • Domain Adaptation: The current augmentations (adding/dropping POIs) are specific to urban data. To apply this to other domains, new augmentation strategies would need to be designed.
  • Scalability: The study focused on Manhattan (270 regions). Scaling this to a global or even a city-wide level (thousands of regions) would require more computational resources and potentially more complex graph structures.
  • Temporal Dynamics: While mobility data includes time, the model primarily treats it as a static heatmap. Future work could more deeply integrate the evolving nature of cities over months or years.

Summary updated

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.
BibTeX
@article{zhang2023region,
  author = {Liang Zhang and Cheng Long and Gao Cong},
  title = {{Region Embedding With Intra and Inter-View Contrastive Learning}},
  journal = {IEEE Trans. Knowl. Data Eng.},
  year = {2023},
  doi = {10.1109/tkde.2022.3220874},
}

Loading related content...

Figure