Accurate prediction of the next point of interest (POI) that a user will visit is important to location-based services for personalized and context-aware recommendation. Traditional approaches, particularly those based on collaborative filtering, struggle to fully utilize the semantic, temporal, and spatial information embedded in user behavior data. Although the emergence of large language models (LLMs) can capture semantic information, large-scale LLMs are infeasible for large-scale data due to computational constraints, as our experiment shows. In this paper, we propose an eficient and lightweight POI recommendation framework, KP4POI, that leverages small-scale LLM enhanced with knowledge prompting, which encodes knowledge graphs of venue and user into natural language. By transforming user and venue knowledge graphs into natural language prompts, our method enables LLMs to incorporate semantic, geographic, and social information efectively. Experiments on real-world datasets demonstrate that our approach improves recommendation quality while maintaining scalability.
The proliferation of location-based services has precipitated an increasing demand for systems capable of
recommending the next Point of Interest (POI) that is relevant and useful to the user, conditioned on their
spatio-temporal check-in history. Such predictive capabilities are pivotal not only for improving
usercentric experiences, but also for optimizing marketing strategies through context-aware personalization
[
Subsequently, many hybrid methods have been proposed that incorporate categorical, temporal, or
contextual information [
With the rise of large language models (LLMs), new avenues for context-aware recommendation have emerged [14, 15, 16]. Two main types of LLM-based POI recommendation methods have gained attention: prompt-based methods and embedding-based methods, as shown in Figs. 1 (a) and (b). Prompt-based methods convert recommendation tasks into natural language format and use compact LLMs to generate candidates [17]. These methods are computationally eficient but often lack the ability to integrate structured knowledge such as venue categories or spatial relationships. Embedding-based methods generate semantic user embeddings from user check-in history and POI embeddings from POI attributes by using large-scale LLMs [18, 19, 20, 21], and calculate the similarity between them at Workshop on Recommenders in Tourism (RecTour 2025), September 22, 2025, co-located with the 19th ACM Conference on
CEUR
ceur-ws.org the cost of heavy fine-tuning and inference. Although expressive, these methods are computationally demanding and impractical for large-scale datasets [21, 22].
To address this challenge, we designed a framework, KP4POI, that enriches prompt-based methods with structured knowledge through knowledge prompting (KP) [23, 24], as shown in Fig. 1 (c). Specifically, we transform venue and user knowledge graphs (KGs) into natural language prompts, allowing compact LLMs to efectively utilize not only venue attributes (e.g., category, location) but also user context (e.g., social connections) during prediction. In contrast to embedding-based approaches, our method achieves eficiency by leveraging compact LLMs, while the integration of KGs can help mitigate hallucinations in LLM predictions and improve accuracy [25]. Moreover, because our method uses structured knowledge, such as venue categories, addresses, or user communities, to substitute for raw location coordinates, it partially supports privacy-aware settings where fine-grained location data may not be available [ 26]. In such scenarios, user-level knowledge, such as social groupings, can further complement missing venue information, enabling robust recommendations even under limited or anonymized metadata.
In this paper, we compare our KP4POI with embedding-based methods in terms of fine-tuning eficiency on large-scale POI datasets. We examine the contribution of diferent types of knowledge—specifically, categorical afinity and geographic proximity—and assess the utility of user KP derived from social graphs [27]. Our findings demonstrate that compact LLMs augmented with structured knowledge ofer a practical and scalable solution for personalized POI recommendation.
As discussed in Section 1, LLM-empowered POI recommendation methods can be categorized into prompt-based and embedding-based approaches. One of the most important prompt-based methods is P5. The P5 uses a prompt to reformulate the recommendation task as a natural language task by encoding the IDs of the recommendation items [17] with the masked personalized prompt (MPP) [17, 23]. Fig. 2 shows an example of MPP templates, where denotes a set of prompt templates to encode historical check-ins as natural language. The MPP (, ; ) is constructed for the user based on the IDs of previously visited venues = { 1, … , }. The next POI target is masked using multiple templates , where the wider variety of templates is important [28]. Based on this prompt, an LLM can predict the next POI ID, denoted +1 , that the user may be interested in. Combining the venue knowledge prompt with the standard MPP in the following final input: (, ) = [ ] (, ; )[ ].
(1) In Eq. (1), the special token [SP] serves as a delimiter separating the user’s check-in history from the appended knowledge prompts. To minimize vocabulary size and improve generalization, eficient tokenization represents IDs split into two-digit segments, e.g., 1298 as <12><98>. Since the output in P5 is ID, high generative capacity is not required and performance and speed are balanced using a compact LLM such as T5 [29]. Advanced recommendation methods such as Prompt Distillation for Eficient LLM-based Recommendation (POD) [ 28] and KP [23] are based on P5. POD distills discrete using diferent types of templates . prompt templates into task-specific continuous prompt vectors, thus improving the training eficiency of LLM-based recommendation systems. Functionally, POD and P5 difer only in that POD uses multiple prompt templates [28].
Instead of prompting, embedding-based methods directly use LLM embeddings to make recommendations [21]. However, embedding-based methods that use large-scale LLMs demand a high computational load because the entire check-in history must be inputted to LLM for inference as a natural language prompt. Both fine-tuning and inference require high computational load. Section 4.2.1 compares the ifne-tuning time of two types of methods.
Building on the motivation outlined in Section 1, we present KP4POI, a prompt-based POI recommendation framework enhanced by structured KP. KP4POI leverages knowledge about the venue and user in a natural language format to guide LLMs for efective recommendation. This section details the architecture of KP4POI and describes how KPs are generated and integrated into the LLM pipeline.
The knowledge about the venues can be represented as KGs. Public datasets used in the experiment may provide venue information, such as category or latitude / longitude (lat/lon), which can be converted to addresses. For POI recommendation, categorical afinity and geographical proximity play an important role [30]. Categorical afinity [ 31, 32] refers to co-occurrence of categories (e.g., visiting a train station before an ofice), while geographical proximity [ 33, 34] involves nearby locations (e.g., dining at a restaurant close to the ofice). This knowledge can be concatenated into MPPs in the form of a KP. To measure geometrical proximity, to address privacy concerns, reverse-geocoded address features are aggregated at the street level (NYC) or chome (block) (Tokyo), ensuring that no fine-grained and personally identifiable location information is used. At inference time, these features are retrieved from venue metadata and do not require real-time or private user data.
KGs consist of triples (head / relation / tail) represented as (ℎ, , ) . To incorporate KG knowledge into LLM, a KP transforms a KG triple using a relation-specific template , as illustrated in Fig. 3. A KG = ( , category, ) is converted using the relation-related template () = “The category of is ” when = category. Let be the set of KGs related to the venue . A venue KP can be generated from the KGs of the venues visited as follows: ( , ) = ⋃∈ ⋃∈ (), where ⋃ denotes the concatenation of prompts. The final prompt combining venue
For example, for a user who visited 1 ( 1 = ( 1, category, ) ) then 2 ( 2 = ( 2, category, ) )), the final prompt becomes: “[ ] [ ]
The category of 1 is . The category of 2 is . [ ] ”.
User-specific information has been shown to improve the performance of recommendation [ 27]. Attributes such as gender and age can ofer valuable personalization signals. However, such personal data are rarely available in POI recommendation due to heightened privacy concerns, especially because high-resolution location data can inadvertently expose sensitive details such as a user’s residence or workplace. In contrast, location-based social networks like Foursquare often include user social graphs, which are less sensitive and are more readily available. These social graphs can implicitly capture behavioral traits and allow for unsupervised grouping of users with similar characteristics. When a social graph is available, user clustering can be efective for POI recommendation [ 35].
To take advantage of this, we apply graph partitioning methods [36] to identify user communities in the social graph. Each user is assigned a community that represents interpretable group-level characteristics. Given that the POI recommendation benefits from approaches using graph neural networks [7, 8] or matrix factorization [9, 10, 11], we derive community labels using node embeddings [37] using node2vec, and cluster them [38] by k-means++ clustering [39]. Each cluster ID serves as a communitylevel pseudo-attribute label that represents the user’s social context, where each user is assigned to exactly one cluster. These community labels can be encoded as KG triples (, community, ) . We can convert them into natural language prompts using a relation-specific template as in Fig. 4, and incorporate them into the model via user KP. This allows the compact LLM to benefit from social information even in privacy-aware settings where fine-grained personal metadata are not accessible 1. and the final prompt combining user knowledge is given by:
Let be the set of KG triples related to user . Then, a user KP is (, ) = ⋃∈ () (2) (3) (4) In KP4POI, the loss function maximizes the likelihood of the next POI +1 in the model as (, , , ) = (, , ) (,
)[ ].
=1 ℒ = − ∑ ∑ log ( +1 | (, , , )).
For POI recommendation tasks, LLM-based methods have generally been efective [ 21]. Among those utilizing large-scale LLM such as Llama, SeCor and soft prompting [ 40] demonstrated the highest performance. Both SeCor and our method are fine-tuned from a pre-trained Llama2-7B and T5-small, respectively. SeCor requires an additional step of CF training. However, as shown in Table 2, the use of 1Note that our approach derives community labels by clustering user embeddings obtained from the social graph. Therefore, user KP is applicable only to datasets that include explicit social relations (WWW2019 and Gowalla). For datasets without social graphs (NYC and TKY), user KP cannot be constructed. large-scale LLMs is impractical for large-scale datasets2. Except for methods using large-scale LLMs, the POD and collaborative filter-based seq2graph [ 41] performed comparable.
We compare our method 3 with P54 and SeCor, as they represent two contrasting baselines: P5 as a prompt-based approach and SeCor as an embedding-based approach using large-scale LLM. We primarily compare with P5 because it represents a compact, prompt-based LLM recommender that is closest in spirit to our framework. While traditional baselines such as popularity or matrix factorization would also be informative, we leave their integration for future work since our focus here is to evaluate the eficiency and efectiveness of knowledge prompting relative to existing LLM-based approaches.
We used T5-small (60.5M parameters) as the backbone, which is much smaller than Llama-7B. The model used 512 dimensions and eight multihead attention. AdamW was used with a maximum learning rate of 0.001, batch size 64, and 100 epochs. The best model was selected on the basis of the validation performance. Following previous work [21, 42], the check-ins were sorted chronologically by user. Users with fewer than 10 check-ins and venues visited only once or twice were removed. Three types of venue KG were used: address, lat/lon5, and category such as restaurant. KP was implemented based on KP4SR6. The evaluation used NDCG@ and Recall@ for = {5, 10, 50} in a leave-one-out setting7. The last two venues were reserved for testing and predicted by models fine-tuned on the remaining data. We used a paired t-test with Bonferroni correction to assess the significance between P5 and our KP4POI.
Table 1 shows the statistics of four Foursquare datasets8 (NYC, TKY, TIST2015, and WWW2019) and one Gowalla dataset9, where Foursquare and Gowalla are location-based social networks with user check-ins. NYC and TKY are small, commonly used, and suitable for parametric studies. TIST2015 and WWW2019 are used to test scalability, because they are large-scale. Among the four datasets, only WWW2019 and Gowalla contain explicit social graphs. Since our method requires social graphs to derive community labels via clustering, user KP is only available for these two datasets
The NYC and TKY dataset [43] contains check-ins in New York City (NYC) and Tokyo (TKY), where 2Although we used the oficial implementation ( https://github.com/siri-ya/SeCor), we could not reproduce the results reported in [21]. 3https://github.com/DensoITLab/KP4POI 4P5 and POD are functionally equivalent, except that POD uses multiple prompt templates [28], while P5 in our implementation employs 11 templates to ensure comparability. 5In [21], lat/lon was converted to address. We validated the efectiveness of this. 6https://github.com/zhaijianyang/KP4SR 7We focus on accuracy metrics (NDCG, Recall) as in prior work, leaving diversity/coverage for future work. 8https://sites.google.com/site/yangdingqi/home/foursquare-dataset 9https://snap.stanford.edu/data/loc-gowalla.html
Recall
NDCG
Recall Method P5(POD) 0.2383 0.2437 0.3084 0.3250 0.2135 0.2243 0.2861 0.3188 KP4POI(add) 0.2395 0.2431 0.3121 0.3232 0.2162 0.2287† 0.2904† 0.3280** KP4POI(lat/lon) 0.2385 0.2436 0.3102 0.3259 0.2175 0.2287† 0.2883† 0.3227* KP4POI(add+lat/lon) 0.2399 0.2451 0.3121 0.3278 0.2144 0.2256 0.2887 0.3227 KP4POI(cat) 0.2457* 0.2516* 0.3075 0.3250 0.2197* 0.2307* 0.2909** 0.3245** KP4POI(cat+add) 0.2521** 0.2562** 0.3112** 0.3232 0.2196* 0.2291† 0.2918* 0.3210† KP4POI(cat+add+lat/lon) 0.2508** 0.2560** 0.3130** 0.3287** 0.2197* 0.2290† 0.2913** 0.3201 each check-in includes a timestamp, lat/lon, and category. The addresses were obtained by reverse geocoding: NYC via Nominatim10 (“street” level), TKY via Yahoo!11 (“chome” level). To avoid invading privacy and reduce the number of reverse geocoding, the coordinates were rounded to three decimal places. The TIST 2015 dataset [44] includes check-ins from 415 cities in 77 countries, where each city had at least 10K check-ins, and the city names were used as addresses. The WWW 2019 dataset [45] covers global check-ins12 and includes check-ins and user social graphs13. The Gowalla dataset [46] covers global check-ins14 and includes a social graph, where because many users had few check-ins, we retained only those with 10+ check-ins (69,705 of 196,591).
We designed four types of research questions: one concerning fine-tuning time and the applicability of LLM (4.2.1), two that address venue KP (4.2.2 and 4.2.3), and one exploring user KP (4.2.4). 4.2.1. RQ1: Can large-scale LLMs be applied to large-scale POI recommendation? Table 2 shows the fine-tuning time required by our KP4POI using T5-small and SeCor using Llama2-7B. The gap between the two is significant. In fact, even in the experiments of [ 21], only 850,010 check-ins out of 5 million were used from the Gowalla dataset. Although parallelization makes it feasible to run experiments on datasets of the NYC/TKY scale, applying large-scale LLMs solely to obtain embeddings for large-scale data is impractical. Therefore, it is essential to improve performance by using compact 10https://nominatim.org/ 11https://developer.yahoo.co.jp/webapi/map/openlocalplatform/v1/reversegeocoder.html 12This does not include city names but does include lat/lon and country codes. We used 629 major cities (TIST2015 cities + capital cities) and matched check-ins to the closest city with the same country code. 13Although two types of social graph are provided, we used the newer one. 14Because country codes are not provided, we matched each check-in to the nearest city using lat/lon. LLMs. In addition to the significant diference in the fine-tuning time (Table implications emerge.
• Scalability: Our method handles tens of millions of check-ins with fine-tuning times under 10 minutes per epoch, making it practical for real-time model updates in production. • Cost-eficiency : Unlike large-scale LLMs, which require substantial GPU memory and power, T5-small requires only modest computational resources, making it accessible to institutions with limited hardware. • Maintainability: The lightweight architecture allows for rapid fine-tuning as user behavior evolves, facilitating continuous learning without downtime or extensive costs. • Broader applicability: The method is suitable not only for centralized cloud systems, but also for edge deployment scenarios, promoting POI personalization in privacy-sensitive or ofline-first environments.
These findings reafirm that KP4POI that enhances small LLMs with structured knowledge is a practical and scalable direction for future POI recommendation systems. 4.2.2. RQ2: Which is more efective: address or latitude and longitude? We compare KP4POI(add) and KP4POI(lat/lon) to assess the contribution of diferent types of geographical knowledge15. Although Table 3 shows only small diferences between address and lat/lon in the NYC dataset due to the limited granularity of US street-level addresses, significant gains in TKY highlight the usefulness of the hierarchical structure of Japanese addresses (e.g., prefecture → city → chome (block)), which encodes spatial clustering and provides richer cues than raw coordinates. This also explains why the overall contribution of address-based knowledge may appear small in some datasets: when addresses lack hierarchical depth or are too coarse, they fail to capture the fine-grained distances that strongly influence tourist behaviour. In practice, however, address-based knowledge remains easier to interpret, more stable against noise, and often more accessible in privacy-sensitive or anonymized datasets where lat/lon is not disclosed. In contrast, latitude/longitude directly preserves precise spatial proximity. Thus, while their relative performance diferences are modest, the two variants capture complementary aspects of geographical information, and address-level knowledge can sometimes substitute for exact coordinates in real-world applications. 4.2.3. RQ3: Which is more efective: geographical information or category? Table 3 also includes performance metrics when KP4POI(cat) is applied to the NYC and TKY datasets. Address information contributes mainly to improvements in Recall, while category information mainly enhances NDCG, which shows that geographical information and category can have diferent types of useful information. The last two rows of Table 3 show the results when combining the address or lat/lon with the category. In the NYC dataset, this combination significantly improved performance in all metrics (with 1% significance), demonstrating the efectiveness of using both types of knowledge. KP4POI(cat+add+lat/lon) improved NDCG@5 by 1.3% compared to P5(POD) in NYC.
Table 4 presents the results of similar experiments conducted on a large-scale dataset. A consistent trend is also observed here: The combination of both leads to the best results from both perspectives. These results suggest that category-based knowledge tends to improve ranking precision (NDCG), while geographical knowledge often improves recall. Although not absolute, this complementary efect was observed consistently across datasets. Importantly, combining both sources of knowledge results in statistically significant improvements in all metrics (Table 3, Table 4). This demonstrates the complementary nature of the two: while semantic similarity guides “what” the user wants, the spatial context informs “where” it is realistically accessible. This robustness is particularly beneficial in large-scale settings like TIST2015 and WWW2019, where venue diversity and user heterogeneity are 15KP4POI without add is functionally equivalent to P5, as it only uses sequential prompts without additional knowledge. For this reason, we do not report it separately.
NDCG
TIST2015 0.1863 0.1928** 0.1937** 0.2003** 0.2386 0.2413** 0.2394** 0.2440**
Recall 0.2533 0.2753** 0.2565** 0.2775**
WWW2019 NDCG 0.1905 0.1915** 0.1970** 0.1988** 0.1927 0.1988** 0.1999** 0.2058** 0.2480 0.2508** 0.2481** 0.2529** 0.2548 0.2730** 0.2570** 0.2743**
NDCG
WWW2019 0.1927 0.1946** 0.1988** 0.1993** 0.2061** 0.2480 0.2490** 0.2508** 0.2513** 0.2526**
Recall 0.2548 0.2612** 0.2730** 0.2732** 0.2741**
Gowalla NDCG
Recall 0.1929 0.1930 0.1924 0.1927 0.1950 0.1956† 0.1979** 0.1983** pronounced. For that case, enriching language models with multifaceted knowledge, both semantic and geographic, yields more reliable and context-aware POI recommendations. 4.2.4. RQ4: Does user information improve performance? Table 5 shows the performance results when the user KP (KP4POI(u)) is applied to the WWW2019 dataset. Except for NDCG@5, all metrics show statistically significant improvements at the 1% level. This indicates that even when only facility IDs are available without lat/lon due to privacy issues, user KP significantly improves performance. When used together with category or address information, the added benefit of user KP is minimal, suggesting that social grouping can be learned implicitly through check-in records.
Table 5 also shows the corresponding results in the Gowalla dataset. However, incorporating address information leads to statistically significant improvements in NDCG@10 and Rec@10 (1% level). Similarly, the user KP significantly improves NDCG@10 and Rec@10, confirming its efectiveness. The combination of address and user information does not provide an additional benefit.
These results demonstrate the efectiveness of KP4POI(u), particularly in environments where venue metadata are limited or unavailable. Community-level embeddings serve as surrogates for missing contextual cues, enabling the model to generalize based on social afinity and collective behavior patterns. Although user KP significantly improves performance when used alone, its marginal benefit in combination with venue information suggests some redundancy between social and spatial signals. For privacy-sensitive contexts, user KP can be efective when explicit social graphs are available (e.g., in Gowalla or WWW2019). However, its utility for cold-start users in real-world scenarios without such graphs may be limited. Its eficiency and interpretability further support its inclusion in scalable POI recommendation pipelines.
Although the contribution of user-knowledge prompts appears small, this can be partly explained by the characteristics of the datasets. Only a fraction of users are connected on social graphs, and clustering from sparse interactions yields coarse community labels. Consequently, the additional signal provided by user KP is weaker than the venue-based knowledge. However, user-level information remains valuable in privacy-sensitive contexts or datasets where precise location data are unavailable, as it can still capture group-level behavioral tendencies.
In this paper, we proposed KP4POI, a knowledge-enhanced POI recommendation framework that uses compact LLMs with knowledge prompting derived from venues and users. By reformulating POI recommendation as a natural language task, our method integrates semantic and contextual knowledge without incurring the heavy computational costs of large-scale embedding-based approaches. Extensive experiments on four real-world datasets demonstrated that KP4POI achieves competitive performance while reducing training time and resource requirements. In addition, our ablation study showed that category- and geography-based knowledge prompts contribute diferently to recommendation performance, and their complementary efects improve both ranking precision and recall. Although we leave for future work a more comprehensive evaluation, including additional baselines (e.g., popularity, matrix factorization), user cold-start scenarios, and broader recommendation metrics such as diversity and catalog coverage, these results highlight the practicality of incorporating structured knowledge into lightweight LLMs for POI recommendation.
Our findings indicate that KP4POI ofers a scalable, interpretable, and privacy-aware solution to POI recommendation. This work opens promising directions for building more intelligent, adaptive, and eficient location-based services using LLMs augmented by external knowledge.
During the preparation of this work, the author used ChatGPT-4 and writefull in order to: Grammar and spelling check. After using these services, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content. [22] J. Chen, H. Wang, J. Shang, GeoMamba: Toward eficient geography-aware sequential POI recommendation, IEEE Access 12 (2024) 167906–167918. doi:10.1109/ACCESS.2024.3482116. [23] J. Zhai, X. Zheng, C.-D. Wang, H. Li, Y. Tian, Knowledge prompt-tuning for sequential recommendation, in: Proceedings of the 31st ACM International Conference on Multimedia, MM ’23, Association for Computing Machinery, New York, NY, USA, 2023, pp. 6451–6461. URL: https://doi.org/10.1145/3581783.3612252. doi:10.1145/3581783.3612252. [24] Z. Zhao, W. Fan, J. Li, Y. Liu, X. Mei, Y. Wang, Z. Wen, F. Wang, X. Zhao, J. Tang, Q. Li, Recommender systems in the era of large language models (LLMs), IEEE Transactions on Knowledge and Data Engineering 36 (2024) 6889–6907. doi:10.1109/TKDE.2024.3392335. [25] G. Agrawal, T. Kumarage, Z. Alghamdi, H. Liu, Can knowledge graphs reduce hallucinations in LLMs? : A survey, in: K. Duh, H. Gomez, S. Bethard (Eds.), Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Association for Computational Linguistics, Mexico City, Mexico, 2024, pp. 3947–3960. URL: https://aclanthology.org/2024.naacl-long.219/. doi:10.18653/ v1/2024.naacl-long.219. [26] C. Xu, X. Mei, D. Liu, K. Zhao, A. S. Ding, An eficient privacy-preserving point-of-interest recommendation model based on local diferential privacy, Complex & Intelligent Systems 9 (2023) 3277–3300. URL: https://doi.org/10.1007/s40747-022-00917-0. doi:10.1007/s40747-022-00917-0. [27] Y. Tachioka, User knowledge prompt for sequential recommendation, in: Proceedings of the 18th ACM Conference on Recommender Systems, RecSys ’24, Association for Computing Machinery, New York, NY, USA, 2024, pp. 1142–1146. URL: https://doi.org/10.1145/3640457.3691714. doi:10. 1145/3640457.3691714. [28] L. Li, Y. Zhang, L. Chen, Prompt distillation for eficient LLM-based recommendation, in: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM ’23, Association for Computing Machinery, New York, NY, USA, 2023, pp. 1348–1357. URL: https://doi.org/10.1145/3583780.3615017. doi:10.1145/3583780.3615017. [29] C. Rafel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, P. J. Liu, Exploring the limits of transfer learning with a unified text-to-text transformer, J. Mach. Learn. Res. 21 (2020). [30] H.-C. Lai, Y.-S. Lu, M.-F. Wang, Y.-C. Chen, W.-Y. Shih, J.-L. Huang, SPENT+: A category- and region-aware successive POI recommendation model, in: 2021 22nd Asia-Pacific Network Operations and Management Symposium (APNOMS), 2021, pp. 230–233. doi:10.23919/APNOMS52696. 2021.9562645. [31] F. Yu, L. Cui, W. Guo, X. Lu, Q. Li, H. Lu, A category-aware deep model for successive POI recommendation on sparse check-in data, in: Proceedings of The Web Conference 2020, WWW ’20, Association for Computing Machinery, New York, NY, USA, 2020, pp. 1264–1274. URL: https: //doi.org/10.1145/3366423.3380202. doi:10.1145/3366423.3380202. [32] X. Wang, Y. Liu, X. Zhou, Z. Leng, X. Wang, Next POI recommendation method based on category preference and attention mechanism in LBSNs, in: Web and Big Data: 6th International Joint Conference, APWeb-WAIM 2022, Nanjing, China, November 25–27, 2022, Proceedings, Part II, SpringerVerlag, Berlin, Heidelberg, 2022, pp. 12–19. URL: https://doi.org/10.1007/978-3-031-25198-6_2. doi:10.1007/978-3-031-25198-6_2. [33] C. Song, J. Wen, S. Li, Personalized POI recommendation based on check-in data and geographicalregional influence, in: Proceedings of the 3rd International Conference on Machine Learning and Soft Computing, ICMLSC ’19, Association for Computing Machinery, New York, NY, USA, 2019, pp. 128–133. URL: https://doi.org/10.1145/3310986.3311034. doi:10.1145/3310986.3311034. [34] J. Tang, J. Jin, Z. Miao, B. Zhang, Q. An, J. Zhang, Region-aware POI recommendation with semantic spatial graph, in: 2021 IEEE 24th International Conference on Computer Supported Cooperative Work in Design (CSCWD), 2021, pp. 214–219. doi:10.1109/CSCWD49262.2021.9437810. [35] X. Song, Z. Liu, L. Meng, D. Chu, J. Yu, Q. Z. Sheng, Accurate POI recommendation for random groups with improved graph neural networks and a multi-negotiation model, Scientific Reports 15 (2025) 7531. URL: https://doi.org/10.1038/s41598-025-91805-3. doi:10.1038/ s41598-025-91805-3. [36] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, E. Lefebvre, Fast unfolding of communities in large networks, Journal of Statistical Mechanics: Theory and Experiment 2008 (2008) P10008. URL: https://dx.doi.org/10.1088/1742-5468/2008/10/P10008. doi:10.1088/1742-5468/2008/10/P10008. [37] H. Cai, V. W. Zheng, K. C.-C. Chang, A comprehensive survey of graph embedding: Problems, techniques, and applications, IEEE Transactions on Knowledge and Data Engineering 30 (2018) 1616–1637. doi:10.1109/TKDE.2018.2807452. [38] A. Grover, J. Leskovec, node2vec: Scalable feature learning for networks, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, Association for Computing Machinery, New York, NY, USA, 2016, pp. 855–864. URL: https://doi.org/10.1145/2939672.2939754. doi:10.1145/2939672.2939754. [39] D. Arthur, S. Vassilvitskii, k-means++: the advantages of careful seeding, in: Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, Society for Industrial and Applied Mathematics, USA, 2007, pp. 1027–1035. [40] Z. Zhang, B. Wang, Prompt learning for news recommendation, in: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, Association for Computing Machinery, New York, NY, USA, 2023, pp. 227–237. URL: https://doi.org/10.1145/3539618.3591752. doi:10.1145/3539618.3591752. [41] Y. Li, T. Chen, Y. Luo, H. Yin, Z. Huang, Discovering collaborative signals for next POI recommendation with iterative seq2graph augmentation, in: Z.-H. Zhou (Ed.), Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, International Joint Conferences on Artificial Intelligence Organization, 2021, pp. 1491–1497. URL: https://doi.org/10.24963/ijcai.2021/206. doi:10.24963/ijcai.2021/206, main Track. [42] X. Wang, F. Fukumoto, J. Cui, Y. Suzuki, J. Li, D. Yu, EEDN: Enhanced encoder-decoder network with local and global context learning for POI recommendation, in: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, Association for Computing Machinery, New York, NY, USA, 2023, pp. 383–392. URL: https://doi.org/10.1145/3539618.3591678. doi:10.1145/3539618.3591678. [43] D. Yang, D. Zhang, V. W. Zheng, Z. Yu, Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs, IEEE Transactions on Systems, Man, and Cybernetics: Systems 45 (2015) 129–142. doi:10.1109/TSMC.2014.2327053. [44] D. Yang, D. Zhang, B. Qu, Participatory cultural mapping based on collective behavior data in location-based social networks, ACM Trans. Intell. Syst. Technol. 7 (2016). URL: https://doi.org/10. 1145/2814575. doi:10.1145/2814575. [45] D. Yang, B. Qu, J. Yang, P. Cudre-Mauroux, Revisiting user mobility and social relationships in LBSNs: A hypergraph embedding approach, in: The World Wide Web Conference, WWW ’19, Association for Computing Machinery, New York, NY, USA, 2019, pp. 2147–2157. URL: https: //doi.org/10.1145/3308558.3313635. doi:10.1145/3308558.3313635. [46] E. Cho, S. A. Myers, J. Leskovec, Friendship and mobility: User movement in location-based social networks, in: Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’11, Association for Computing Machinery, New York, NY, USA, 2011, pp. 1082–1090. URL: https://doi.org/10.1145/2020408.2020579. doi:10.1145/2020408. 2020579.