pith. machine review for the scientific record. sign in

arxiv: 2605.05771 · v1 · submitted 2026-05-07 · 💻 cs.IR

Recognition: unknown

Beyond Long Tail POIs: Transition-Centered Generalization for Human Mobility Prediction

Dingyang Lyu, Jey Han Lau, Jianzhong Qi, Zhengjia Xu

Pith reviewed 2026-05-08 06:07 UTC · model grok-4.3

classification 💻 cs.IR
keywords human mobility predictionnext POI predictionlong-tail transitionscompositional generalizationtransition reconstructiontrajectory modelinglocation recommendation
0
0 comments X

The pith

Transition sparsity between locations, not just rare destinations, is the core limit on accurate next-POI forecasts.

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

The paper claims that human mobility prediction often fails even on common POIs because the specific source-to-destination transitions are rare or absent in training data. It reframes this as a compositional generalization task and introduces the RECAP framework to rebuild those missing transitions from two signals that transfer more readily: multi-hop paths through the global transition graph and a user's own revisit patterns. Warm-transition holdout training is added to stop the model from simply memorizing frequent pairs. If this reconstruction works, prediction accuracy rises across datasets, with the largest gains on the tail transitions that previously blocked progress.

Core claim

We argue that a core bottleneck in human mobility prediction lies in transition-level long-tail generalization. We formulate this problem as compositional generalization and propose RECAP, which reconstructs long-tail transitions from multi-hop transitivity in the global transition graph and revisit evidence from a user's historical trajectory. It further uses warm-transition holdout training to discourage memorization of frequent transitions and encourage generalization from transferable signals. Experiments show consistent accuracy gains, especially on tail transitions.

What carries the argument

RECAP framework that reconstructs unseen source-destination transitions by combining multi-hop paths in the global transition graph with user-specific revisit records.

If this is right

  • Prediction accuracy improves on real-world trajectory datasets, with clearest gains when the source-destination pair has never been seen in training.
  • Models stop over-relying on memorizing popular transitions and instead use transferable graph and user signals.
  • New visits to popular POIs become predictable when they arrive via previously unseen but reconstructible paths.
  • The same reconstruction approach applies to multiple mobility datasets without domain-specific adjustments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the reconstruction holds, similar sparsity problems in other sequence tasks such as session-based recommendation or route planning could be addressed by extracting multi-hop and personal-history signals.
  • Urban-planning tools that simulate city flows might gain reliability on unusual but plausible routes once transitions are rebuilt rather than treated as impossible.
  • The warm-holdout trick could be tested as a general regularizer in any graph-based sequence model to push generalization over memorization.

Load-bearing premise

Multi-hop paths through the transition graph plus a user's past visits can rebuild accurate long-tail transitions without adding systematic errors or needing extra tuning.

What would settle it

Run RECAP on a dataset where all long-tail transitions are deliberately removed from the graph and user histories; if accuracy on those cases does not drop below a strong baseline that uses only frequency, the reconstruction claim is false.

Figures

Figures reproduced from arXiv: 2605.05771 by Dingyang Lyu, Jey Han Lau, Jianzhong Qi, Zhengjia Xu.

Figure 1
Figure 1. Figure 1: An example of prediction failure of LoTNext on a popular destination due to an unseen transition on NYC. (a) Transition Frequency (b) Destination Frequency view at source ↗
Figure 3
Figure 3. Figure 3: (1) multi-hop transition-graph encoding for finding plausible next POIs through observed view at source ↗
Figure 3
Figure 3. Figure 3: RECAP combines ❶ transitive completion token as part of the input, ❷ user-history revisit calibration to adjust the prediction logits produced by the backbone model logits, and ❸ warm-transition holdout training to optimize the full framework. POIs that the user has visited before and adjusting their probability scores as being the next POI under the current trajectory context, and (3) warm-transition hold… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation results. Each bar shows the paired drop in percentage points from full RECAP view at source ↗
Figure 5
Figure 5. Figure 5: Impact of Graph hop number N.  view at source ↗
Figure 7
Figure 7. Figure 7: Retrieval gains from user his￾tory and two-hop graph paths. Retrieval Gains from User History and Two-Hop Graph Paths view at source ↗
Figure 8
Figure 8. Figure 8: A prediction example for an unseen transition. Prediction on Unseen Transitions view at source ↗
Figure 9
Figure 9. Figure 9: Sorted train-frequency distribution of unique test source–destination transitions. The shaded view at source ↗
Figure 10
Figure 10. Figure 10: Per-dataset HR@20 under exact train transition-frequency bins. The shaded region marks view at source ↗
Figure 11
Figure 11. Figure 11: Sensitivity to the number of trajectory tokens encoded by the Transformer backbone. The view at source ↗
Figure 12
Figure 12. Figure 12: Effect of the staged training curriculum on RECAP, evaluated on NYC with seed 3407. view at source ↗
read the original abstract

Human mobility prediction forecasts a user's next Point of Interest (POI) from historical trajectories, supporting applications from recommendation to urban planning. Recent studies have recognized the problem with long-tail POIs in human mobility prediction, which are POIs with few visit records, making new visits to such POIs difficult to predict. Our analysis shows that many predictions fail even for visits to popular POIs. The underlying cause is often transition-level sparsity: the corresponding source-destination transition appears rarely, or never appears, in the training set. We therefore argue that a core bottleneck in human mobility prediction lies in transition-level long-tail generalization. We formulate this problem as compositional generalization and propose a tRansition rEconstruction framework for Compositional generAlization in next-POI prediction (RECAP). RECAP reconstructs long-tail transitions from two generalizable signals: multi-hop transitivity in the global transition graph and revisit evidence from a user's historical trajectory. It further uses warm-transition holdout training to discourage memorization of frequent transitions and encourage generalization from transferable signals. Experiments on multiple real-world datasets show that RECAP consistently improves prediction accuracy, with clear gains on tail transitions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper argues that human mobility prediction suffers primarily from transition-level long-tail sparsity (rare or unseen source-destination pairs) rather than solely POI long-tail, formulates this as a compositional generalization task, and proposes the RECAP framework. RECAP reconstructs unseen transitions via multi-hop transitivity in the global transition graph combined with per-user revisit evidence, trained using a warm-transition holdout strategy to discourage memorization. Experiments on multiple real-world datasets report consistent accuracy gains, especially on tail transitions.

Significance. If the reconstruction mechanism holds without systematic bias, the work could meaningfully redirect mobility prediction research toward transition-centric generalization, with downstream value for recommendation and urban planning. Credit is due for the multi-dataset evaluation and explicit focus on tail-transition performance; these elements strengthen the practical relevance of the central claim.

major comments (3)
  1. [§4 Experiments] §4 Experiments: the reported accuracy improvements on tail transitions are presented without ablation studies that isolate the contribution of multi-hop transitivity versus user revisit evidence, nor with statistical significance tests or exact baseline specifications. This information is load-bearing for verifying that gains arise from the proposed compositional reconstruction rather than other modeling choices.
  2. [§3.3 and §4.2] §3.3 and §4.2: the central assumption that global multi-hop paths plus sparse revisit evidence faithfully reconstruct user-specific long-tail transitions is not directly tested; no metric (e.g., precision or recall of reconstructed edges against held-out ground truth) is reported to quantify reconstruction fidelity or to diagnose cases where the two signals conflict.
  3. [§3.2] §3.2: the warm-transition holdout training strategy is motivated as a guard against memorization, yet no sensitivity analysis on the holdout ratio or empirical check that it actually induces compositional rather than correlational learning is provided, leaving the generalization claim under-supported.
minor comments (2)
  1. [Problem formulation] The precise frequency threshold used to define 'long-tail transitions' should be stated explicitly in the problem formulation section for reproducibility.
  2. [Figure 2] Figure 2 (transition reconstruction diagram) would benefit from an accompanying equation or pseudocode that formalizes how multi-hop paths are aggregated with revisit counts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas to strengthen the empirical support for RECAP's transition reconstruction approach. We address each major comment point by point below and will revise the manuscript accordingly to incorporate the requested analyses.

read point-by-point responses

  1. Referee: [§4 Experiments] §4 Experiments: the reported accuracy improvements on tail transitions are presented without ablation studies that isolate the contribution of multi-hop transitivity versus user revisit evidence, nor with statistical significance tests or exact baseline specifications. This information is load-bearing for verifying that gains arise from the proposed compositional reconstruction rather than other modeling choices.

    Authors: We agree that isolating the individual contributions is essential for validating the source of the gains. In the revised version, we will add ablation studies that separately disable the multi-hop transitivity module and the user revisit evidence module, reporting their effects specifically on tail-transition accuracy. We will also include statistical significance testing (e.g., paired t-tests over multiple random seeds) and provide complete baseline specifications, including exact hyperparameter settings and implementation references, to confirm that improvements derive from the compositional reconstruction mechanism. revision: yes

  2. Referee: [§3.3 and §4.2] §3.3 and §4.2: the central assumption that global multi-hop paths plus sparse revisit evidence faithfully reconstruct user-specific long-tail transitions is not directly tested; no metric (e.g., precision or recall of reconstructed edges against held-out ground truth) is reported to quantify reconstruction fidelity or to diagnose cases where the two signals conflict.

    Authors: We acknowledge that a direct fidelity evaluation would better substantiate the reconstruction assumption. We will add quantitative metrics in the revised paper, specifically precision and recall of reconstructed transitions evaluated against held-out ground-truth transitions. We will also analyze and report instances where the multi-hop and revisit signals conflict, providing diagnostic insights into the reconstruction process and its limitations. revision: yes

  3. Referee: [§3.2] §3.2: the warm-transition holdout training strategy is motivated as a guard against memorization, yet no sensitivity analysis on the holdout ratio or empirical check that it actually induces compositional rather than correlational learning is provided, leaving the generalization claim under-supported.

    Authors: We agree that further validation of the holdout strategy would strengthen the generalization claims. In the revision, we will include a sensitivity analysis varying the holdout ratio and report its impact on performance. We will also add empirical checks that contrast results on transitions requiring compositional inference versus those amenable to memorization, to demonstrate that the strategy promotes compositional rather than correlational learning. revision: yes

Circularity Check

0 steps flagged

No significant circularity in RECAP derivation or claims

full rationale

The paper proposes RECAP as a new framework that reconstructs long-tail transitions via multi-hop transitivity in the global graph plus per-user revisit evidence, trained under warm-transition holdout to promote generalization. No equations or steps reduce by construction to fitted parameters or prior self-citations; the reconstruction signals are external to the target predictions and the accuracy gains are reported from experiments on real-world datasets. The central claim of transition-level compositional generalization is therefore self-contained and externally falsifiable rather than tautological.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that graph transitivity and revisit patterns provide reliable signals for unseen transitions and that holdout training forces genuine generalization rather than alternative memorization.

free parameters (1)
  • warm-transition holdout ratio
    Hyperparameter chosen to balance prevention of memorization against retention of sufficient training signal.
axioms (2)
  • domain assumption Multi-hop paths in the global transition graph can be used to infer direct source-destination transitions for reconstruction.
    Invoked when stating that long-tail transitions are reconstructed from transitivity in the transition graph.
  • domain assumption User revisit evidence provides an independent generalizable signal for transition reconstruction.
    Used to justify the second reconstruction source alongside the graph.

pith-pipeline@v0.9.0 · 5513 in / 1331 out tokens · 55795 ms · 2026-05-08T06:07:50.983482+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    DeepMove: Predicting Human Mobility with Attentional Recurrent Networks

    Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng Jin. DeepMove: Predicting Human Mobility with Attentional Recurrent Networks. InProceedings of the 2018 World Wide Web Conference, pages 1459–1468, 2018

    2018

  2. [2]

    Dingqi Yang, Benjamin Fankhauser, Paolo Rosso, and Philippe Cudre-Mauroux. Location Prediction over Sparse User Mobility Traces Using RNNs: Flashback in Hidden States! In Proceedings of the 29th International Joint Conference on Artificial Intelligence, pages 2184– 2190, 2020

    2020

  3. [3]

    Taming the Long Tail in Human Mobility Prediction

    Xiaohang Xu, Renhe Jiang, Chuang Yang, Zipei Fan, and Kaoru Sezaki. Taming the Long Tail in Human Mobility Prediction. InProceedings of the 38th International Conference on Neural Information Processing Systems, pages 54748–54771, 2024

    2024

  4. [4]

    STAN: Spatio-Temporal Attention Network for Next Location Recommendation

    Yingtao Luo, Qiang Liu, and Zhaocheng Liu. STAN: Spatio-Temporal Attention Network for Next Location Recommendation. InProceedings of the Web Conference, pages 2177–2185, 2021

    2021

  5. [5]

    Going Where, by Whom, and at What Time: Next Location Prediction Considering User Preference and Temporal Regularity

    Tianao Sun, Ke Fu, Weiming Huang, Kai Zhao, Yongshun Gong, and Meng Chen. Going Where, by Whom, and at What Time: Next Location Prediction Considering User Preference and Temporal Regularity. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2784–2793, 2024

    2024

  6. [6]

    Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation

    Tianhao Huang, Xuan Pan, Xiangrui Cai, Ying Zhang, and Xiaojie Yuan. Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation. InProceedings of the 38th AAAI Conference on Artificial Intelligence, pages 8535–8543, 2024

    2024

  7. [7]

    ROTAN: A Rotation- based Temporal Attention Network for Time-Specific Next POI Recommendation

    Shanshan Feng, Feiyu Meng, Lisi Chen, Shuo Shang, and Yew Soon Ong. ROTAN: A Rotation- based Temporal Attention Network for Time-Specific Next POI Recommendation. InProceed- ings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 759–770, 2024

    2024

  8. [8]

    Graph-Flashback Network for Next Location Recommendation

    Xuan Rao, Lisi Chen, Yong Liu, Shuo Shang, Bin Yao, and Peng Han. Graph-Flashback Network for Next Location Recommendation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1463–1471, 2022

    2022

  9. [9]

    GETNext: Trajectory Flow Map Enhanced Trans- former for Next POI Recommendation

    Song Yang, Jiamou Liu, and Kaiqi Zhao. GETNext: Trajectory Flow Map Enhanced Trans- former for Next POI Recommendation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1144–1153, 2022

    2022

  10. [10]

    Adaptive Graph Representation Learning for Next POI Recommendation

    Zhaobo Wang, Yanmin Zhu, Chunyang Wang, Wenze Ma, Bo Li, and Jiadi Yu. Adaptive Graph Representation Learning for Next POI Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 393–402, 2023

    2023

  11. [11]

    Bi-Level Graph Structure Learning for Next POI Recommendation.IEEE Transactions on Knowledge and Data Engineering, 36(11):5695–5708, 2024

    Liang Wang, Shu Wu, Qiang Liu, Yanqiao Zhu, Xiang Tao, Mengdi Zhang, and Liang Wang. Bi-Level Graph Structure Learning for Next POI Recommendation.IEEE Transactions on Knowledge and Data Engineering, 36(11):5695–5708, 2024

    2024

  12. [12]

    Hyperbolic Variational Graph Auto-Encoder for Next POI Recommendation

    Yuwen Liu, Lianyong Qi, Xingyuan Mao, Weiming Liu, Fan Wang, Xiaolong Xu, Xuyun Zhang, Wanchun Dou, Xiaokang Zhou, and Amin Beheshti. Hyperbolic Variational Graph Auto-Encoder for Next POI Recommendation. InProceedings of the ACM on Web Conference, pages 3267–3275, 2025

    2025

  13. [13]

    Adaptive Location Hierarchy Learning for Long-Tailed Mobility Prediction

    Yu Wang, Junshu Dai, Yuchen Ying, Hanyang Yuan, Zunlei Feng, Tongya Zheng, and Min- gli Song. Adaptive Location Hierarchy Learning for Long-Tailed Mobility Prediction. In Proceedings of the ACM Web Conference, pages 6493–6504, 2026

    2026

  14. [14]

    Disentangled Graph Debiasing for Next POI Recommendation

    Hailun Zhou, Jiajie Xu, Qiaoming Zhu, and Chengfei Liu. Disentangled Graph Debiasing for Next POI Recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1779–1788, 2025. 10

    2025

  15. [15]

    Modeling User Activity Preference by Leveraging User Spatial Temporal Characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 45(1):129–142, 2015

    Dingqi Yang, Daqing Zhang, Vincent W Zheng, and Zhiyong Yu. Modeling User Activity Preference by Leveraging User Spatial Temporal Characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 45(1):129–142, 2015

    2015

  16. [16]

    When Does Compositional Structure Yield Composi- tional Generalization? A Kernel Theory

    Samuel Lippl and Kim Stachenfeld. When Does Compositional Structure Yield Composi- tional Generalization? A Kernel Theory. InThe 13th International Conference on Learning Representations, 2025

    2025

  17. [17]

    Predicting the Next Location: A Recurrent Model with Spatial and Temporal Contexts

    Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. Predicting the Next Location: A Recurrent Model with Spatial and Temporal Contexts. InProceedings of the 30th AAAI Conference on Artificial Intelligence, pages 194–200, 2016

    2016

  18. [18]

    REPLAY: Modeling Time-Varying Temporal Regularities of Human Mobility for Location Prediction over Sparse Trajectories.IEEE Transactions on Mobile Computing, 24(10):9428–9440, 2025

    Bangchao Deng, Bingqing Qu, Pengyang Wang, Dingqi Yang, Benjamin Fankhauser, and Philippe Cudre-Mauroux. REPLAY: Modeling Time-Varying Temporal Regularities of Human Mobility for Location Prediction over Sparse Trajectories.IEEE Transactions on Mobile Computing, 24(10):9428–9440, 2025

    2025

  19. [19]

    Disentangled And Personalized Representation Learning For Next Point-of-Interest Recommendation

    Xuan Rao, Shuo Shang, Lisi Chen, Renhe Jiang, and Peng Han. Disentangled And Personalized Representation Learning For Next Point-of-Interest Recommendation. InProceedings of the 34th International Joint Conference on Artificial Intelligence, pages 7697–7705, 2025

    2025

  20. [20]

    Next Point-of-Interest Recommendation with Temporal and Multi-Level Context Attention

    Ranzhen Li, Yanyan Shen, and Yanmin Zhu. Next Point-of-Interest Recommendation with Temporal and Multi-Level Context Attention. In2018 IEEE International Conference on Data Mining, pages 1110–1115, 2018

    2018

  21. [21]

    Geography-aware Sequential Location Recommendation

    Defu Lian, Yongji Wu, Yong Ge, Xing Xie, and Enhong Chen. Geography-aware Sequential Location Recommendation. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2009–2019, 2020

    2009

  22. [22]

    Spatio-temporal Hypergraph Learning for Next POI Recommendation

    Xiaodong Yan, Tengwei Song, Yifeng Jiao, Jianshan He, Jiaotuan Wang, Ruopeng Li, and Wei Chu. Spatio-temporal Hypergraph Learning for Next POI Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 403–412, 2023

    2023

  23. [23]

    Disentangled Contrastive Hypergraph Learning for Next POI Recommendation

    Yantong Lai, Yijun Su, Lingwei Wei, Tianqi He, Haitao Wang, Gaode Chen, Daren Zha, Qiang Liu, and Xingxing Wang. Disentangled Contrastive Hypergraph Learning for Next POI Recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1452–1462, 2024

    2024

  24. [24]

    Multifaceted Scenario-Aware Hypergraph Learning for Next POI Recommendation

    Yuxi Lin, Yongkang Li, Jie Xing, and Zipei Fan. Multifaceted Scenario-Aware Hypergraph Learning for Next POI Recommendation. InProceedings of the 40th AAAI Conference on Artificial Intelligence, pages 15269–15277, 2026

    2026

  25. [25]

    Bias and Debias in Recommender System: A Survey and Future Directions.ACM Transactions on Information Systems, 41:1–39, 2023

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. Bias and Debias in Recommender System: A Survey and Future Directions.ACM Transactions on Information Systems, 41:1–39, 2023

    2023

  26. [26]

    Causal Intervention for Leveraging Popularity Bias in Recommendation

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. Causal Intervention for Leveraging Popularity Bias in Recommendation. In Proceedings of the 44th international ACM SIGIR Conference on Research and Development in Information Retrieval, pages 11–20, 2021

    2021

  27. [27]

    Improving Long-tail Item Recommen- dation with Graph Augmentation

    Sichun Luo, Chen Ma, Yuanzhang Xiao, and Linqi Song. Improving Long-tail Item Recommen- dation with Graph Augmentation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 1707–1716, 2023

    2023

  28. [28]

    MELT: Mutual Enhancement of Long-Tailed User and Item for Sequential Recommendation

    Kibum Kim, Dongmin Hyun, Sukwon Yun, and Chanyoung Park. MELT: Mutual Enhancement of Long-Tailed User and Item for Sequential Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 68–77, 2023

    2023

  29. [29]

    Cross-Task Multimodal Rein- forcement for Long Tail Next POI Recommendation.IEEE Transactions on Multimedia, 26: 1996–2005, 2024

    Jiangfeng Du, Silin Zhou, Jie Yu, Peng Han, and Shuo Shang. Cross-Task Multimodal Rein- forcement for Long Tail Next POI Recommendation.IEEE Transactions on Multimedia, 26: 1996–2005, 2024. 11

    1996

  30. [30]

    How well does generative recommendation generalize?arXiv preprint arXiv:2603.19809, 2026

    Yijie Ding, Zitian Guo, Jiacheng Li, Letian Peng, Shuai Shao, Wei Shao, Xiaoqiang Luo, Luke Simon, Jingbo Shang, Julian McAuley, and Yupeng Hou. How Well Does Generative Recommendation Generalize?arXiv preprint arXiv:2603.19809, 2026

    work page arXiv 2026

  31. [31]

    IM-POI: Bridging ID and Multi-modal Gaps in Next POI Recommendation

    Siyuan Huang, Jiahui Jin, Xin Lin, Xigang Sun, and Yukun Ban. IM-POI: Bridging ID and Multi-modal Gaps in Next POI Recommendation. InProceedings of the 33rd ACM International Conference on Multimedia, pages 5979–5987, 2025

    2025

  32. [32]

    Not Too Little, Not Too Much: A Theoretical Analysis of Graph (Over) Smoothing

    Nicolas Keriven. Not Too Little, Not Too Much: A Theoretical Analysis of Graph (Over) Smoothing. InProceedings of the 36th International Conference on Neural Information Processing System, pages 2268–2281, 2022

    2022

  33. [33]

    Time-Aware Point-of-Interest Recommendation

    Quan Yuan, Gao Cong, Zongyang Ma, Aixin Sun, and Nadia Magnenat-Thalmann. Time-Aware Point-of-Interest Recommendation. InProceedings of The 36th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 363–372, 2013

    2013

  34. [34]

    Where Would I Go Next? Large Language Models as Human Mobility Predictors.arXiv preprint arXiv:2308.15197, 2023

    Xinglei Wang, Meng Fang, Zichao Zeng, and Tao Cheng. Where Would I Go Next? Large Language Models as Human Mobility Predictors.arXiv preprint arXiv:2308.15197, 2023

    work page arXiv 2023

  35. [35]

    Large Language Models for Next Point-of-Interest Recommendation

    Peibo Li, Maarten de Rijke, Hao Xue, Shuang Ao, Yang Song, and Flora D Salim. Large Language Models for Next Point-of-Interest Recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1463–1472, 2024

    2024

  36. [36]

    NextLo- cLLM: Location Semantics Modeling and Coordinate-Based Next Location Prediction with LLMs

    Shuai Liu, Ning Cao, Yile Chen, Yue Jiang, George Rosario Jagadeesh, and Gao Cong. NextLo- cLLM: Location Semantics Modeling and Coordinate-Based Next Location Prediction with LLMs. InProceedings of The International Workshop on Spatio-Temporal Data Intelligence and Foundation Models, The 34th ACM International Conference on Information and Knowledge Mana...

    2025

  37. [37]

    Generative Next POI Recommendation with Semantic ID

    Dongsheng Wang, Yuxi Huang, Shen Gao, Yifan Wang, Chengrui Huang, and Shuo Shang. Generative Next POI Recommendation with Semantic ID. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2904–2914, 2025

    2025

  38. [38]

    GeoGR: A Generative Retrieval Framework for Spatio-Temporal Aware POI Recommendation.arXiv preprint arXiv:2602.10411, 2026

    Fangye Wang, Haowen Lin, Yifang Yuan, Siyuan Wang, Xiaojiang Zhou, Song Yang, and Pengjie Wang. GeoGR: A Generative Retrieval Framework for Spatio-Temporal Aware POI Recommendation.arXiv preprint arXiv:2602.10411, 2026

    work page arXiv 2026

  39. [39]

    R3-VAE: Reference Vector-Guided Rating Residual Quantization VAE for Generative Recommendation

    Qiang Wan, Ze Yang, Dawei Yang, Ying Fan, Xin Yan, Siyang Liu, Yicong Liu, Chenwei Zhang, Wei Xu, Jiahao Qin, and Ke Wang. R3-V AE: Reference Vector-Guided Rating Residual Quantization V AE for Generative Recommendation.arXiv preprint arXiv:2604.11440, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    Comapoi: A collaborative multi-agent framework for next poi prediction bridging the gap between trajectory and language

    Lin Zhong, Lingzhi Wang, Xu Yang, and Qing Liao. Comapoi: A collaborative multi-agent framework for next poi prediction bridging the gap between trajectory and language. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1768–1778, 2025

    2025

  41. [41]

    MAS4POI: A Multi-Agent Col- laboration System for Next POI Recommendation

    Yuqian Wu, Yuhong Peng, Jiapeng Yu, and Raymond Lee. MAS4POI: A Multi-Agent Col- laboration System for Next POI Recommendation. InProceedings of the 29th Pacific-Asia Conference on Knowledge Discovery and Data Mining, pages 356–367. Springer, 2025. 12 A Additional Details of RECAP A.1 Notation Table 3 summarizes the symbols used in the paper. Table 3: Sym...

    work page arXiv 2025