Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 89%
Last extracted: 8/31/2026, 3:29:48 AM
Summary
This paper empirically evaluates cross-city Point-of-Interest (POI) recommendation using the large-scale Trip World benchmark. It identifies three main bottlenecks in state-of-the-art methods: (1) hometown-aware models rely on destination-region popularity priors rather than transferring user-specific preferences due to low home-destination region overlap; (2) accuracy-efficiency trade-offs degrade at scale, with simpler models like MatTrip outperforming complex ones like SPOT-Trip; and (3) semantic metadata integration mechanisms yield little benefit. A diagnostic pilot on LLM-based agentic methods shows they also underperform simple popularity priors, highlighting the need for task-specific designs for preference transfer, semantic grounding, and scalable reasoning.
Entities (13)
Relation Signals (7)
Trip World → usedforevaluation → Cross-city POI recommendation
confidence 98% · Using the recently proposed large-scale benchmark Trip World, we empirically re-examine whether conclusions drawn on small prior benchmarks still hold
Trip World → hashighscale → Cross-city POI recommendation
confidence 90% · Trip World is an order of magnitude larger than previous benchmarks
MatTrip → outperforms → SPOT-Trip
confidence 90% · MatTrip, the simplest model... attains the best accuracy, while the most compositional model, SPOT-Trip... ranks lowest
LLMMove → underperforms → Popularity prior
confidence 90% · Every agentic variant trails the simple in-candidate Popularity prior by a wide margin
SPOT-Trip → relieson → destination-region priors
confidence 85% · hometown-aware models appear to rely more on destination-region priors than on user-specific preference transfer
AgentMove → underperforms → LLMMove
confidence 85% · the more elaborate AgentMove underperforms the lighter LLMMove prompt
Semantic metadata → yieldslittlebenefit → SPOT-Trip
confidence 80% · existing mechanisms for integrating semantic metadata yield little benefit... SPOT-Trip no-review outperforms SPOT-Trip full
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2608.27840v1
- Canonical: https://arxiv.org/abs/2608.27840v1
Trouble viewing inline? Open PDF directly →
Full Text
30,348 characters extracted from source content.
Expand or collapse full text
An Empirical Evaluation of Cross-City POI Recommendation on a Large-Scale Benchmark 411Conference: The 34th ACM International Conference on Advances in Geographic Information Systems; November 3–6, 2026; Riverside, CA, USAThe 34th ACM International Conference on Advances in Geographic Information Systems (SIGSPATIAL ’26), November 3–6, 2026, Riverside, CA, USACCS: Information systems Recommender systems Peibo Li Affiliation: University of New South Wales, Sydney, Australia email: peibo.li@student.unsw.edu.au , Yang Song Affiliation: University of New South Wales, Sydney, Australia email: yang.song1@unsw.edu.au , Hao Xue Affiliation: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China email: haoxue@hkust-gz.edu.cn , Maarten de Rijke Affiliation: University of Amsterdam, The Netherlands email: m.derijke@uva.nl and Flora D. Salim Affiliation: University of New South Wales, Sydney, Australia email: flora.salim@unsw.edu.au 2026 Abstract. Cross-city POI (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 (CRUISE Research Group, 2026), 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. Keywords: Point-of-interest recommendation 1. Introduction Point-of-interest (POI) recommendation underpins location-based services, helping users discover relevant places (Zhao et al., 2016; Zhang et al., 2025). Most work targets the in-city or next-POI setting, where users repeatedly interact with POIs in their home region (Wang et al., 2017; Li and Gong, 2020). Many real travel scenarios break this assumption: a traveler arriving in an unfamiliar city has no prior interactions with the destination POIs, only a short and sparse trajectory, and preferences that shift from everyday routines to tourism. Cross-city POI recommendation therefore requires transferring user preferences from a home city to a destination with a largely unseen POI inventory (Ding et al., 2019; Li and Gong, 2020; Sun et al., 2024), while contending with habitual histories that do not translate to travel behavior, geographically and semantically disjoint POIs, and limited sequential evidence (Xin et al., 2022; Liu et al., 2025). Progress on this setting has been limited by the lack of large-scale, realistic, and semantically rich benchmarks. Existing datasets are designed for single-city or in-city next-POI prediction (Chen et al., 2016; Thomee et al., 2016), or are small, geographically narrow, and metadata-poor (Yang et al., 2015; Yang et al., 2016; Wang et al., 2017). Some provide only coordinates and coarse categories, while others offer rich reviews but rely on review-derived interactions that need not correspond to real visit sequences. These limitations make it hard to evaluate whether a method generalizes across cities, exploits semantic POI information, or handles the extreme cold start travelers face. We conduct our evaluation on Trip World (CRUISE Research Group, 2026), a recently proposed large-scale benchmark that provides worldwide out-of-town check-ins together with enriched POI metadata and user reviews, to rigorously study cross-city POI recommendation. Its scale and coverage support a unified comparison of classical, neural, and recent agentic methods, and enable testing whether models leverage mobility history, destination context, POI semantics, and travel-specific behavior beyond simple proximity or popularity signals. Our contributions are: • We evaluate representative cross-city POI recommendation methods on Trip World, spanning popularity-based, sequential, neural, semantic-aware, and LLM (LLM)-agentic approaches. • We surface three bottlenecks: hometown-aware methods appear to rely more on destination-region priors than on transferable user preferences; the simplest model is among the strongest while the most compositional ranks lowest at much higher cost; and existing semantic-integration mechanisms yield little benefit. • As a diagnostic pilot, we benchmark agentic methods adapted from next-POI recommendation and show that naive transfer is insufficient, charting directions for cross-city-native designs. 2. Related Work 2.1. Trip and Cross-City POI Recommendation Intra-city trip recommendation predicts a sequence of intermediate POIs given an origin, destination, and budget within one region. Learning-based methods evolved from combinatorial optimization to deep sequence models: MatTrip (Zhang et al., 2024) (a dual-LSTM encoder-decoder), Graph-Trip (Gao et al., 2023) (heterogeneous POI graphs), AR-Trip (Shu et al., 2024) (a Transformer mitigating repetition), and PPROC (Iakovlev and Lähdesmäki, 2025) (neural spatiotemporal point processes). All rely on rich localized history and degrade under the cold start of out-of-town travel. Cross-city (out-of-town) recommenders transfer preferences from a data-rich hometown to an unfamiliar target city across disjoint POI spaces. Early generative and content-aware models share a city-agnostic latent space (Yin et al., 2014; Gao et al., 2015; Yin et al., 2017); graph and causal approaches disentangle hometown preference from destination constraints (Xin et al., 2021; Liu et al., 2024), and meta-learning transfers preference initializations across cities (Wang et al., 2019; Chen et al., 2021; Ding et al., 2022). Most target isolated next-POI prediction; SPOT-Trip (Liu et al., 2025) instead unifies trip modeling with cross-city transfer via a POI attribute knowledge graph and neural ODEs for preference drift, reaching state of the art in data-scarce regions. 2.2. Datasets for Trip and Cross-City Recommendation Table 1 summarizes six widely used datasets. Edinburgh, Glasgow, Osaka, and Toronto are reconstructed from Flickr photos (Thomee et al., 2016; Chen et al., 2016) and each cover a single region with only hometown check-ins. Only Foursquare (Yang et al., 2015; Yang et al., 2016) and Yelp11 1 https://w.yelp.com/dataset contain genuine cross-city activity, so we adopt them for comparison; both are small and largely US-centric. Foursquare carries no auxiliary POI information, whereas Yelp offers rich reviews but its check-ins are review artifacts rather than real visits, with roughly 50% of out-of-town behavior being “phantom travel” within the same metropolitan area. These gaps motivate a large-scale, worldwide, and semantically rich benchmark such as Trip World (CRUISE Research Group, 2026), which we adopt for our evaluation. 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 Regions POIs CIs Home OOT Real CI Edinburgh 1,4541,454 11 2828 33,94433,944 33,94433,944 00 ✗ Glasgow 601601 11 2727 11,43411,434 11,43411,434 00 ✗ Osaka 450450 11 2727 7,7477,747 7,7477,747 00 ✗ Toronto 1,3951,395 11 2929 39,41939,419 39,41939,419 00 ✗ Foursquare 3,0073,007 2121 23,88423,884 126,219126,219 109,225109,225 16,99416,994 ✓ Yelp 4,4174,417 214214 29,93029,930 78,88278,882 58,40358,403 20,47920,479 ✗ 3. Experiments Table 2. Statistics of Trip World after processing. CIs: check-ins; Avg |co||c_o|: mean number of POIs per out-of-town trip. Dataset #Users #POIs #Regions #Home CIs #Travel CIs Avg |co||c_o| Trip World 63,896 336,102 890 6,242,544 576,504 4.97 3.1. Experimental setup 3.1.1. Task Given a user’s hometown check-in history HuH_u and an out-of-town trip specified by its origin o, destination d, target region r, and length L, the task is to predict the ordered intermediate POIs co[1:−1]c_o[1:-1] visited between o and d. All models observe HuH_u, o, d, r, and L; semantic-aware variants additionally observe POI categories, coordinates, and review-derived signals. Evaluation is over the intermediate POIs only. 3.1.2. Baselines We benchmark 7 baselines: 4 trip recommendation methods without hometown information (Popularity (Chen et al., 2016), GraphTrip (Gao et al., 2023), MatTrip (Zhang et al., 2024), AR-Trip (Shu et al., 2024)) and 3 with hometown information (KDDC (Liu et al., 2024), PPROC (Iakovlev and Lähdesmäki, 2024), SPOT-Trip (Liu et al., 2025)), where KDDC and PPROC use the SPOT-Trip backbone. The Popularity baseline ranks POIs by their global training-set visit frequency. All experiments run on a cluster of H200 GPUs. 3.1.3. Evaluation metric We adopt F1 and Pairs-F1, following existing work (Gao et al., 2023; Liu et al., 2025; Shu et al., 2024; Kuo et al., 2023): F1 measures the overlap between predicted and ground-truth intermediate POIs, while Pairs-F1 additionally rewards correct visiting order. Both are computed over the intermediate POIs only, excluding the fixed origin and destination. 3.1.4. Dataset processing We follow (Liu et al., 2025) to process the dataset. We filter out users with fewer than three out-of-town check-in records, and exclude out-of-town trips whose travel durations are shorter than 1 hour or longer than 30 days. We partition by user into training, validation, and testing sets with an 80%/10%/10%80\%/10\%/10\% ratio. For baselines that require knowledge graphs, we generate entity-specific relations following the original papers, using auxiliary information from the dataset. Statistics after processing are shown in Table 2. 3.2. Results and Discussion Table 3. Out-of-town trip recommendation performance, measured by F1 and Pairs-F1 over the intermediate POIs (higher is better, ↑ ). 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 ≤0.003≤0.003 for F1 and ≤0.0013≤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 (—). Foursquare Yelp Trip World Architecture F1 ↑ Pairs-F1 ↑ F1 ↑ Pairs-F1 ↑ F1 ↑ Pairs-F1 ↑ Popularity (Chen et al., 2016) 0.0261 0.0013 0.0257 0.0056 0.0222 ± 0.0012 0.0057 ± 0.0005 MatTrip (Zhang et al., 2024) 0.0311 0.0037 0.0301 0.0119 0.0601 ± 0.0019 0.0166 ± 0.0010 GraphTrip (Gao et al., 2023) 0.0295 0.0048 0.0289 0.0126 0.0582 ± 0.0017 0.0164 ± 0.0005 AR-Trip (Shu et al., 2024) 0.0304 0.0045 0.0307 0.0153 0.0562 ± 0.0028 0.0151 ± 0.0013 KDDC (Liu et al., 2024) 0.0375 0.0079 0.0341 0.0156 0.0575 ± 0.0005 0.0159 ± 0.0001 PPROC (Iakovlev and Lähdesmäki, 2024) 0.0330 0.0071 0.0334 0.0159 0.0554 ± 0.0008 0.0148 ± 0.0008 SPOT-Trip no-review (Liu et al., 2025) — — — — 0.0572 ± 0.0004 0.0159 ± 0.0006 SPOT-Trip full (Liu et al., 2025) 0.0400 0.0109 0.0399 0.0190 0.0557 ± 0.0013 0.0151 ± 0.0010 We report baseline performance in Table 3. Using the published Foursquare/Yelp results as a contextual reference rather than a controlled re-run in our pipeline, we observe that the relative ranking on Trip World differs: there, the simplest model (MatTrip) is among the strongest while the most compositional, hometown-aware SPOT-Trip ranks lowest. We examine possible explanations below. 3.2.1. Transferable preferences, or region priors? On Trip World, methods that use hometown information do not outperform those relying solely on out-of-town signals, unlike on Foursquare and Yelp (Liu et al., 2025). The standard framing casts these methods as learning user preferences from home history and transferring them to the destination. However, as shown in Table 4, only 30.9% of destination regions in Trip World also appear as home regions for other users, far below Foursquare and Yelp. This pattern suggests that, in this low-overlap regime, the methods rely more on region-conditional POI distributions (effectively per-region popularity priors) than on user preferences that transfer across regions. We test this directly on SPOT-Trip, the most complete hometown-aware model (KDDC and PPROC share its backbone): at test time we replace each trip’s hometown history with that of a random other user, holding the trained model and decoder fixed. Across three seeds, this shuffling leaves F1 unchanged (relative change <0.55%<0.55\%, within the cross-seed standard deviation), so the model is invariant to whose home it sees and does not exploit user-specific hometown preference. We benchmark the most complete model; extending the ablation to every architecture is left to future work. Table 4. Region overlap between home regions (ℛhR_h) and destination regions (ℛoR_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. Dataset Dest-coverage |ℛh∩ℛo|/|ℛo||R_h _o|/|R_o| Dest-only regions Foursquare 95.2% 1 Yelp 69.9% 56 Trip World 30.9% 603 3.2.2. Accuracy–efficiency trade-off at scale. Trip World is an order of magnitude larger than previous benchmarks, raising the question of how the baselines behave at this scale. MatTrip, the simplest model (a Bi-LSTM with Bahdanau attention), attains the best accuracy, while the most compositional model, SPOT-Trip (a knowledge graph with an ODE decoder), ranks lowest while being the costliest to train (Table 5). This is consistent with the additional inductive biases adding variance rather than reducing it in a regime where the per-POI gradient signal is sparse (336,102336,102 POIs across 890890 regions). We cannot fully separate architecture from tuning or implementation here, and leave a controlled scaling study (e.g., training-size sweeps) to future work; nonetheless, the architectures that pay the most overhead also score lowest. Table 5. Per-epoch training time on Trip World (single H200 GPU). Architecture s / epoch Relative to MatTrip MatTrip 58 1.0× GraphTrip 178 3.1× KDDC 137 2.4× AR-Trip 304 5.2× PPROC 330 5.7× SPOT-Trip full 422 7.3× 3.2.3. Integrating semantic metadata Trip World provides rich semantic features. We probe one common integration mechanism: SPOT-Trip full adds a knowledge-graph relation encoding each POI’s average rating, while SPOT-Trip no-review omits it. Counterintuitively, SPOT-Trip no-review outperforms SPOT-Trip full, indicating that this particular mechanism, a single review-derived rating relation, does not help. We stress that this probes one mechanism rather than semantic metadata in general; the richer review text and user-taste signals are examined in our agentic pilot (Section 3.3). Existing methods were not designed for such features, as they were unavailable in prior benchmarks, and we hope Trip World’s semantics motivate models that can harness them. 3.3. Can Agentic Methods Transfer to Cross-City Recommendation? As a diagnostic pilot, we test whether LLM-based agentic methods, which reason over textual context and natively consume semantic metadata, alleviate the bottlenecks above. We adapt two representative methods from next-POI prediction: LLMMove (Feng et al., 2024), a direct LLM ranking prompt, and AgentMove (Feng et al., 2025), a full agent with memory and world-model reasoning, both backed by Claude Haiku 4.5 (claude-haiku-4-5). Adaptation and setup. Open-vocabulary generation over the 336,102336,102-POI inventory is infeasible for an LLM, a first finding in itself. We thus recast the task as in-candidate selection: each query receives a menu of 100 candidates guaranteed to contain the ground-truth intermediates, which the agent selects and orders. Negatives are sampled to approximately match ground-truth popularity (residual GT-vs-negative popularity AUC ≈0.55≈0.55), so popularity cannot trivially separate the answer; the in-candidate Popularity baseline ranks candidates by leave-one-out training frequency. Because this restriction raises the chance level, absolute F1 here is not comparable to Table 3; the fair references are the in-candidate Popularity and Random baselines. Beyond trajectory context, we supply two review-grounded signals. The review summary is a short LLM-written description of what a POI is known for and whom it suits, distilled from its reviews and attached to every candidate. The user-taste profile is a concise LLM-written description of a user’s preferred cuisines, venue types, vibe, and price level, distilled from the review summaries of the POIs the user visited at home, and prepended to the prompt. 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. Method F1 ↑ Pairs-F1 ↑ 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 Naive transfer underperforms a popularity prior. Table 6 reports the in-candidate results on data-rich destinations. Every agentic variant trails the simple in-candidate Popularity prior by a wide margin (e.g., 0.07180.0718 vs. 0.10690.1069 F1). Echoing the scaling result for neural baselines, the more elaborate AgentMove underperforms the lighter LLMMove prompt, indicating that additional agentic machinery again adds variance rather than signal. Strikingly, injecting the candidates’ popularity into the prompt (LLMMove, + popularity) helps more than any semantic signal, yet still does not close the gap to the popularity baseline. Naive transfer of agentic methods is therefore insufficient for cross-city POI recommendation. Table 7. Disentangling the two semantic signals on cold-start destinations (≤30≤ 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. Variant F1 ↑ Pairs-F1 ↑ 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 Semantics are not exploited, but the signal is present. To probe the semantic bottleneck, we disentangle the user-taste and review signals on the cold-start slice (destinations with ≤30≤ 30 training trips), where transfer matters most. Table 7 shows that both signals independently hurt: removing either improves F1, and the variant with no semantics is the strongest LLM configuration. This is not a data-quality artifact, cold-start POI summaries are in fact well sourced (a median of 342 reviews per summarized POI). Rather, the agent cannot convert the metadata into correct selections and is instead distracted away from the popularity prior. Crucially, the information is present: the ground-truth intermediates are measurably aligned with the user’s hometown taste, matching the user’s home POI categories 40.6%40.6\% of the time versus 25.3%25.3\% for the popularity-matched negatives. The limitation is therefore one of method, not of available signal. Where the headroom lies. The dominance of popularity is itself partly an artifact of the benchmark and metric. Popularity still wins overall only because the candidate matching is approximate and most queries have popular ground truth: stratifying by ground-truth popularity, the prior is strong where the answer is itself popular, but on the long-tail third of queries it falls to near the Random floor, where the LLM agents exceed it. This is where semantic reasoning should help, pointing to directions for cross-city-native methods: (i) anchor on the popularity prior and use semantics only to refine among comparable candidates, rather than letting free-form generation override a reliable prior; (i) align POI semantics to the user-taste vocabulary, since current review summaries describe service sentiment largely orthogonal to the preferences that distinguish the ground truth; and (i) develop scalable retrieval and grounding to relax the in-candidate restriction toward open-vocabulary cross-city recommendation. These mirror the three bottlenecks found for neural baselines, preference transfer, semantic grounding, and scale. 4. Conclusion Using Trip World, a large-scale, worldwide, and semantically rich benchmark, we re-examined state-of-the-art methods for cross-city POI recommendation. Our evaluation surfaces three bottlenecks. First, under Trip World’s low home–destination region overlap, hometown-aware methods rely on per-region popularity priors rather than transferable user preferences: replacing a user’s hometown history with a random other user’s leaves SPOT-Trip—the most complete such model—essentially unchanged, confirming directly that it does not exploit user-specific home signal. Second, their accuracy–efficiency trade-off degrades at this scale: the simplest model (MatTrip) is among the strongest, while the most compositional (SPOT-Trip) ranks lowest at the highest cost. Third, a common mechanism for integrating semantic metadata yields no benefit. A diagnostic pilot on agentic LLM methods fares no better: they trail a simple in-candidate popularity prior, even though the relevant semantic signal is present in the data. Overall, progress will require task-specific designs for cross-city preference transfer, semantic grounding, and scalable reasoning over unseen destination inventories. References Chen et al. (2016) D. Chen, C. S. Ong, and L. Xie Learning points and routes to recommend trajectories. In Proceedings of the 25th ACM international on conference on information and knowledge management, p. 2227–2232. Cited by: §1, §2.2, §3.1.2, Table 3. Chen et al. (2021) Y. Chen, X. Wang, M. Fan, J. Huang, S. Yang, and W. Zhu Curriculum meta-learning for next POI recommendation. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD ’21, p. 2692–2702. Cited by: §2.1. CRUISE Research Group (2026) CRUISE Research Group TripWorld: a large-scale benchmark for cross-city POI recommendation. Note: Hugging Face datasethttps://huggingface.co/datasets/CRUISEResearchGroup/TripWorld Cited by: §1, §2.2, Abstract.. Ding et al. (2019) J. Ding, G. Yu, Y. Li, D. Jin, and H. Gao Learning from hometown and current city: cross-city poi recommendation via interest drift and transfer learning. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3 (4), p. 1–28. Cited by: §1. Ding et al. (2022) Y. Ding X. Wang et al. Cross-city recommendation via spatio-social meta-learning. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’22, p. 284–293. Cited by: §2.1. Feng et al. (2025) J. Feng, Y. Du, J. Zhao, and Y. Li Agentmove: a large language model based agentic framework for zero-shot next location prediction. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 1322–1338. Cited by: §3.3, Table 6, Table 6. Feng et al. (2024) S. Feng, H. Lyu, F. Li, Z. Sun, and C. Chen Where to move next: zero-shot generalization of llms for next poi recommendation. In 2024 ieee conference on artificial intelligence (cai), p. 1530–1535. Cited by: §3.3, Table 6, Table 6. Gao et al. (2015) H. Gao, J. Tang, X. Hu, and H. Liu Content-aware point of interest recommendation on location-based social networks. In Proceedings of the 29th AAAI Conference on Artificial Intelligence, AAAI ’15, p. 172–178. Cited by: §2.1. Gao et al. (2023) Q. Gao, W. Wang, L. Huang, X. Yang, T. Li, and H. Fujita Dual-grained human mobility learning for location-aware trip recommendation with spatial-temporal graph knowledge fusion. Information Fusion 92, p. 46–63. Cited by: §2.1, §3.1.2, §3.1.3, Table 3. Iakovlev and Lähdesmäki (2025) V. Iakovlev and H. Lähdesmäki Learning spatiotemporal dynamical systems from point process observations. In Proceedings of the 13th International Conference on Learning Representations, ICLR ’25. Cited by: §2.1. Iakovlev and Lähdesmäki (2024) V. Iakovlev and H. Lähdesmäki Learning spatiotemporal dynamical systems from point process observations. arXiv preprint arXiv:2406.00368. Cited by: §3.1.2, Table 3. Kuo et al. (2023) A. Kuo, H. Chen, and W. Ku BERT-trip: effective and scalable trip representation using attentive contrast learning. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), p. 612–623. Cited by: §3.1.3. Li and Gong (2020) D. Li and Z. Gong A deep neural network for crossing-city poi recommendations. IEEE Transactions on Knowledge and Data Engineering 34 (8), p. 3536–3548. Cited by: §1. Liu et al. (2025) Y. Liu, H. Miao, G. Shen, Y. Zhao, X. Kong, and I. Lee SPOT-trip: dual-preference driven out-of-town trip recommendation. arXiv preprint arXiv:2506.01705. Cited by: §1, §2.1, §3.1.2, §3.1.3, §3.1.4, §3.2.1, Table 3, Table 3, Table 3, Table 3. Liu et al. (2024) Y. Liu, G. Shen, C. Cui, Z. Zhao, X. Han, J. Du, X. Zhao, and X. Kong KDDC: knowledge-driven disentangled causal metric learning for pre-travel out-of-town recommendation. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence, IJCAI ’24, p. 4–9. Cited by: §2.1, §3.1.2, Table 3. Shu et al. (2024) W. Shu, K. Xu, W. Tai, T. Zhong, Y. Wang, and F. Zhou Analyzing and mitigating repetitions in trip recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’24, p. 2276–2280. Cited by: §2.1, §3.1.2, §3.1.3, Table 3. Sun et al. (2024) K. Sun, C. Li, and T. Qian City matters! a dual-target cross-city sequential poi recommendation model. ACM Transactions on Information Systems 42 (6), p. 1–27. Cited by: §1. Thomee et al. (2016) B. Thomee, D. A. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li YFCC100M: the new data in multimedia research. Communications of the ACM 59 (2), p. 64–73. Cited by: §1, §2.2. Wang et al. (2017) H. Wang, Y. Fu, Q. Wang, H. Yin, C. Du, and H. Xiong A location-sentiment-aware recommender system for both home-town and out-of-town users. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, p. 1135–1143. Cited by: §1, §1. Wang et al. (2019) H. Wang, G. Tang, X. Liu, and Z. Wang Cross-city transfer learning for deep spatio-temporal prediction. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, IJCAI ’19, p. 3821–3827. Cited by: §2.1. Xin et al. (2021) H. Xin, X. Lu, T. Xu, H. Liu, J. Gu, D. Dou, and H. Xiong Out-of-town recommendation with travel intention modeling. In Proceedings of the 35th AAAI Conference on Artificial Intelligence, AAAI ’21, p. 4490–4497. Cited by: §2.1. Xin et al. (2022) H. Xin, X. Lu, N. Zhu, T. Xu, D. Dou, and H. Xiong Captor: a crowd-aware pre-travel recommender system for out-of-town users. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 1174–1184. Cited by: §1. Yang et al. (2015) D. Yang, D. Zhang, L. Chen, and B. Qu NationTelescope: monitoring and visualizing large-scale collective behavior in lbsns. Journal of Network and Computer Applications 55, p. 170–180. Cited by: §1, §2.2. Yang et al. (2016) D. Yang, D. Zhang, and B. Qu Participatory cultural mapping based on collective behavior data in location-based social networks. ACM Transactions on Intelligent Systems and Technology (TIST) 7 (3), p. 1–23. Cited by: §1, §2.2. Yin et al. (2014) H. Yin, B. Cui, X. Zhou, W. Wang, Z. Huang, and S. Sadiq Joint modeling of user check-in behaviors for out-of-town recommendation. In Proceedings of the 23rd ACM International Conference on Information and Knowledge Management, CIKM ’14, p. 361–370. Cited by: §2.1. Yin et al. (2017) H. Yin, W. Wang, H. Wang, L. Chen, and X. Zhou Spatial-aware hierarchical collaborative deep learning for poi recommendation. IEEE Transactions on Knowledge and Data Engineering 29 (11), p. 2537–2551. Cited by: §2.1. Zhang et al. (2024) J. Zhang, M. Ma, X. Gao, and G. Chen Encoder-decoder based route generation model for flexible travel recommendation. IEEE Transactions on Services Computing. Cited by: §2.1, §3.1.2, Table 3. Zhang et al. (2025) Q. Zhang, P. Yang, J. Yu, H. Wang, X. He, S. Yiu, and H. Yin A survey on point-of-interest recommendation: models, architectures, and security. IEEE Transactions on Knowledge and Data Engineering. Cited by: §1. Zhao et al. (2016) S. Zhao, I. King, and M. R. Lyu A survey of point-of-interest recommendation in location-based social networks. arXiv preprint arXiv:1607.00647. Cited by: §1.