An Empirical Evaluation of Cross-City POI Recommendation on a Large-Scale Benchmark

Peibo Li; Yang Song; Hao Xue; Maarten de Rijke; Flora D. Salim

cs.AIAug 28, 2026

An Empirical Evaluation of Cross-City POI Recommendation on a Large-Scale Benchmark

Authors: Peibo Li, Yang Song, Hao Xue, Maarten de Rijke, Flora D. Salim

Organizations: University of New South Wales Sydney, Australia · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · University of Amsterdam The Netherlands

Abstract

Cross-city point-of-interest (POI) recommendation is crucial for navigating unfamiliar urban environments, yet its progress has historically been constrained by data limitations. Using the recently proposed large-scale benchmark Trip World, we empirically re-examine whether conclusions drawn on small prior benchmarks still hold under worldwide coverage, low home-destination region overlap, and large, semantically rich POI inventories. Our evaluation surfaces three bottlenecks of representative state-of-the-art methods: (1) hometown-aware models appear to rely more on destination-region priors than on user-specific preference transfer; (2) their accuracy-efficiency trade-off degrades at this scale, where the simplest model is among the strongest; and (3) existing mechanisms for integrating semantic metadata yield little benefit. We further include a diagnostic pilot on agentic methods adapted from next-POI recommendation, finding that naive adaptation trails a simple popularity prior even though the relevant semantic signal is present in the data. These results highlight the need for task-specific designs that support cross-city preference transfer, semantic grounding, and scalable reasoning over unseen destination inventories.

Figures & tables

Dataset	Users	Regions	POIs	CIs	Home	OOT	Real CI
Edinburgh	$1{,}454$	$1$	$28$	$33{,}944$	$33{,}944$	$0$	✗
Glasgow	$601$	$1$	$27$	$11{,}434$	$11{,}434$	$0$	✗
Osaka	$450$	$1$	$27$	$7{,}747$	$7{,}747$	$0$	✗
Toronto	$1{,}395$	$1$	$29$	$39{,}419$	$39{,}419$	$0$	✗
Foursquare	$3{,}007$	$21$	$23{,}884$	$126{,}219$	$109{,}225$	$16{,}994$	✓
Yelp	$4{,}417$	$214$	$29{,}930$	$78{,}882$	$58{,}403$	$20{,}479$	✗

Table 1. Statistics of prior datasets. CIs denote check-ins; Home and OOT are hometown and out-of-town check-ins. Real CI indicates whether the dataset contains genuine check-in events (✓) versus reconstructed or proxy records (✗). The four city-specific datasets cover a single region and contain only hometown check-ins by construction.

Dataset	#Users	#POIs	#Regions	#Home CIs	#Travel CIs	Avg $\|c_{o}\|$
Trip World	63,896	336,102	890	6,242,544	576,504	4.97

Table 2. Statistics of Trip World after processing. CIs: check-ins; Avg

|c_{o}|

: mean number of POIs per out-of-town trip.

	Foursquare		Yelp		Trip World
Architecture	F1 $\uparrow$	Pairs-F1 $\uparrow$	F1 $\uparrow$	Pairs-F1 $\uparrow$	F1 $\uparrow$	Pairs-F1 $\uparrow$
Popularity ( Chen et al., 2016 )	0.0261	0.0013	0.0257	0.0056	0.0222 $\pm$ 0.0012	0.0057 $\pm$ 0.0005
MatTrip ( Zhang et al., 2024 )	0.0311	0.0037	0.0301	0.0119	0.0601 $\pm$ 0.0019	0.0166 $\pm$ 0.0010
GraphTrip ( Gao et al., 2023 )	0.0295	0.0048	0.0289	0.0126	0.0582 $\pm$ 0.0017	0.0164 $\pm$ 0.0005
AR-Trip ( Shu et al., 2024 )	0.0304	0.0045	0.0307	0.0153	0.0562 $\pm$ 0.0028	0.0151 $\pm$ 0.0013
KDDC ( Liu et al., 2024 )	0.0375	0.0079	0.0341	0.0156	0.0575 $\pm$ 0.0005	0.0159 $\pm$ 0.0001

Table 3. Out-of-town trip recommendation performance, measured by F1 and Pairs-F1 over the intermediate POIs (higher is better,

\uparrow

). Foursquare and Yelp numbers are reproduced from Liu et al. ( Liu et al., 2025 ) (single published values), not re-run in our pipeline; Trip World numbers are averaged over 3 random seeds (per-seed std

{\leq}0.003

for F1 and

{\leq}0.0013

for Pairs-F1, so the neural baselines differ by within roughly one standard deviation). Bold marks the best value per column. SPOT-Trip no-review is a Trip-World-specific ablation that removes the review-rating relation and has no counterpart in the original benchmark (—).

Dataset	Dest-coverage $\|\mathcal{R}_{h}\cap\mathcal{R}_{o}\|/\|\mathcal{R}_{o}\|$	Dest-only regions
Foursquare	95.2%	1
Yelp	69.9%	56
Trip World	30.9%	603

Table 4. Region overlap between home regions (

\mathcal{R}_{h}

) and destination regions (

\mathcal{R}_{o}

) on the three benchmarks. Dest-coverage is the fraction of destination regions that also appear as a home region. Dest-only regions count regions that appear only as destinations and never as homes.

Architecture	s / epoch	Relative to MatTrip
MatTrip	58	1.0 $\times$
GraphTrip	178	3.1 $\times$
KDDC	137	2.4 $\times$
AR-Trip	304	5.2 $\times$
PPROC	330	5.7 $\times$
SPOT-Trip full	422	7.3 $\times$

Table 5. Per-epoch training time on Trip World (single H200 GPU).

Method	F1 $\uparrow$	Pairs-F1 $\uparrow$
Random (in-candidate)	0.0361	0.0040
AgentMove ( Feng et al., 2025 ) , no semantics	0.0627	0.0102
AgentMove ( Feng et al., 2025 ) , + semantics	0.0658	0.0116
LLMMove ( Feng et al., 2024 ) , + semantics	0.0718	0.0146
LLMMove ( Feng et al., 2024 ) , + popularity	0.0820	0.0189
Popularity (in-candidate)	0.1069	0.0234

Table 6. Agentic methods on cross-city POI recommendation ( in-candidate setting, data-rich destinations, Claude Haiku 4.5). Values are not comparable to Table 3 ; the references are the in-candidate Popularity and Random baselines. Bold marks the best value per column.

Variant	F1 $\uparrow$	Pairs-F1 $\uparrow$
Random (in-candidate)	0.0324	0.0089
+ taste + review	0.0490	0.0083
+ taste $-$ review	0.0536	0.0097
$-$ taste + review	0.0524	0.0125
$-$ taste $-$ review	0.0591	0.0144
Popularity (in-candidate)	0.0686	0.0176

Table 7. Disentangling the two semantic signals on cold-start destinations (

\leq 30

training trips), LLMMove with Claude Haiku 4.5 (in-candidate). Removing either signal improves F1; the variant with no semantics is best among the LLM , and all trail the Popularity prior.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 521 papers

Category-Anchored Large Language Models for Next POI Recommendation
Oct 29, 2026Lexi Bai, Xiaopeng Zeng, Xun Zhou +2Sequential Recommendation Large Language Model-Based Recommendation
Harbin Institute of Technology, Shenzhen, China · City University of Hong Kong, Hong Kong SAR, China.
CaST-POI: Candidate-Conditioned Spatiotemporal Modeling for Next POI Recommendation
Feb 10, 2026Zhenyu Yu, Chunlei Meng, Yangchen Zeng +3Sequential Recommendation Recommender Systems
College of Computer Science and Artificial Intelligence Fudan University Shanghai, China · College of Intelligent Robotics and Advanced Manufacturing Fudan University Shanghai, China · School of Cyber Science and Engineering Southeast University Nanjing, China +2
Why Users Go There: World Knowledge-Augmented Generative Next POI Recommendation
May 12, 2026Qiuyu Ding, Heng-Da Xu, Wei Zhang +4Sequential Recommendation Knowledge Augmentation for Language Models
Amap, Alibaba Group · Xi’an Jiaotong University

Oct 29, 2026J/K move · Enter open · S save

Category-Anchored Large Language Models for Next POI Recommendation

Lexi Bai, Xiaopeng Zeng, Xun Zhou, Yuanshao Zhu+1

Harbin Institute of Technology, Shenzhen, China · City University of Hong Kong, Hong Kong SAR, China.

Sequential Recommendation Large Language Model-Based Recommendation LLM Grounding

Open paper AskPDF arXiv

Next point-of-interest (POI) recommendation aims to predict a user's next destination from spatio-temporal check-in data, and serves as the foundation for many location-based services. Recent advances incorporate large language models (LLMs) into this task by textualizing user trajectories to exploit their semantic reasoning ability. However, directly generating POI identifiers with such models is unreliable: POIs are sparse, localized, and long-tailed entities with weak natural-language semantics, which creates a semantic-to-entity gap between high-level mobility intent and concrete POI prediction. To bridge this gap, we propose ANCHOR, a category-anchored framework for semantic-to-entity grounding in next POI recommendation. Specifically, ANCHOR first estimates category-level mobility intent in a compact and stable semantic space, rather than predicting concrete POI entities directly. The inferred intent is then grounded into a candidate set by integrating three complementary sources of mobility evidence: category anchors, historical visit memory, and user preferences. Finally, a history-aware reranker selects the target POI by matching each candidate with the user's historical mobility patterns and current context. Experiments on three real-world check-in datasets show that ANCHOR consistently outperforms representative traditional, structure-enhanced, and LLM-based baselines on the reported metrics. Further analyses validate the effectiveness of semantic-to-entity grounding and show favorable trade-offs among candidate coverage, ranking quality, and inference efficiency.

Abstract

Recommender Systems

Explore similar work

Category-Anchored Large Language Models for Next POI Recommendation

CaST-POI: Candidate-Conditioned Spatiotemporal Modeling for Next POI Recommendation

Why Users Go There: World Knowledge-Augmented Generative Next POI Recommendation