LE4Mob: Towards Inductive, Distance-Aware and General-Purpose Location Embedding for Human Mobility Modelling

Xinglei Wang; Stephen Law; Zichao Zeng; Junyuan Liu; Guangsheng Dong; Tao Cheng

cs.LGAug 19, 2026

LE4Mob: Towards Inductive, Distance-Aware and General-Purpose Location Embedding for Human Mobility Modelling

Authors: Xinglei Wang, Stephen Law, Zichao Zeng, Junyuan Liu, Guangsheng Dong, Tao Cheng

Organizations: SpaceTimeLab, Department of Civil, Environmental and Geomatic Engineering, University College London, London, United Kingdom

Abstract

Location representations provide mobility models with fundamental information about the spatial position, functional characteristics, and relationships of places. However, existing embeddings are often dependent on mobility observations, unable to represent unseen locations, and weakly constrained to retain geographic distance. This limits their reuse across datasets and mobility tasks. To address these limitations, we propose LE4Mob, an inductive, distance-aware, and geography-derived location embedding framework for mobility modelling. LE4Mob extends contrastive language-location pre-training while introducing a distance-aware regularisation objective that encourages the embedding space to preserve spatial relationships. Pre-trained from geographic context, LE4Mob can encode rich spatial-semantic information and generate embeddings for unseen locations inductively. Its independence from downstream mobility task supervision also makes it transferable across different mobility tasks. We evaluate LE4Mob on individual-level next location prediction and population-level commuter flow generation. Experiments across multiple datasets and study areas show that LE4Mob outperforms strong baselines, with particular advantages in inductive settings and when downstream models rely directly on interactions between location embeddings. These findings demonstrate the potential of distance-aware, geography-derived location representations as reusable foundations for human mobility modelling.

Figures & tables

Fig. 1: Framework of LE4Mob. (a) In the pre-training phase, batches of POIs are encoded through a dual-stream architecture that aligns coordinate-based location embeddings with POI-derived textual semantics using contrastive learning, while a distance-preserving loss regularises sampled location pairs to retain geographic distance information. (b) After pre-training, the frozen location encoder generates embeddings for downstream mobility tasks, including individual-level next location prediction and population-level flow generation.

	FSQ-NYC	FSQ-TKY	Gowalla-LD	Geolife
#Users	535	1643	68	39
#Days tracked	319	318	367	1703
#Total unique locations	4019	6959	1065	1248
#Stays per user	$178.3\pm 206.1$	$225.1\pm 220.9$	$127.1\pm 123.3$	$358.6\pm 383.6$
#Stays per user per day	$2.1\pm 2.1$	$2.7\pm 2.7$	$2.4\pm 3.0$	$2.4\pm 1.5$
#Unique locations per user	$34.7\pm 21.4$	$56.6\pm 33.5$	$46.3\pm 44.2$	$61.5\pm 52.8$

TABLE I: Basic statistics of the mobility datasets after preprocessing. The mean and standard deviation across users are reported.

		Conventional					Inductive
Data	Model	Acc@1 $\uparrow$	Acc@5 $\uparrow$	nDCG@10 $\uparrow$	GDE@1 $\downarrow$	mGDE@5 $\downarrow$	Acc@1 $\uparrow$	Acc@5 $\uparrow$	nDCG@10 $\uparrow$	GDE@1 $\downarrow$	mGDE@5 $\downarrow$
FSQ-NYC	Vanilla-E2E	19.94 $\pm$ 0.19	46.84 $\pm$ 0.44	37.52 $\pm$ 0.25	4.45 $\pm$ 0.09	5.36 $\pm$ 0.10	9.80 $\pm$ 1.18	24.41 $\pm$ 2.10	19.23 $\pm$ 1.93	7.53 $\pm$ 0.75	8.47 $\pm$ 0.69
	Skip-gram	19.81 $\pm$ 0.34	45.76 $\pm$ 0.19	36.85 $\pm$ 0.29	4.43 $\pm$ 0.06	5.20 $\pm$ 0.03	8.68 $\pm$ 1.19	21.71 $\pm$ 2.92	17.35 $\pm$ 2.40	8.07 $\pm$ 0.82	8.96 $\pm$ 1.15
	POI2Vec	19.62 $\pm$ 0.58	44.94 $\pm$ 0.71	36.32 $\pm$ 0.66	4.53 $\pm$ 0.16	5.38 $\pm$ 0.12	8.84 $\pm$ 1.31	21.80 $\pm$ 3.40	17.37 $\pm$ 2.70	8.26 $\pm$ 0.73	8.74 $\pm$ 0.72
	Geo-Teaser	19.43 $\pm$ 0.73	45.46 $\pm$ 0.84	36.48 $\pm$ 0.59	4.52 $\pm$ 0.08	5.30 $\pm$ 0.08	8.08 $\pm$ 1.47	20.33 $\pm$ 3.13	16.16 $\pm$ 2.74	8.77 $\pm$ 0.48	9.26 $\pm$ 0.61
	TALE	20.20 $\pm$ 0.40	45.80 $\pm$ 0.12	36.98 $\pm$ 0.12	4.46 $\pm$ 0.06	5.35 $\pm$ 0.05	8.52 $\pm$ 1.47	21.04 $\pm$ 3.41	16.83 $\pm$ 2.76	8.23 $\pm$ 0.99	9.07 $\pm$ 0.99

TABLE II: Performance comparison of different embedding methods on the next location prediction task. The best and second-best performance are marked in bold and underlined , respectively, based on the unrounded mean values. For better readability, the ranking metric values are scaled by a factor of

10^{2}

, while GDE and mGDE are reported in kilometres.

		DeepGravity				Bilinear
Data	Model	CPC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	JSD $\downarrow$	CPC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	JSD $\downarrow$
Manchester	Handcrafted	0.765 $\pm$ 0.008	0.693 $\pm$ 0.055	0.871 $\pm$ 0.083	0.202 $\pm$ 0.005	0.757 $\pm$ 0.011	0.717 $\pm$ 0.063	0.915 $\pm$ 0.091	0.207 $\pm$ 0.008
	LDA	0.784 $\pm$ 0.017	0.649 $\pm$ 0.080	0.835 $\pm$ 0.116	0.183 $\pm$ 0.013	0.769 $\pm$ 0.021	0.694 $\pm$ 0.095	0.904 $\pm$ 0.142	0.192 $\pm$ 0.019
	SPPE	0.778 $\pm$ 0.012	0.669 $\pm$ 0.066	0.866 $\pm$ 0.097	0.189 $\pm$ 0.009	0.796 $\pm$ 0.009	0.610 $\pm$ 0.058	0.795 $\pm$ 0.089	0.173 $\pm$ 0.006
	Urban2Vec	0.769 $\pm$ 0.017	0.693 $\pm$ 0.083	0.895 $\pm$ 0.119	0.196 $\pm$ 0.014	0.809 $\pm$ 0.017	0.584 $\pm$ 0.078	0.793 $\pm$ 0.109	0.157 $\pm$ 0.012
	Space2Vec	0.764 $\pm$ 0.016	0.705 $\pm$ 0.080	0.903 $\pm$ 0.116	0.199 $\pm$ 0.013	0.783 $\pm$ 0.022	0.660 $\pm$ 0.095	0.892 $\pm$ 0.140	0.183 $\pm$ 0.019

TABLE III: Performance comparison of different embedding methods on the flow generation task. The best and second-best performances are determined using unrounded mean values and marked in bold and underlined , respectively.

Model	Pearson’s $r$	Spearman’s $\rho$
CaLLiPer	0.268	0.315
LE4Mob	0.369	0.438

TABLE IV: Correspondence between geographic similarity and embedding cosine similarity for OA pairs in the Greater London test set. All correlations are statistically significant at

p<0.0001

Fig. 2: Flow generation performance with and without explicit distance inputs in the downstream DeepGravity model. Error bars indicate standard deviations across repeated runs. Percentages denote the relative CPC change after removing the explicit distance feature. HC, U2V, S2V, and CaL denote Hand-Crafted, Urban2Vec, Space2Vec, and CaLLiPer, respectively.

Geospatial Representation Learning

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

See all 133 papers

Enhancing Human Mobility Prediction with Spatially Aware LLM-based Multi-Agent Systems
Sep 13, 2026Shangyu Lou, Ziqi CuiTrajectory Prediction Spatial Reasoning
University of California & Santa Barbara & San Diego State University California, USA · Politecnico di Milano Milan, Italy
ARMove: Learning to Predict Human Mobility through Agentic Reasoning
Apr 19, 2026Chuyue Wang, Jie Feng, Yuxi Wu +2Transfer Learning Agentic Reasoning
Tencent · Zhongguancun Academy · Alibaba Group +1
Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent
Jun 3, 2026Linyao Chen, Qinlao Zhao, Zechen Li +7LLM Agents Human Behavior Prediction
The University of Tokyo · Huazhong University of Science and Technology · University of New South Wales, Sydney +3

Sep 13, 2026J/K move · Enter open · S save

Enhancing Human Mobility Prediction with Spatially Aware LLM-based Multi-Agent Systems

Shangyu Lou, Ziqi Cui

University of California & Santa Barbara & San Diego State University California, USA · Politecnico di Milano Milan, Italy

Trajectory Prediction Spatial Reasoning Multi-Agent LLM Systems

Open paper AskPDF arXiv

Predicting a user's next POI is a task in human mobility modeling, yet LLM-based approaches focus on semantic reasoning from previous mobility records, while neglecting real-world spatial context. However, human mobility is inherently shaped by spatial cognition, including geographic distance and neighborhood context. This issue is further compounded by prior evidence that LLMs often struggle with spatial reasoning tasks, including distance estimation and geographically biased prediction. To address these limitations, we propose our framework, a multi-agent LLM framework that decomposes next-POI prediction into three stages: Firstly, a Pattern Extraction Agent that captures temporal and categorical mobility patterns from trajectory history; Secondly, a Spatial Reasoning Agent that structures candidate activity choices by combining behavioral preferences with real-world spatial constraints, including geographic distance, road network distance, and neighborhood affiliation; and Thirdly, a Decision Synthesis Agent that integrates behavioral patterns and spatial reasoning for final prediction. Experiments on the NYC benchmark dataset with two LLM backbones show improvements over baseline methods, with up to 493% Hit@1 improvement and 37% relative improvement in Hit@5. Ablations show that combining neighborhood affiliation with distance-based features generally outperforms distance-only settings, and that the Spatial Reasoning Agent plays a crucial role in final prediction by integrating behavioral preferences with real-world spatial constraints, especially for smaller models. Overall, the results highlight the importance of spatial reasoning in mobility prediction. Accurate next-POI prediction requires combining behavioral patterns with explicit real-world spatial constraints, and multi-agent decomposition provides an effective structure for organizing these forms of context.

Abstract

Geospatial Representation Learning

Explore similar work

Enhancing Human Mobility Prediction with Spatially Aware LLM-based Multi-Agent Systems

ARMove: Learning to Predict Human Mobility through Agentic Reasoning

Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent