cs.LGFeb 17, 2026

UrbanVerse: Learning Urban Region Representation Across Cities and Tasks

Authors: Fengze Sun, Egemen Tanin, Shanika Karunasekera, Zuqing Li, Flora D. Salim, Jianzhong Qi

Organizations: School of Computing and Information System, University of Melbourne, Melbourne, Australia · School of Computer Science and Engineering, University of New South Wales, Sydney, Australia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advances in urban region representation learning have enabled a wide range of applications in urban analytics, yet existing methods remain limited in their capabilities to generalize across cities and analytic tasks. We aim to generalize urban representation learning beyond city- and task-specific settings, towards a foundation-style model for urban analytics. To this end, we propose UrbanVerse, a model for cross-city urban representation learning and cross-task urban analytics. For cross-city generalization, UrbanVerse focuses on features local to the target regions and structural features of the nearby regions rather than the entire city. We model regions as nodes on a graph, which enables a random walk-based procedure to form "sequences of regions" that reflect both local and neighborhood structural features for urban region representation learning. For cross-task generalization, we propose a cross-task learning module named HCondDiffCT. This module integrates region-conditioned prior knowledge and task-conditioned semantics into the diffusion process to jointly model multiple downstream urban prediction tasks. HCondDiffCT is generic. It can also be integrated with existing urban representation learning models to enhance their downstream task effectiveness. Experiments on real-world datasets show that UrbanVerse consistently outperforms state-of-the-art methods across six tasks under cross-city settings, achieving up to 35.89% improvements in prediction accuracy.

Figures & tables

Figure 1. Region representation learning frameworks.
Figure 2. UrbanVerse model overview. The model supports cross-city urban representation learning and cross-task urban analytics through two components: (1) CELearning takes a set CC of grid cells and first learns cell embeddings E\mathbf{E} via cell sequences formed by random walks over a graph of cells. It then aggregates cell embeddings to generate region embeddings H\mathbf{H} across multiple cities, facilitating cross-city generalization. (2) HCondDiffCT jointly models multiple tasks within a diffusion process to achieve cross-task generalization.
Figure 3. Prior knowledge generation.
Figure 4. Task-conditioned denoiser.
 ytu=αˉt y0u  +  1−αˉt ϵ  +  (1−αˉt) y~u\boldsymbol{y}_{t}^{u}=\sqrt{\bar{\alpha}_{t}}\,\boldsymbol{y}_{0}^{u}\;+\;\sqrt{1-\bar{\alpha}_{t}}\,\boldsymbol{\epsilon}\;+\;\bigl(1-\sqrt{\bar{\alpha}_{t}}\bigr)\,\widetilde{\boldsymbol{y}}^{u} 
Algorithm 1 HCondDiffCT Training
 y^0u=1αˉt(ytu−(1−αˉt)y~u−1−αˉt ϵ^)\hat{\boldsymbol{y}}_{0}^{u}=\frac{1}{\sqrt{\bar{\alpha}_{t}}}\Bigl(\boldsymbol{y}_{t}^{u}-(1-\sqrt{\bar{\alpha}_{t}})\widetilde{\boldsymbol{y}}^{u}-\sqrt{1-\bar{\alpha}_{t}}\,\hat{\boldsymbol{\epsilon}}\Bigr) 
Algorithm 2 HCondDiffCT Inference
CHI & SF (X) →\rightarrow NYC (Y)HREP ( Zhou et al., 2023 )RegionDCL ( Li et al., 2023 )UrbanCLIP ( Yan et al., 2024 )CityFM ( Balsebre et al., 2024 )GeoHG ( Zou et al., 2025 )GURPP ( Jin et al., 2025 )FlexiReg ( Sun et al., 2025 )UrbanVerseImprovement
R2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrow
Crime0.301  ± \,\pm\, 0.0370.211  ± \,\pm\, 0.0280.241  ± \,\pm\, 0.0050.297  ± \,\pm\, 0.0170.339  ± \,\pm\, 0.0470.545  ± \,\pm\, 0.0390.663  ± \,\pm\, 0.0240.724  ± \,\pm\, 0.008+9.20%
Check-in0.478  ± \,\pm\, 0.0410.434  ± \,\pm\, 0.0140.426  ± \,\pm\, 0.0310.445  ± \,\pm\, 0.0220.329  ± \,\pm\, 0.0580.716  ± \,\pm\, 0.0210.754  ± \,\pm\, 0.0090.781  ± \,\pm\, 0.012+3.58%
Service Call0.173  ± \,\pm\, 0.0310.071  ± \,\pm\, 0.0160.155  ± \,\pm\, 0.0120.124  ± \,\pm\, 0.0110.231  ± \,\pm\, 0.0390.247  ± \,\pm\, 0.0550.513  ± \,\pm\, 0.0310.589  ± \,\pm\, 0.021+14.81%
Population0.262  ± \,\pm\, 0.0440.185  ± \,\pm\, 0.0100.224  ± \,\pm\, 0.0110.219  ± \,\pm\, 0.0090.206  ± \,\pm\, 0.0220.477  ± \,\pm\, 0.0240.569  ± \,\pm\, 0.0180.626  ± \,\pm\, 0.016+10.02%
Carbon0.136  ± \,\pm\, 0.0260.025  ± \,\pm\, 0.0160.051  ± \,\pm\, 0.0330.085  ± \,\pm\, 0.0360.139  ± \,\pm\, 0.0570.009  ± \,\pm\, 0.0240.324  ± \,\pm\, 0.0220.389  ± \,\pm\, 0.027+20.06%
Table 1. Overall Cross-city Prediction Accuracy Results (‘ ↑\uparrow ’ indicates that larger values are preferred. The best results are in boldface, and the second-best results are underlined. Same for the tables below).
CHI & SF (X) →\rightarrow Staten Island (Y)Check-inPopulationCarbonNightlight
R2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )0.148  ± \,\pm\, 0.0310.181  ± \,\pm\, 0.0240.136  ± \,\pm\, 0.0380.226  ± \,\pm\, 0.033
RegionDCL ( Li et al., 2023 )0.098  ± \,\pm\, 0.0270.056  ± \,\pm\, 0.0180.106  ± \,\pm\, 0.0380.394  ± \,\pm\, 0.023
UrbanCLIP ( Yan et al., 2024 )0.061  ± \,\pm\, 0.0190.092  ± \,\pm\, 0.0140.514  ± \,\pm\, 0.0180.577  ± \,\pm\, 0.019
CityFM ( Balsebre et al., 2024 )0.023  ± \,\pm\, 0.0670.073  ± \,\pm\, 0.0170.013  ± \,\pm\, 0.0230.063  ± \,\pm\, 0.026
GeoHG ( Zou et al., 2025 )0.057  ± \,\pm\, 0.0220.027  ± \,\pm\, 0.0150.051  ± \,\pm\, 0.0380.077  ± \,\pm\, 0.027
Table 2. Model Applicability to Suburban Areas.
NYCHREPHREP-DiffCTUrbanCLIPUrbanCLIP-DiffCTHAFusionHAFusion-DiffCTGURPPGURPP-DiffCT
R2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrow
Crime0.681 ±\pm 0.0140.784 ±\pm 0.013 ( +15.1% )0.267 ±\pm 0.0120.471 ±\pm 0.039 ( +76.4% )0.734 ±\pm 0.0150.823 ±\pm 0.011 ( +12.1% )0.589 ±\pm 0.0390.722 ±\pm 0.029 ( +22.6% )
Check-in0.700 ±\pm 0.0220.850 ±\pm 0.016 ( +21.4% )0.458 ±\pm 0.0050.642 ±\pm 0.025 ( +40.2% )0.844 ±\pm 0.0120.900 ±\pm 0.009 ( +6.6% )0.757 ±\pm 0.0190.837 ±\pm 0.017 ( +10.6% )
Service Call0.398 ±\pm 0.0190.589 ±\pm 0.014 ( +49.0% )0.232 ±\pm 0.0050.369 ±\pm 0.033 ( +59.1% )0.493 ±\pm 0.0140.667 ±\pm 0.006 ( +35.3% )0.405 ±\pm 0.0480.528 ±\pm 0.009 ( +30.4% )
Population0.571 ±\pm 0.0210.667 ±\pm 0.019 ( +16.8% )0.276 ±\pm 0.0020.392 ±\pm 0.024 ( +42.0% )0.616 ±\pm 0.0190.692 ±\pm 0.019 ( +12.3% )0.381 ±\pm 0.0260.420 ±\pm 0.021 ( +10.2% )
Carbon0.184 ±\pm 0.0500.373 ±\pm 0.036 ( +102.7% )0.113 ±\pm 0.0130.238 ±\pm 0.033 ( +110.6% )0.189 ±\pm 0.0190.358 ±\pm 0.017 ( +89.4% )0.021 ±\pm 0.0010.204 ±\pm 0.019 ( +871.4% )
Table 3. Prediction Accuracy Results When Powering Existing Models with Our HCondDiffCT Module (NYC).
Figure 5. Ablation study results (NYC).
Figure 6. Adaptability to new downstream tasks.
Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

SymbolDescription
SSA spatial area of interest
RRA set of regions (non-overlapping space partitions)
NNThe number of regions
CCA set of spatial grid cells (basic space partition units)
nnThe number of grid cells
pi\boldsymbol{p}_{i}POI feature of cell cic_{i}
Appendix
Table 4. Frequently Used Symbols.
Figure 7. Grid cell construction.
 NYC ( New York Dataset, 2020 )CHI ( Chicago Dataset, 2020 )SF ( San Francisco Dataset, 2020 )
#regions18077175
#grid cells4387201032
#POIs24,49657,89128,578
#POI categories151515
#crime records35,33518,20048,489
(data collection time)unknown12/2022 - 12/202201/2022 - 12/2022
Appendix
Table 5. Dataset Statistics: New York City, Chicago, and San Francisco.
CHI & SF (X) →\rightarrow NYC (Y)CrimeCheck-inService Call
MAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )91.2  ± \,\pm\, 3.1123.6  ± \,\pm\, 3.40.301  ± \,\pm\, 0.037382.2  ± \,\pm\, 12.5591.3  ± \,\pm\, 16.50.478  ± \,\pm\, 0.0311638  ± \,\pm\, 432493  ± \,\pm\, 520.173  ± \,\pm\, 0.029
RegionDCL ( Li et al., 2023 )100.7  ± \,\pm\, 1.6131.3  ± \,\pm\, 2.40.211  ± \,\pm\, 0.028398.1  ± \,\pm\, 5.5615.3  ± \,\pm\, 9.20.434  ± \,\pm\, 0.0171842  ± \,\pm\, 152643  ± \,\pm\, 200.071  ± \,\pm\, 0.014
UrbanCLIP ( Yan et al., 2024 )98.5  ± \,\pm\, 1.2128.9  ± \,\pm\, 2.30.241  ± \,\pm\, 0.005394.1  ± \,\pm\, 4.8619.9  ± \,\pm\, 6.30.426  ± \,\pm\, 0.0081693  ± \,\pm\, 212520  ± \,\pm\, 320.155  ± \,\pm\, 0.015
CityFM ( Balsebre et al., 2024 )93.5  ± \,\pm\, 1.4124.1  ± \,\pm\, 2.30.297  ± \,\pm\, 0.017388.1  ± \,\pm\, 17.7608.7  ± \,\pm\, 29.80.445  ± \,\pm\, 0.0241778  ± \,\pm\, 92567  ± \,\pm\, 160.124  ± \,\pm\, 0.011
GeoHG ( Zou et al., 2025 )88.6  ± \,\pm\, 3.6120.1  ± \,\pm\, 5.90.339  ± \,\pm\, 0.047420.7  ± \,\pm\, 33.1674.8  ± \,\pm\, 62.30.329  ± \,\pm\, 0.0411587  ± \,\pm\, 432404  ± \,\pm\, 610.231  ± \,\pm\, 0.039
Appendix
Table 6. Overall Cross-city Prediction Accuracy Results on Crime, Check-in, and Service Call Prediction Tasks (‘ ↑\uparrow ’ indicates that large values are preferred. The best results are in boldface, and the second-best results are underlined. Same below).
CHI & SF (X) →\rightarrow NYC (Y)PopulationCarbonNightlight
MAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )3595  ± \,\pm\, 914542  ± \,\pm\, 970.262  ± \,\pm\, 0.034668  ± \,\pm\, 171223  ± \,\pm\, 210.136  ± \,\pm\, 0.02656.6  ± \,\pm\, 0.876.6  ± \,\pm\, 1.10.028  ± \,\pm\, 0.027
RegionDCL ( Li et al., 2023 )3801  ± \,\pm\, 464774  ± \,\pm\, 570.185  ± \,\pm\, 0.012826  ± \,\pm\, 91298  ± \,\pm\, 110.025  ± \,\pm\, 0.01656.5  ± \,\pm\, 0.977.2  ± \,\pm\, 1.10.011  ± \,\pm\, 0.029
UrbanCLIP ( Yan et al., 2024 )3596  ± \,\pm\, 694657  ± \,\pm\, 810.224  ± \,\pm\, 0.014783  ± \,\pm\, 41282  ± \,\pm\, 120.051  ± \,\pm\, 0.00750.8  ± \,\pm\, 0.663.8  ± \,\pm\, 1.60.324  ± \,\pm\, 0.022
CityFM ( Balsebre et al., 2024 )3601  ± \,\pm\, 194672  ± \,\pm\, 270.219  ± \,\pm\, 0.008761  ± \,\pm\, 71258  ± \,\pm\, 110.085  ± \,\pm\, 0.01356.9  ± \,\pm\, 0.476.7  ± \,\pm\, 0.60.026  ± \,\pm\, 0.014
GeoHG ( Zou et al., 2025 )3583  ± \,\pm\, 1494711  ± \,\pm\, 1690.206  ± \,\pm\, 0.041645  ± \,\pm\, 331219  ± \,\pm\, 690.139  ± \,\pm\, 0.04655.3  ± \,\pm\, 0.972.2  ± \,\pm\, 1.10.137  ± \,\pm\, 0.023
Appendix
Table 7. Overall Cross-city Prediction Accuracy Results on Population, Carbon, and Nightlight Prediction Tasks.
NYCCrimeCheck-inService Call
MAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )62.8  ± \,\pm\, 2.183.1  ± \,\pm\, 2.30.681  ± \,\pm\, 0.014276.3  ± \,\pm\, 11.7448.2  ± \,\pm\, 16.90.700  ± \,\pm\, 0.0221430  ± \,\pm\, 292127  ± \,\pm\, 330.398  ± \,\pm\, 0.019
RegionDCL ( Li et al., 2023 )98.7  ± \,\pm\, 0.7127.9  ± \,\pm\, 1.10.251  ± \,\pm\, 0.025371.2  ± \,\pm\, 10.3595.5  ± \,\pm\, 5.90.471  ± \,\pm\, 0.0231783  ± \,\pm\, 222597  ± \,\pm\, 180.103  ± \,\pm\, 0.026
UrbanCLIP ( Yan et al., 2024 )97.4  ± \,\pm\, 2.6126.1  ± \,\pm\, 1.90.267  ± \,\pm\, 0.012393.6  ± \,\pm\, 5.9602.4  ± \,\pm\, 3.10.458  ± \,\pm\, 0.0051409  ± \,\pm\, 72401  ± \,\pm\, 150.232  ± \,\pm\, 0.005
CityFM ( Balsebre et al., 2024 )95.5  ± \,\pm\, 1.4122.4  ± \,\pm\, 1.80.315  ± \,\pm\, 0.010380.2  ± \,\pm\, 3.8594.9  ± \,\pm\, 6.40.471  ± \,\pm\, 0.0111780  ± \,\pm\, 272577  ± \,\pm\, 190.117  ± \,\pm\, 0.013
GeoHG ( Zou et al., 2025 )79.5  ± \,\pm\, 1.5116.9  ± \,\pm\, 2.10.374  ± \,\pm\, 0.019388.1  ± \,\pm\, 11.9621.5  ± \,\pm\, 34.10.422  ± \,\pm\, 0.0321484  ± \,\pm\, 292325  ± \,\pm\, 370.281  ± \,\pm\, 0.023
Appendix
Table 8. Overall Same-city Prediction Accuracy Results on Crime, Check-in, and Service Call Prediction Tasks.
NYCPopulationCarbonNightlight
MAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )2656  ± \,\pm\, 593461  ± \,\pm\, 830.571  ± \,\pm\, 0.021719  ± \,\pm\, 351188  ± \,\pm\, 370.184  ± \,\pm\, 0.02558.7  ± \,\pm\, 1.178.7  ± \,\pm\, 1.30.026  ± \,\pm\, 0.024
RegionDCL ( Li et al., 2023 )3753  ± \,\pm\, 474734  ± \,\pm\, 590.198  ± \,\pm\, 0.019817  ± \,\pm\, 311272  ± \,\pm\, 420.064  ± \,\pm\, 0.03755.8  ± \,\pm\, 0.575.4  ± \,\pm\, 1.00.057  ± \,\pm\, 0.015
UrbanCLIP ( Yan et al., 2024 )3338  ± \,\pm\, 114499  ± \,\pm\, 160.276  ± \,\pm\, 0.002730  ± \,\pm\, 131238  ± \,\pm\, 190.113  ± \,\pm\, 0.01348.9  ± \,\pm\, 0.363.2  ± \,\pm\, 1.00.337  ± \,\pm\, 0.019
CityFM ( Balsebre et al., 2024 )3515  ± \,\pm\, 184545  ± \,\pm\, 260.261  ± \,\pm\, 0.002751  ± \,\pm\, 121253  ± \,\pm\, 210.092  ± \,\pm\, 0.01456.2  ± \,\pm\, 0.475.8  ± \,\pm\, 0.80.047  ± \,\pm\, 0.016
GeoHG ( Zou et al., 2025 )3578  ± \,\pm\, 524685  ± \,\pm\, 420.215  ± \,\pm\, 0.011618  ± \,\pm\, 101236  ± \,\pm\, 160.117  ± \,\pm\, 0.00951.7  ± \,\pm\, 1.070.5  ± \,\pm\, 1.60.176  ± \,\pm\, 0.038
Appendix
Table 9. Overall Same-city Prediction Accuracy Results on Population, Carbon, and Nightlight Prediction Tasks.
Staten IslandCheck-inPopulationCarbonNightlight
R2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrow
HREP ( Zhou et al., 2023 )0.136  ± \,\pm\, 0.0380.192  ± \,\pm\, 0.0280.126  ± \,\pm\, 0.0150.239  ± \,\pm\, 0.021
RegionDCL ( Li et al., 2023 )0.061  ± \,\pm\, 0.0560.028  ± \,\pm\, 0.0210.237  ± \,\pm\, 0.0290.399  ± \,\pm\, 0.021
UrbanCLIP ( Yan et al., 2024 )0.072  ± \,\pm\, 0.0140.088  ± \,\pm\, 0.0130.557  ± \,\pm\, 0.0160.605  ± \,\pm\, 0.015
CityFM ( Balsebre et al., 2024 )0.067  ± \,\pm\, 0.0470.098  ± \,\pm\, 0.0150.079  ± \,\pm\, 0.0150.079  ± \,\pm\, 0.031
GeoHG ( Zou et al., 2025 )0.094  ± \,\pm\, 0.0350.023  ± \,\pm\, 0.0120.047  ± \,\pm\, 0.0280.103  ± \,\pm\, 0.049
Appendix
Table 10. Model Applicability to Suburban Areas under Same-city Setting.
 Embedding LearningDownstream Task
 NYCCHISFNYCCHISF
HREP971029892 (0.003)146 (0.005)91 (0.004)
RegionDCL2,0994751,9250.017 (0.001)0.054 (0.002)0.023 (0.001)
UrbanCLIP5,8174,9114,00186 (0.005)86 (0.005)84 (0.005)
CityFM15,30114,57115,427104 (0.006)102 (0.005)104 (0.006)
GeoHG9010710596 (0.004)95 (0.004)98 (0.004)
Appendix
Table 11. Embedding Learning and Test Times (seconds).
Number of Cells2565121,0242,0484,0968,19216,384
Training time2750971913767521,495
Appendix
Table 12. Impact of the Number of Cells on Embedding Learning Time (seconds, based on SF data).
Figure 8. Predictive distribution of sample regions.
 Singapore ( Singapore Dataset, 2019 )Lisbon ( Lisbon Region Boundary, 2024 )
#regions32453
#spatial partition units748690
#POIs65,08243,961
#POI categories1515
#check-ins355,46324,327
(data collection time)04/2012 - 09/201304/2012 - 09/2013
Appendix
Table 13. Dataset Statistics: Singapore and Lisbon.
 NYC & CHI & SF (X) →\rightarrow Singapore (Y)NYC & CHI & SF (X) →\rightarrow Lisbon (Y)
 Check-inPopulationCheck-inPopulation
 R2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrowR2↑R^{2}\uparrow
FlexiReg0.223 ±\pm 0.0150.518 ±\pm 0.0180.716 ±\pm 0.0140.881 ±\pm 0.005
UrbanVerse0.368 ±\pm 0.0160.625 ±\pm 0.0060.781 ±\pm 0.0110.909 ±\pm 0.010
Improvement+65.0%+20.7%+9.1%+3.2%
Appendix
Table 14. Prediction Accuracy over Cities in Different Countries
#RW\#RW24816
Crime0.721  ± \,\pm\, 0.0210.724  ± \,\pm\, 0.0080.726  ± \,\pm\, 0.0070.711  ± \,\pm\, 0.011
Check-in0.776  ± \,\pm\, 0.0290.781  ± \,\pm\, 0.0110.779  ± \,\pm\, 0.0120.772  ± \,\pm\, 0.009
Service Call0.585  ± \,\pm\, 0.0230.589  ± \,\pm\, 0.0090.584  ± \,\pm\, 0.0110.569  ± \,\pm\, 0.012
Population0.624  ± \,\pm\, 0.0260.626  ± \,\pm\, 0.0060.621  ± \,\pm\, 0.0130.613  ± \,\pm\, 0.005
Carbon0.383  ± \,\pm\, 0.0290.389  ± \,\pm\, 0.0060.390  ± \,\pm\, 0.0110.363  ± \,\pm\, 0.019
Nightlight0.498  ± \,\pm\, 0.0310.492  ± \,\pm\, 0.0130.488  ± \,\pm\, 0.0110.475  ± \,\pm\, 0.009
Appendix
Table 15. Impact of kk on NYC.
#SR\#SR11050100
Crime0.645  ± \,\pm\, 0.0280.724  ± \,\pm\, 0.0080.715  ± \,\pm\, 0.0080.710  ± \,\pm\, 0.006
Check-in0.754  ± \,\pm\, 0.0220.781  ± \,\pm\, 0.0130.776  ± \,\pm\, 0.0130.779  ± \,\pm\, 0.009
Service Call0.538  ± \,\pm\, 0.0390.589  ± \,\pm\, 0.0090.590  ± \,\pm\, 0.0110.583  ± \,\pm\, 0.009
Population0.537  ± \,\pm\, 0.0340.626  ± \,\pm\, 0.0060.621  ± \,\pm\, 0.0090.626  ± \,\pm\, 0.005
Carbon0.401  ± \,\pm\, 0.0510.389  ± \,\pm\, 0.0060.381  ± \,\pm\, 0.0050.385  ± \,\pm\, 0.005
Nightlight0.430  ± \,\pm\, 0.0410.492  ± \,\pm\, 0.0130.502  ± \,\pm\, 0.0100.485  ± \,\pm\, 0.008
Appendix
Table 16. Impact of #SR\#SR ( R2↑R^{2}\uparrow on NYC).
 Crime PredictionCheck-in PredictionService Call Prediction
 MAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrowMAE ↓\downarrowRMSE ↓\downarrowR2↑R^{2}\uparrow
HREP62.8 ±\pm 2.183.1 ±\pm 2.30.681 ±\pm 0.014276.3 ±\pm 11.7448.2 ±\pm 16.90.700 ±\pm 0.0221430 ±\pm 292128 ±\pm 340.398 ±\pm 0.019
HREP-DiffCT50.8 ±\pm 1.668.8 ±\pm 2.10.784 ±\pm 0.013205.9 ±\pm 7.8316.5 ±\pm 17.10.850 ±\pm 0.0161188 ±\pm 311756 ±\pm 330.589 ±\pm 0.014
Improvement+19.1%+17.2%+15.1%+25.5%+29.4%+21.4%+16.9%+17.5%+48.0%
UrbanCLIP97.4 ±\pm 2.6126.1 ±\pm 1.90.267 ±\pm 0.012393.6 ±\pm 5.9602.4 ±\pm 3.10.458 ±\pm 0.0051409 ±\pm 72401 ±\pm 160.232 ±\pm 0.005
UrbanCLIP-DiffCT77.6 ±\pm 3.2107.4 ±\pm 4.20.471 ±\pm 0.039348.1 ±\pm 13.5489.2 ±\pm 17.70.642 ±\pm 0.0251338 ±\pm 522179 ±\pm 640.369 ±\pm 0.033
Appendix
Table 17. Overall Prediction Accuracy Results When Powering Existing Models with Our HCondDiffCT Module.
More in this topic

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 126 papers

Explore similar work

CardsList
May 25, 2026

CITYREP: A Unified Benchmark for Urban Representations Across Cities, Tasks, and Modalities

Urban representation learning encodes complex urban environments into general-purpose embeddings for diverse downstream tasks and emerging urban foundation models. However, current evaluations are limited, typically focusing on one or two cities and tasks and relying on random splits that introduce spatial leakage, leading to inflated performance and weak support for cross-location generalization and fair comparison. To address this, we propose CityRep, a unified benchmark that evaluates urban representations across data modalities, cities, and tasks using spatially structured splits. CityRep consists of three key components: (1) a spatial unit-agnostic evaluation framework that supports heterogeneous urban representations through a standardized alignment module; (2) a unified evaluation protocol using block-based spatial splits to mitigate spatial leakage and enable rigorous model comparison; and (3) an extensible multi-city, multi-task benchmark suite spanning 8 cities and 8 tasks across regression, classification, and distribution prediction. We evaluate 11 representative urban representation models. Results show that performance is highly sensitive to the split protocol, with random splits inflating scores and altering model rankings. We also observe substantial variability across cities and tasks, underscoring the need for generalization-aware evaluation. CityRep is released as a reproducible benchmark with datasets, evaluation pipelines, and diagnostic tools to facilitate fair comparison and support future research in urban representation learning towards urban foundation models.

Junyuan Liu, Xinglei Wang, Zichao Zeng +5

SpaceTimeLab, University College London, UK · 3DIMPact, University College London, UK · School of Resource and Environmental Sciences, Wuhan University, China +1

Representation LearningGeospatial Representation LearningData Contamination
AskPDF
Sep 14, 2026

When Correlations Mislead: Confounder-Aware Multi-View Urban Region Representation Learning

Urban region representation learning commonly combines heterogeneous data sources, such as mobility flows, points of interest, and land-use information, to support tasks including mobility analysis, public safety forecasting, and service demand estimation. Existing multi-view methods typically improve region embeddings by strengthening interactions across views. However, such methods often overlook view-specific regional structures and may propagate correlations induced by shared latent factors, which can reduce the stability of downstream predictions. To overcome this major limitation, we propose CURE, a confounder-aware framework for multi-view urban region representation learning. CURE first encodes each view with its regional graph structure, estimates a shared latent component, and then reduces its projected influence before cross-view interaction. A hierarchical graph-aware fusion module subsequently aggregates the residual view representations using local and global regional contexts Experiments on three real-world cities show that CURE improves predictive performance, remains robust under missing and noisy input views, and provides reliable cross-view integration through shared component separation and context-dependent view weighting.

Sean Bin Yang, Ying Sun, Zongyi Xu +6

Aalborg University, Aalborg, Denmark · Chongqing University of Posts and Telecommunications, Chongqing, China · East China Normal University, Shanghai, China

Causal Representation LearningGraph Representation LearningGeospatial Representation LearningMulti-View Learning
AskPDF
Aug 21, 2026

CoST: Semantic-Aware Urban Understanding via Spatial-Temporal Alignment

Geospatial representation learning from satellite imagery is a fundamental problem for large-scale urban analysis and real-world applications. Despite recent advances, current methods struggle with cross-region generalization and semantic interpretability due to their reliance on region-specific auxiliary data and the neglect of semantic alignment within multi-temporal urban imagery. Therefore, we present CoST, a novel Contrastive-based Spatial-Temporal framework that aligns spatial context with multi-temporal semantics to extract universal geographic regularities shared across regions. Specifically, CoST explicitly models spatial correlations to capture transferable geographic structures and exploits multi-year urban change semantics to align learned representations with high-level geo-semantics. Extensive experiments demonstrate that CoST consistently achieves superior performance across various downstream tasks and in unseen scenario, yielding an average relative gain of 8.7% over the strongest competing methods across eight city-indicator settings. The code is available in this repo.

Yutian Jiang, Jiabo Liu, Xixuan Hao +1

The Hong Kong University of Science and Technology (Guangzhou)

Geospatial Representation LearningRemote Sensing Image UnderstandingSpatial GeneralizationSpatial Context ModelingContrastive Learning
AskPDF