Forecasting Unseen Points of Interest Visits Using Context and Proximity Priors

Cyrus Shahabi; Shang-Ling Hsu; Ziyao Li

REVIEW 5 major objections 7 minor 24 references

Forecasting Unseen Points of Interest Visits Using Context and Proximity Priors

T0 review · 5 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that next-POI forecasting can be reorganized into semantic-category prediction plus a proximity prior, which lets the model assign nonzero scores to points of interest that never appeared in training.

desk verdict Simple but sound idea for scoring unseen POIs via category + distance prior; the write-up and evaluation need real work, but it deserves referee attention. read the letter →

arxiv 2411.15285 v1 pith:42I3XR55 submitted 2024-11-22 cs.LG cs.AI

Ziyao Li , Shang-Ling Hsu , Cyrus Shahabi This is my paper · ORCID

classification cs.LGcs.AI

keywords point-of-interestforecastingunseenPOIpredictionnextvisitsemanticcontextproximitypriorhumanmobility

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to solve a blind spot in next-POI forecasting: models trained as POI classifiers cannot assign probability to a POI that was absent from training, such as a newly opened restaurant. The proposed method instead predicts the semantic category of the next visit and multiplies that by an empirical distance prior, producing a score for every candidate POI, seen or unseen. Results on New York City check-in data show a reported 17% improvement in accuracy and a slower accuracy decline as unseen POIs accumulate. If correct, the model would let location recommenders suggest new venues without continuous retraining.

What carries the argument

The load-bearing object is the joint probability approximation defined by Eqs. (8)–(14): a binary distance-similarity operator that compares candidate POIs by distance from the user's last location, a semantic-context classifier over POI categories, and a global discrete proximity prior built by histogramming distances between consecutive visits. The factorization multiplies the category probability and the proximity prior for each candidate POI, and this single mechanism is what gives unseen POIs nonzero ranking scores.

What would settle it

Compare per-category histograms of distances between consecutive visits in the real check-in data; if the distributions differ materially, the single global prior in Eq. (14) is false. A direct follow-up is to rerun the unseen-POI evaluation with category-conditioned distance priors; if accuracy does not improve, the independence assumption is not the bottleneck.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that replacing the POI classifier with a factored objective—predict the category of the next visit, then combine that category distribution with a global proximity prior—lets a model generalize to POIs it has never seen. Concretely, the approximation in Eq. (10) factors the probability of the next POI given the user's history into the probability of the next semantic category times the probability that the candidate POI sits at a typical next-trip distance. Because every candidate POI with a known category and location can be plugged into both factors, unseen POIs receive nonzero probability, whereas a softmax over training POIs assigns them essentially zero. The paper's experiments support this by showing nonzero accuracy on unseen POIs where the baseline gives zero, and a slower drop in accuracy as the fraction of unseen POIs increases.

Load-bearing premise

The approach assumes that once a user's history is known, the category of the next place and how far away it is are independent, and that one global distance distribution fits all users and times; if people travel farther for certain categories, the combined scores are miscalibrated.

Editorial extensions

If this is right

A recommender can rank a newly opened venue as soon as its category and coordinates are known, with no retraining.
As the share of unseen POIs grows, accuracy falls more slowly than it does for the baseline, so the model degrades more gracefully in an evolving urban environment.
Because the category space is far smaller than the POI space, training signals transfer across POIs of the same category.
The factored objective could be combined with a separate seen-POI classifier to improve overall accuracy, as the paper itself suggests.
The method can be evaluated at any candidate POI, seen or unseen, so it directly addresses the open-set aspect of next-POI prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The conditional-independence assumption in Eq. (10) suggests a direct extension: condition the distance prior on the predicted category, which would test whether people travel farther for certain categories.
The method transfers to any open-set recommendation problem where instances have categorical labels and coordinates, not just points of interest.
The global proximity prior is estimated once from one city's check-ins; in cities with different transport modes its shape may change, so a per-city or per-user prior is a natural robustness check.
Errors in category prediction multiply into the final ranking, so calibrating the category softmax could improve unseen-POI accuracy.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the paper has a genuinely simple and useful idea—decompose next-POI prediction into category prediction plus a distance prior to score POIs, including ones never seen in training—but the current write-up overstates the results and leaves the exact scoring formula underspecified. Worth sending to referees, though they will need to demand serious revisions.

What is new: instead of training a POI classifier that only scores known POIs, the model predicts the next semantic category and multiplies that probability by an empirical distance prior to score every candidate POI. That mechanism is clean and, as far as I can tell, not present in the cited next-POI papers. The temporal split for creating unseen POIs is a reasonable evaluation protocol.

What it does well: the motivation is clearly explained—why classifier logits for unseen classes collapse to zero—and reusing MobTCast's encoder for the category head is a pragmatic choice. The paper is short and readable, and the math is mostly standard.

Soft spots, in order of severity: (1) The headline results are misreported. Table I shows MobTCast at 0.0000 on unseen POIs, so "17% improvement" is relative to zero—meaningless. They likely mean an absolute 17.89 percentage-point difference, but they need to say that. (2) The 635.417% slope decrease in Fig. 4 has no regression details and the phrasing is confusing. "Average decrease in slope" is not a standard quantity. (3) The final scoring rule is underspecified: after multiplying the two probabilities, how are scores normalized across all candidate POIs? How are the distance buckets defined? This has to be written down precisely. (4) The independence assumption in Eq. (10) and the global distance prior in Eq. (14) are plausible but never validated. A per-category distance baseline or a simple popularity-plus-proximity baseline would help. (5) The experimental comparison is only against MobTCast; no ablation of the proximity prior, no other next-POI methods, and one dataset.

None of these are fatal to the core idea. The mechanism genuinely allows nonzero probability for unseen POIs, which is the paper's real selling point. But the evidence as presented is thin, and the paper needs a fair amount of rewriting before it is convincing.

Who it is for: researchers working on next-POI forecasting or location-based recommendation who care about the cold-start / new-POI problem. I would bring it to a reading group—the independence assumption is a good discussion point—and I would cite it if I wrote on this topic.

Recommendation: send it to peer review. The novelty is real and the topic is timely, but I would expect referees to require a fully specified method, stronger baselines, and corrected reporting before acceptance.

Referee Report

5 major / 7 minor

Summary. The paper proposes a next-POI forecasting method that predicts the semantic category of the next visit and combines it with a distance-based proximity prior, enabling predictions for POIs that were unseen during training. The method encodes a user's visit sequence with a Transformer following MobTCast, trains a category classifier, and multiplies the predicted category distribution by a precomputed distance histogram. Experiments on Foursquare NYC use a temporal split to create unseen POIs, reporting top-k accuracy improvements over MobTCast and a slower accuracy decline as the fraction of unseen POIs grows.

Significance. If the claims are substantiated, the paper offers a practically relevant direction for location-based recommendation in dynamic urban environments with newly opened POIs. The core idea of predicting semantic context rather than POI identities is reasonable, and the temporal-split evaluation protocol for unseen POIs is a useful contribution. However, the current evaluation does not support the headline quantitative claims: the reported relative improvement on unseen POIs is undefined because the baseline score is zero, the robustness slope claim lacks regression details, and the load-bearing independence assumptions in the probabilistic formulation are not empirically validated.

major comments (5)

[Table I and Section VI-B] The headline '17% higher top-20 accuracy when evaluating with the set of unseen POIs' is not well-defined: MobTCast scores 0.0000 on the unseen-POI set, so a relative percentage improvement cannot be computed. The 17% figure appears to be the relative improvement on the All-POIs row (0.1986 vs. 0.1689), but the text explicitly attributes it to the unseen set. Please state the exact comparison and metric, and avoid percentage improvements relative to a zero baseline.
[Fig. 4 and Section VI-B] The claim of a '635.417% decrease in slope' is unexplained. The paper does not specify the regression model, the variables (e.g., accuracy vs. percentage of unseen POIs), whether the slope is computed per accuracy metric or aggregated, or any measure of uncertainty. Without these details and a statistical test, the robustness claim is unsubstantiated.
[Section V-A, Eq. (10)] Eq. (10) factorizes the joint probability of the next POI's category and its distance from the current location given the visit history into a product of marginals. This conditional independence assumption is load-bearing because the final score is a product of these two factors, and it is not empirically validated. I request a comparison against a model that estimates the joint distribution without this factorization, or at least a report of conditional distance distributions per category to assess whether the assumption is reasonable.
[Section V-C, Eq. (14)] The global proximity prior P assumes that the distance between consecutive visits follows one distribution for all users, times, and POI categories. This is a strong assumption for mobility data; for example, users may travel systematically farther for airports or specialty venues than for coffee shops. Since unseen POIs have no learned signal other than this prior, any systematic category-distance correlation would miscalibrate the rankings. Please provide evidence that this single-distribution approximation is adequate, or evaluate sensitivity to category-, user-, or time-conditioned priors.
[Section V-C, final paragraph] The combination rule is not fully specified. The text says 'we obtain Pr[...] by combining the approximated proximity prior distribution, the input visit sequence, and the candidate POI', but no equation defines how the predicted category distribution and the proximity prior are normalized and multiplied. Without the exact scoring function, the experiments are not reproducible. Please provide the full formula, including how the discrete distance buckets map to candidate POIs and how the two terms are weighted or normalized.

minor comments (7)

[Abstract and Section VI-B] The abstract states a '17% improvement in accuracy' without specifying the metric; please report the exact metric and the comparison set.
[Conclusion] Typo: 'absense' should be 'absence'.
[Fig. 4] The legend label 'MobT Cast' contains an erroneous space; it should be 'MobTCast'.
[References] References [4] and [16] are the same paper; duplicate entries should be removed.
[Section V-B] The sentence 'we leverage enhance a visit sequence representations' is ungrammatical and should be revised.
[Eq. (8)] The binary operator is displayed as 'i ≡ ≡ j', which appears to be a formatting corruption; please fix the notation.
[Section VI-A] Please state the number of distinct POI categories in the FS-NYC dataset, since this directly affects the semantic classification task.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: category predictor and proximity prior are independent inputs; self-citations are background only.

full rationale

The paper's central claim is that a next POI can be forecast without training a POI classifier by predicting a category distribution (Section V-B) and multiplying by a proximity prior (Section V-C). The category classifier is trained on POI categories, not on unseen POI IDs, and the proximity prior is a histogram of consecutive-visit distances estimated from the training split. Neither term is defined in terms of the target POI labels. The factorization in Eq. (10) and the global-distance assumption in Eq. (14) are stated modeling assumptions, not circular reductions: they could be wrong (e.g., if category and distance correlate), but being wrong is a correctness risk, not circularity. The only self-citations ([9], [11]) appear in related-work surveys and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work, and the MobTCast encoder is externally cited. Evaluation against MobTCast, which cannot emit unseen POIs, may be an easy baseline, but that is an experimental-design concern, not a circular derivation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central method rests on a factorization of the next-POI probability into a category term and a distance term, plus a global distance prior fitted to training data. These are unvalidated modeling choices rather than derived results, and the category-coverage assumption limits the model to new POIs with familiar categories.

free parameters (3)

Proximity prior histogram P = Histogram over distance buckets (Fig. 2); bucket widths not reported
Estimated from distances between consecutive visits in the training set (Section V-C) and used directly as the proximity term in Eq. (10). The choice of bucket size changes every predicted score.
Number and width of distance buckets = Not reported
The 'counting buckets' in Section V-C are chosen by hand and are not documented; the granularity of the distance prior is a free choice that affects results.
Model hyperparameters = window=20, hidden dim=128, embeddings: POI=80, category=24, time=24
Hand-chosen values in Section VI-A; they affect the semantic context classifier but are not the main source of the generalization claim.

assumptions (3)

ad hoc to paper Conditional independence of next-POI category and distance to current location given the visit history, Eq. (10)
The factorization in Eq. (10) is asserted without justification. If category preference correlates with travel distance, the combined score is miscalibrated.
domain assumption A single global distance distribution P applies to all users, times, and POIs, Eqs. (13)-(14)
The distance between consecutive visits is modeled as one distribution regardless of user habits, time of day, or city area. This discards user-specific mobility constraints and is likely violated in practice.
domain assumption Every candidate POI, including unseen ones, has a category among the CS categories seen in training
The semantic context classifier in Eq. (11) only assigns probabilities to categories present in training; a genuinely new category gets zero probability by construction. The paper does not discuss this limitation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Forecasting Unseen Points of Interest Visits Using Context and Proximity Priors." pith.science (2026). https://pith.science/paper/42I3XR55

@misc{pith2026241115285,
  author       = {Pith},
  title        = {Pith review of: Forecasting Unseen Points of Interest Visits Using Context and Proximity Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42I3XR55}},
  note         = {Machine review of arXiv:2411.15285}
}

read the original abstract

Understanding human mobility behavior is crucial for numerous applications, including crowd management, location-based recommendations, and the estimation of pandemic spread. Machine learning models can predict the Points of Interest (POIs) that individuals are likely to visit in the future by analyzing their historical visit patterns. Previous studies address this problem by learning a POI classifier, where each class corresponds to a POI. However, this limits their applicability to predict a new POI that was not in the training data, such as the opening of new restaurants. To address this challenge, we propose a model designed to predict a new POI outside the training data as long as its context is aligned with the user's interests. Unlike existing approaches that directly predict specific POIs, our model first forecasts the semantic context of potential future POIs, then combines this with a proximity-based prior probability distribution to determine the exact POI. Experimental results on real-world visit data demonstrate that our model outperforms baseline methods that do not account for semantic contexts, achieving a 17% improvement in accuracy. Notably, as new POIs are introduced over time, our model remains robust, exhibiting a lower decline rate in prediction accuracy compared to existing methods.

Figures

Figures reproduced from arXiv: 2411.15285 by the authors.

**Figure 1.** Our proposed model. visited POIs and timestamps, but the semantic context of the POIs. The key idea is that the semantic similarity between different POIs provides valuable insight for predicting future visits, especially in a dynamic context. For example, one who usually visits a restaurant on Friday evenings might be interested in visiting a newly-opened restaurant on a future Friday evening. Although those restau… view at source ↗

**Figure 2.** The proximity priors histogram representing the num [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** The training, testing, validation, and unseen POI [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The comparison of MobTCast and our model accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

[1]

To- wards rich mobile phone datasets: Lausanne data collection campaign,

N. Kiukkonen, J. Blom, O. Dousse, D. Gatica-Perez, and J. Laurila, “To- wards rich mobile phone datasets: Lausanne data collection campaign,” Proc. ICPS, Berlin , vol. 68, no. 7, 2010

work page 2010
[2]

Zheng, H

Y . Zheng, H. Fu, X. Xie, W.-Y . Ma, and Q. Li, Geolife GPS trajectory dataset - User Guide , geolife gps trajectories 1.1 ed., July 2011. [Online]. Available: https://www.microsoft.com/en-us/research/publication/geolife-gps-trajectory-dataset-user-guide/

work page 2011
[3]

Foursquare check-in dataset,

Foursquare, “Foursquare check-in dataset,” 2018, accessed: 2024-06-04. [Online]. Available: https://sites.google.com/site/yangdingqi/home/foursquare-dataset

work page 2018
[4]

Mobtcast: Levera ging auxiliary trajectory forecasting for human mobility prediction,

H. Xue, F. Salim, Y . Ren, and N. Oliver, “Mobtcast: Levera ging auxiliary trajectory forecasting for human mobility prediction,” Advances in Neural Information Processing Systems , vol. 34, pp. 30 380–30 391, 2021

work page 2021
[5]

Getnext: trajectory ﬂow map enhanced transformer for next poi recommendation,

S. Y ang, J. Liu, and K. Zhao, “Getnext: trajectory ﬂow map enhanced transformer for next poi recommendation,” in Proceedings of the 45th International ACM SIGIR Conference on research and develop ment in information retrieval, 2022, pp. 1144–1153

work page 2022
[6]

Context-aware deep m odel for joint mobility and time prediction,

Y . Chen, C. Long, G. Cong, and C. Li, “Context-aware deep m odel for joint mobility and time prediction,” in Proceedings of the 13th International Conference on W eb Search and Data Mining , 2020, pp. 106–114

work page 2020
[7]

Hierarchical transformer with spati o-temporal context aggregation for next point-of-interest recommend ation,

J. Xie and Z. Chen, “Hierarchical transformer with spati o-temporal context aggregation for next point-of-interest recommend ation,” ACM Transactions on Information Systems , vol. 42, no. 2, pp. 1–30, 2023

work page 2023
[8]

Attention is all you need,

A. V aswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones , A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

2017

Show all 24 references

[9]

Generat ing realistic and representative trajectories with mobility behavior cl ustering,

H. Lin, S. Shaham, Y .-Y . Chiang, and C. Shahabi, “Generat ing realistic and representative trajectories with mobility behavior cl ustering,” in Proceedings of the 31st ACM International Conference on Adv ances in Geographic Information Systems , 2023, pp. 1–4

2023
[10]

Practical synthetic human trajectories ge neration based on variational point processes,

Q. Long, H. Wang, T. Li, L. Huang, K. Wang, Q. Wu, G. Li, Y . L iang, L. Y u, and Y . Li, “Practical synthetic human trajectories ge neration based on variational point processes,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023, ...

2023
[11]

Trajgpt: Controlled synthetic trajectory generation using a multit ask transformer- based spatiotemporal model,

S.-L. Hsu, E. Tung, J. Krumm, C. Shahabi, and K. Shaﬁque, “Trajgpt: Controlled synthetic trajectory generation using a multit ask transformer- based spatiotemporal model,” in Proceedings of the 32nd ACM Inter- national Conference on Advances in Geographic Information Systems (...

2024
[12]

Friendship and mob ility: user movement in location-based social networks,

E. Cho, S. A. Myers, and J. Leskovec, “Friendship and mob ility: user movement in location-based social networks,” in Proceedings of the 17th ACM SIGKDD international conference on Knowledge discover y and data mining , 2011, pp. 1082–1090

2011
[13]

Deep- move: Predicting human mobility with attentional recurren t networks,

J. Feng, Y . Li, C. Zhang, F. Sun, F. Meng, A. Guo, and D. Jin , “Deep- move: Predicting human mobility with attentional recurren t networks,” in Proceedings of the 2018 world wide web conference , 2018, pp. 1459– 1468

2018
[14]

Predicting the next lo cation: A recurrent model with spatial and temporal contexts,

Q. Liu, S. Wu, L. Wang, and T. Tan, “Predicting the next lo cation: A recurrent model with spatial and temporal contexts,” in Proceedings of the AAAI conference on artiﬁcial intelligence , vol. 30, no. 1, 2016

2016
[15]

Mining geogra phic-temporal- semantic patterns in trajectories for location prediction ,

J. J.-C. Ying, W.-C. Lee, and V . S. Tseng, “Mining geogra phic-temporal- semantic patterns in trajectories for location prediction ,” ACM Transac- tions on Intelligent Systems and Technology (TIST) , vol. 5, no. 1, pp. 1–33, 2014

2014
[16]

Mobtcast: Lever aging auxiliary trajectory forecasting for human mobility prediction,

H. Xue, F. Salim, Y . Ren, and N. Oliver, “Mobtcast: Lever aging auxiliary trajectory forecasting for human mobility prediction,” Advances in Neural Information Processing Systems , vol. 34, pp. 30 380–30 391, 2021

2021
[17]

Where to go next: A spatio-temporal gated network f or next poi recommendation,

P . Zhao, A. Luo, Y . Liu, J. Xu, Z. Li, F. Zhuang, V . S. Sheng , and X. Zhou, “Where to go next: A spatio-temporal gated network f or next poi recommendation,” IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 5, pp. 2512–2524, 2020

2020
[18]

Deep move: Predicting human mobility with attentional recurrent netw orks,

J. Feng, X. Li, J. Zhang, C. Zhang, J. Han, and K. Li, “Deep move: Predicting human mobility with attentional recurrent netw orks,” in Proceedings of the 2018 W orld Wide W eb Conference . ACM, 2018, pp. 1459–1468

2018
[19]

Stan: Spa tio-temporal attention network for next location recommendation,

X. Liu, X. He, B. Tian, J. Wang, and T.-S. Chua, “Stan: Spa tio-temporal attention network for next location recommendation,” in Proceedings of the 28th ACM International Conference on Information and Kn owledge Management. ACM, 2019, pp. 2193–2196

2019
[20]

S tgn: Spatio-temporal gated network for human mobility predicti on,

Y . Wang, J. Feng, Z. Liu, X. Wang, T.-S. Chua, and X. He, “S tgn: Spatio-temporal gated network for human mobility predicti on,” IEEE Transactions on Knowledge and Data Engineering , 2021

2021
[21]

Discovering subsequence patterns for next poi recommenda tion

K. Zhao, Y . Zhang, H. Yin, J. Wang, K. Zheng, X. Zhou, and C . Xing, “Discovering subsequence patterns for next poi recommenda tion.” in IJCAI, vol. 2020, 2020, pp. 3216–3222

2020
[22]

Where would i go n ext? large language models as human mobility predictors,

X. Wang, M. Fang, Z. Zeng, and T. Cheng, “Where would i go n ext? large language models as human mobility predictors,” arXiv preprint arXiv:2308.15197, 2023

2023 arXiv
[23]

A poi cate gorization by composition of onomastic and contextual information,

S. J. Choi, H. J. Song, S. B. Park, and S. J. Lee, “A poi cate gorization by composition of onomastic and contextual information,” i n 2014 IEEE/WIC/ACM International Joint Conferences on W eb Intel ligence (WI) and Intelligent Agent Technologies (IAT) , vol. 2. IEEE, 2014, pp. 38–45

2014
[24]

Modeling user a ctivity preference by leveraging user spatial temporal characteri stics in lbsns,

D. Y ang, D. Zhang, V . W. Zheng, and Z. Y u, “Modeling user a ctivity preference by leveraging user spatial temporal characteri stics in lbsns,” IEEE Transactions on Systems, Man, and Cybernetics: System s, vol. 45, no. 1, pp. 129–142, 2014

2014

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

To- wards rich mobile phone datasets: Lausanne data collection campaign,

N. Kiukkonen, J. Blom, O. Dousse, D. Gatica-Perez, and J. Laurila, “To- wards rich mobile phone datasets: Lausanne data collection campaign,” Proc. ICPS, Berlin , vol. 68, no. 7, 2010

work page 2010

[2] [2]

Zheng, H

Y . Zheng, H. Fu, X. Xie, W.-Y . Ma, and Q. Li, Geolife GPS trajectory dataset - User Guide , geolife gps trajectories 1.1 ed., July 2011. [Online]. Available: https://www.microsoft.com/en-us/research/publication/geolife-gps-trajectory-dataset-user-guide/

work page 2011

[3] [3]

Foursquare check-in dataset,

Foursquare, “Foursquare check-in dataset,” 2018, accessed: 2024-06-04. [Online]. Available: https://sites.google.com/site/yangdingqi/home/foursquare-dataset

work page 2018

[4] [4]

Mobtcast: Levera ging auxiliary trajectory forecasting for human mobility prediction,

H. Xue, F. Salim, Y . Ren, and N. Oliver, “Mobtcast: Levera ging auxiliary trajectory forecasting for human mobility prediction,” Advances in Neural Information Processing Systems , vol. 34, pp. 30 380–30 391, 2021

work page 2021

[5] [5]

Getnext: trajectory ﬂow map enhanced transformer for next poi recommendation,

S. Y ang, J. Liu, and K. Zhao, “Getnext: trajectory ﬂow map enhanced transformer for next poi recommendation,” in Proceedings of the 45th International ACM SIGIR Conference on research and develop ment in information retrieval, 2022, pp. 1144–1153

work page 2022

[6] [6]

Context-aware deep m odel for joint mobility and time prediction,

Y . Chen, C. Long, G. Cong, and C. Li, “Context-aware deep m odel for joint mobility and time prediction,” in Proceedings of the 13th International Conference on W eb Search and Data Mining , 2020, pp. 106–114

work page 2020

[7] [7]

Hierarchical transformer with spati o-temporal context aggregation for next point-of-interest recommend ation,

J. Xie and Z. Chen, “Hierarchical transformer with spati o-temporal context aggregation for next point-of-interest recommend ation,” ACM Transactions on Information Systems , vol. 42, no. 2, pp. 1–30, 2023

work page 2023

[8] [8]

Attention is all you need,

A. V aswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones , A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

2017