Research map / Graph learning / Graph self-supervised learning
Road Network Representation Learning with the Third Law of Geography
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 Garner, a graph contrastive learning framework for road network representation that incorporates the Third Law of Geography via street-view-image-based geographic configuration graph augmentation, spectral negative sampling, and a dual contrastive objective fusing it with the First Law.
Problem
Existing road network representation learning methods overemphasize spatial proximity (First Law of Geography) and ignore geographic configuration similarity, which matters for long-range relationships between road segments.
Approach
Connects road segments with similar street-view-image-derived geographic configurations via a kNN similarity graph, encodes them with SGC so that similar configurations yield similar representations, uses a spectral sparsest-cut-inspired negative sample (a d-regular approximation of the complete graph) to push dissimilar configurations apart, and fuses with a diffusion-based spatial-proximity view
Key results
- On road function prediction in Singapore, Garner reaches 81.40% Micro-F1 and 62.45% Macro-F1 versus 66.61% and 30.67% for the best baseline MVGRL.
- On road traffic inference in Singapore, Garner reports MAE 2.80 versus 3.04 for the best baseline MVGRL.
- On visual road retrieval in Singapore, Garner achieves Recall@10 of 0.4600 versus 0.0143 for the best baseline SARN; baselines perform near random.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add street view images as geographic-configuration inputs (ablation) | Garner · Singapore road function prediction | Macro-F1 (%): 30.67 ± 0.66 -> 50.21 ± 0.60 | Helped |
| Add street view images as geographic-configuration inputs (ablation) | Garner · NYC visual road retrieval | Recall@10: 0.0324 -> 0.2803 | Helped |
| Add geographic-configuration-aware graph augmentation (ablation) | Garner · Singapore road function prediction | Macro-F1 (%): 50.21 ± 0.60 -> 60.57 ± 0.68 | Helped |
| Add geographic-configuration-aware graph augmentation (ablation) | Garner · Singapore visual road retrieval | Recall@10: 0.2002 -> 0.3426 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · Singapore road function prediction | Macro-F1 (%): 60.57 ± 0.68 -> 62.45 ± 0.64 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · Singapore visual road retrieval | Recall@10: 0.3426 -> 0.4600 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · Singapore road traffic inference | MAPE: 0.606 ± 0.040 -> 0.579 ± 0.030 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · NYC road function prediction | Macro-F1 (%): 45.78 ± 0.52 -> 47.22 ± 0.42 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · NYC road traffic inference | MAPE: 0.210 ± 0.002 -> 0.207 ± 0.002 | Helped |
| Use spectral negative sampling instead of feature-shuffling-only negatives (ablation) | Garner · NYC visual road retrieval | Recall@10: 0.4776 -> 0.5531 | Helped |
| Use Garner instead of MVGRL | road function prediction · Singapore | Macro-F1 (%): 30.67 ± 0.66 -> 62.45 ± 0.64 | Helped |
| Use Garner instead of MVGRL | road function prediction · NYC | Macro-F1 (%): 17.39 ± 0.23 -> 47.22 ± 0.42 | Helped |
Surprising findings
- Most baselines perform near random guessing on visual road retrieval, while the street-view-image-based method achieves decent results.
- Several baselines (RFN, SRN2Vec) run out of memory on the NYC dataset.
Limitations
- Method's assumptions may fail where the First or Third Law does not apply, e.g., very large areas or limited data.
- Requires street view images, which may not be available for all cities; roads without SVIs are assigned the average configuration.
- Road function labels come from OSM land-use data available only in some cities.
Experimental setup
Datasets: OpenStreetMap road networks (Singapore, NYC) Google Maps street view images (Singapore, NYC)
Benchmarks: Road function prediction (Singapore, NYC) Road traffic inference (Singapore, NYC) Visual road retrieval (Singapore, NYC)
Metrics: Micro-F1 Macro-F1 AUROC MAE RMSE MAPE Recall@10 MRR
Baselines: Deepwalk MVGRL CCA-SSG GGD RFN SRN2Vec SARN
Builds on: MVGRL Simple Graph Convolution Deep Graph Infomax Tobler's First Law of Geography
Closest papers by meaning
- UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.88)
- Patched Line Segment Learning for Vector Road Mapping (AAAI 2024 · similarity 0.86)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.85)
- RoadSceneBench: A Lightweight Benchmark for Mid-Level Road Scene Understanding (CVPR 2026 · similarity 0.84)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.84)
- Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning (AAAI 2025 · similarity 0.84)
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.