Research map / 3D vision / 3D generation
Urban Region Embedding via Multi-View Contrastive Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ReCP, a multi-view contrastive prediction model that learns consistent urban region embeddings from POI and taxi mobility views via intra-view contrastive/reconstruction losses and inter-view contrastive and dual prediction objectives.
Problem
Urban region representation learning from multiple views (POI attributes and human mobility) typically fuses views only after separate encoding, failing to learn consistent representations across views.
Approach
Replaces posterior attention-based fusion with a consistency learning paradigm: maximize mutual information between attribute and mobility view representations (contrastive) and minimize conditional entropy via dual prediction networks, combined with intra-view contrastive learning and autoencoder reconstruction to avoid trivial solutions.
Key results
- ReCP achieves NMI 0.780, ARI 0.483, F-measure 0.499 on land usage clustering, versus best baseline ReMVC/MGFN (NMI 0.761 for ReMVC)
- On popularity prediction ReCP reports MAE 195.16 and R^2 0.652, improving over HREP by 10.28% MAE and 11.45% R^2
- Ablation: removing inter-view learning (w/o IV) lowers ARI by 29.84% and R^2 by 4.00% relative to full ReCP
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReCP instead of ReMVC | Urban region embeddings · Manhattan land use clustering | NMI: 0.761* ± 0.02 -> 0.780 ± 0.01 | Helped |
| Use ReCP instead of ReMVC | Urban region embeddings · Manhattan land use clustering | ARI: 0.455* ± 0.04 -> 0.483 ± 0.01 | Helped |
| Use ReCP instead of ReMVC | Urban region embeddings · Manhattan land use clustering | F-measure: 0.462* ± 0.04 -> 0.499 ± 0.02 | Helped |
| Use ReCP instead of HREP | Urban region embeddings · Manhattan region popularity prediction | MAE: 217.52* ± 10.98 -> 195.16 ± 18.70 | Helped |
| Use ReCP instead of HREP | Urban region embeddings · Manhattan region popularity prediction | RMSE: 318.41* ± 14.54 -> 291.19 ± 20.04 | Helped |
| Use ReCP instead of HREP | Urban region embeddings · Manhattan region popularity prediction | R²: 0.585* ± 0.04 -> 0.652 ± 0.05 | Helped |
| Add intra-view contrastive learning (ablation) | ReCP · Land use clustering and region popularity prediction | Performance | Helped |
| Add intra-view reconstruction loss (ablation) | ReCP · Land use clustering and region popularity prediction | Performance | Helped |
| Add inter-view learning (ablation) | ReCP · Manhattan land use clustering | ARI: 29.84% | Helped |
| Add inter-view learning (ablation) | ReCP · Manhattan region popularity prediction | R²: 4.00% | Helped |
| Add inter-view dual prediction loss (ablation) | ReCP · Land use clustering and region popularity prediction | Performance | Helped |
Limitations
- Evaluated only on a single city (Manhattan) with two views (POI and mobility)
- Relies on grid-search hyperparameter tuning with a fairly large set of hyperparameters (mu, alpha, lambda1, lambda2)
- Dual prediction may collapse without the intra-view reconstruction loss
Experimental setup
Datasets: NYC Open Data POI data NYC taxi trip records (Feb 2014) NYC check-in data
Benchmarks: Manhattan land usage clustering (29 districts) Manhattan region popularity prediction (check-in counts)
Metrics: NMI ARI F-measure MAE RMSE R^2
Baselines: HDGE ZE-Mob MV-PN CGAL MVURE MGFN ReMVC HREP
Builds on: ReMVC contrastive learning information bottleneck-style mutual information maximization
Closest papers by meaning
- Cross-City Latent Space Alignment for Consistency Region Embedding (ICML 2025 · similarity 0.86)
- Improving Region Representation Learning from Urban Imagery with Noisy Long-Caption Supervision (AAAI 2026 · similarity 0.85)
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.84)
- UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction (AAAI 2025 · similarity 0.84)
- UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.83)
- Road Network Representation Learning with the Third Law of Geography (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.