EASTAR: Event-Anchored Semantic and Task-Adaptive Region Representation Learning
Authors: Peng Liang, Genan Dai, Bowen Zhang
Organizations: College of Applied Science, Shenzhen University, Shenzhen, China · School of Artificial Intelligence, Shenzhen Technology University, Shenzhen, China
Urban region representation learning supports a wide range of urban analytics tasks by encoding regional functions and interactions, with human mobility serving as a central signal because it directly captures inter-region interactions. However, most mobility-based methods learn from aggregated origin-destination matrices, where trips driven by different activity purposes are mixed into purpose-agnostic interaction patterns. Although static urban attributes such as POIs and land use can enrich region-level semantics, they do not explicitly ground mobility interactions in activity-purpose signals. As a result, downstream adaptation is often performed over entangled embeddings, making it difficult to select task-relevant functional semantics. To address these limitations, we propose EASTAR, an event-anchored and task-adaptive framework for urban region representation learning. EASTAR treats offline activity events as sparse semantic anchors and uses them to factorize aggregated mobility representations into multiple purpose-related semantic channels. It further introduces free channels to capture latent mobility patterns not covered by event data and a global channel to preserve overall spatial structure. To adapt the learned channel space to different downstream objectives, EASTAR designs a task-adaptive prompt router that initializes task-channel relevance from LLM-derived priors and refines the routing weights through downstream supervision. Experiments on two real-world urban datasets across four prediction tasks show that EASTAR consistently outperforms representative baselines, achieving up to 39.15% reduction in MAE. Ablation and interpretability analyses further demonstrate the effectiveness of event anchoring, semantic channel construction, and task-adaptive routing.
TABLE I: Activity Categories Grouped into Semantic Anchors
Fig. 1: Overview of EASTAR. Stage 1 learns event-anchored multi-channel region representations through semantic completion and channel decomposition, while Stage 2 adaptively routes the learned channels to downstream tasks.
Dataset
NYC
CHI
Source
Regions
180
77
Open Portal [ 41 , 42 ]
POIs
24,496
57,891
OpenStreetMap [ 38 ]
POI categories
26
26
OpenStreetMap [ 38 ]
Land-use categories
11
12
OpenStreetMap [ 38 ]
Taxi trips
10,953,879
3,381,807
Open Portal [ 41 , 42 ]
Satellite images
180
77
Google Maps [ 43 ]
TABLE II: Dataset Statistics
New York City
Crime
Check-in
Service Call
Population
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
R2↑
MVURE [ 3 ]
67.9 ± 1.1
0 93.8 ± 1.9
0.591 ± 0.016
306.7 ± 0 8.2
499.6 ± 12.9
0.627 ± 0.019
1428 ± 33
2180 ± 0 46
0.367 ± 0.027
2814 ± 67
3564 ± 0 79
0.545 ± 0.008
MGFN [ 7 ]
70.2 ± 2.3
0 89.6 ± 2.5
0.630 ± 0.020
292.6 ± 17.1
451.8 ± 28.1
0.690 ± 0.040
1554 ± 81
2286 ± 115
0.303 ± 0.069
2822 ± 42
3706 ± 0 98
0.509 ± 0.026
ReCP [ 31 ]
83.1 ± 2.4
108.7 ± 2.2
0.459 ± 0.022
246.9 ± 0 4.3
400.1 ± 23.6
0.761 ± 0.029
1516 ± 66
2199 ± 0 49
0.356 ± 0.029
3305 ± 90
4353 ± 153
0.322 ± 0.047
HAFusion [ 5 ]
56.1 ± 1.3
0 76.1 ± 2.2
0.734 ± 0.015
202.8 ± 0 7.2
322.8 ± 12.6
0.844 ± 0.012
1273 ± 20
1951 ± 0 27
0.493 ± 0.014
2497 ± 50
3277 ± 0 82
0.616 ± 0.019
ComSRE [ 32 ]
72.3 ± 1.1
0 99.6 ± 6.5
0.545 ± 0.059
251.7 ± 17.1
434.8 ± 44.4
0.716 ± 0.057
1459 ± 69
2254 ± 0 98
0.324 ± 0.058
2964 ± 95
3916 ± 100
0.452 ± 0.028
TABLE III: Performance comparison on four downstream tasks. (Best: Bold , Second: Underline . ∗ indicates a statistically significant improvement over the best baseline ( p<0.05 ). The ‘Improvement’ is calculated against the best baseline.)
Stage →
Embedding Learning
Downstream Task
Method ↓
NYC
CHI
NYC
CHI
MVURE
12
5
0.02 (0.0005)
0.04 (0.0010)
MGFN
35
42
0.01 (0.0006)
0.04 (0.0011)
HAFusion
26
16
0.02 (0.0005)
0.01 (0.0005)
ComSRE
14
11
0.02 (0.0006)
0.01 (0.0003)
UrbanCLIP
506
874
59 (0.0028)
64 (0.0032)
TABLE IV: Embedding Learning and Testing Times (seconds)
Fig. 2: Ablation study results ( R2 on NYC).
Variant
Crime
Check-in
Service Call
Population
HAFusion
0.734
0.844
0.493
0.616
HAFusion M
0.744
0.857
0.505
0.620
MGFN
0.630
0.690
0.303
0.509
MGFN M
0.632
0.652
0.407
0.488
HREP
0.681
0.700
0.398
0.571
HREP M
0.660
0.667
0.368
0.536
TABLE V: Impact of incorporating Meetup semantics ( R2↑ on NYC)
Fig. 3: Comparison between initial LLM-derived priors and the final learned routing weights across downstream tasks.
Variant
Crime
Check-in
Service Call
Population
EASTAR-POI
0.778
0.874
0.508
0.584
EASTAR-Random
0.731
0.768
0.435
0.519
EASTAR
0.847
0.949
0.658
0.744
TABLE VI: Impact of semantic probe sources ( R2↑ on NYC)
Fig. 4: Region-level semantic consistency on NYC. Activations of the Tech and Social channels from EASTAR and EASTAR-w/o-CC compared with the corresponding ground-truth event distributions.
School of Computing and Information System, University of Melbourne, Melbourne, Australia · School of Computer Science and Engineering, University of New South Wales, Sydney, Australia
Urban AI Institute, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea · Graduate School of Data Science, KAIST, Daejeon, Republic of Korea · Department of Industrial Engineering and Systems, KAIST, Daejeon, Republic of Korea +1
Aalborg University, Aalborg, Denmark · Chongqing University of Posts and Telecommunications, Chongqing, China · East China Normal University, Shanghai, China