Pith. sign in

REVIEW 4 major objections 7 minor 65 references

Mixture-of-Experts for Personalized and Semantic-Aware Next Location Prediction

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Mixture-of-experts model tops next-location prediction on 3 cities

desk verdict Clever architecture and plausible supervised results, but the zero-shot table reuses supervised baseline numbers, so the transfer claim collapses until fixed. read the letter →

arxiv 2505.24597 v1 pith:LJ725DSH submitted 2025-05-30 cs.AI

Shuai Liu , Ning Cao , Yile Chen , Yue Jiang , Gao Cong This is my paper
classification cs.AI
keywords nextlocationpredictionmixtureofexpertslargelanguagemodelssemanticsuserpersonalizationhumanmobilityzero-shottransferinterpretability
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 claims that next-location prediction is held back by two modeling gaps: each location is represented by a single embedding even though real places serve many functions, and all users share one parameter set even though different personas move differently. To close those gaps, it proposes NextLocMoE, an LLM-based architecture with a Location Semantics MoE that routes each current location to a small set of function-specific expert embeddings, and a Personalized MoE that routes each user to a few behavioral persona experts inside the Transformer's feedforward layers. The paper reports that this design beats all tested baselines on three real urban mobility datasets, achieves the best results in zero-shot transfer to an unseen city, and is about 600 times faster at inference than the strongest prior LLM mobility model. A case study argues the activated experts are interpretable, matching a user's persona and a location's functional role.

What carries the argument

The load-bearing mechanism is the dual-level MoE architecture combined with a history-aware router. Location Semantics MoE operates at the embedding level with top-k routing among $K_f$ function experts, each a linear projection of the spatial coordinates initialized from LLM-encoded location-function descriptions; the function-aware embedding $e^{enhanced}_{xy} = e_{xy} + e^{func}_{xy}$ treats the general spatial embedding as a shared expert. Personalized MoE operates inside selected Transformer layers, replacing the feedforward networks with $K_p$ user-persona experts selected until their cumulative routing probability passes a threshold $\tau$, then aggregated by weighted sum. Both routers receive $h_{hist}$, the TCN-encoded historical trajectory, which is what makes the routing history-aware. The model is trained with an $\ell^2$ regression loss on predicted coordinates plus entropy regularization on routing, and inference maps the predicted coordinate to a location ID via KD-Tree nearest-neighbor search.

What would settle it

Replace the coordinate-regression plus KD-Tree retrieval with a standard softmax classification head over the candidate location IDs, keeping the dual MoE, training data, and all other components identical; if the Hit@1 and Hit@5 margins over the baselines collapse or reverse, the retrieval protocol rather than the MoE design explains the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a dual-level Mixture-of-Experts decomposition of a next-location predictor improves accuracy and generalizability simultaneously. The Location Semantics MoE represents a location not as one vector but as a shared spatial embedding plus a weighted sum of top-k function experts (education, commercial, entertainment, and so on), where the experts are initialized with LLM-encoded natural-language descriptions. The Personalized MoE replaces feedforward networks in the upper LLM layers with persona experts, activated by a confidence threshold rather than a fixed count, and a history-aware router feeds a TCN-encoded long-term trajectory into both routers. The paper claims this yields consistent state-of-the-art Hit@1, Hit@5, and Hit@10 on Kumamoto, Shanghai, and Singapore, superior zero-shot transfer between cities, and a 600-fold inference speedup over Llama-Mob. Ablations show both MoE modules contribute, and removing the persona module is especially harmful to zero-shot transfer.

Load-bearing premise

The results assume that predicting a continuous coordinate and then retrieving the nearest location IDs by KD-Tree is equivalent to directly classifying the next location among all candidate IDs, which is the protocol every baseline uses; if the retrieval step inflates the top-k hits, the reported accuracy gains would not survive a matched comparison.

Editorial extensions

If this is right

  • On all three datasets, Hit@1 rises to 17.77% (Kumamoto), 64.93% (Shanghai), and 9.733% (Singapore), beating the strongest LLM baseline Llama-Mob by roughly 2 to 1.2 percentage points.
  • In zero-shot transfer to Kumamoto, NextLocMoE reaches 16.02% Hit@1 when trained on Shanghai and 15.81% when trained on Singapore, ahead of NextLocLLM's 13.14% and 11.73%.
  • Inference on the Kumamoto test set completes in 268 seconds, about 600x faster than Llama-Mob (158,688 s) and 120x faster than LLMMob (33,408 s), because the unified architecture supports batch processing.
  • Ablations show both modules matter: removing Personalized MoE drops Shanghai Hit@1 from 64.92% to 33.68% in the fully-supervised setting, and zero-shot Hit@1 collapses to 1.437%, while removing Location Semantics MoE drops it to 59.72% and to 5.519% in zero-shot.
  • The activated experts are interpretable: in the case study, a trajectory around academic zones routes to education and entertainment location experts and student and teacher persona experts, while a commercial-tourist trajectory routes to entertainment and commercial experts and a tourist expert.

Reading between the lines

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

  • Because prediction is coordinate regression with KD-Tree retrieval, the model is decoupled from city-specific location IDs; this is likely the reason zero-shot transfer works, and it suggests the method could apply to any city whose grid coordinates are available without retraining the ID head.
  • The confidence-threshold router implies a testable robustness property: adding noise to a user's recent trajectory should activate more persona experts, and prediction variance should grow less than with a fixed top-k router; the paper does not run this experiment but its mechanism predicts it.
  • Since location-function and persona experts are initialized from LLM-encoded text, a lightweight adaptation path exists: re-encoding new descriptions for a new city's categories should let the architecture specialize without weight updates; this is an implicit affordance the paper does not test.
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.

Referee Report

4 major / 7 minor

Summary. The paper proposes NextLocMoE, an LLM-based framework for next location prediction that combines two Mixture-of-Experts modules: a Location Semantics MoE that augments coordinate embeddings with function-aware expert embeddings initialized from LLM-encoded descriptions of location categories, and a Personalized MoE that replaces upper-layer FFNs with user-group experts initialized from LLM-encoded persona descriptions. A history-aware router uses TCN-encoded long-term trajectories to guide expert selection in both modules. The model is trained with a coordinate regression loss plus entropy regularization, and at inference predicts a continuous coordinate that is mapped to a discrete location ID via KD-Tree nearest-neighbor search. The authors report fully-supervised results on Kumamoto, Shanghai, and Singapore, zero-shot transfer results on Kumamoto, inference-time comparisons, hyperparameter sensitivity, expert activation statistics, and a case study, claiming consistent state-of-the-art performance and large speedups over LLM baselines.

Significance. If the empirical claims are upheld, the architectural ideas are valuable: the paper is, to my knowledge, the first to integrate dual-level MoE into an LLM for next location prediction, and the use of LLM-encoded semantic priors for expert initialization is a plausible and interpretable design. The reported inference efficiency (268 s vs. 158688 s for Llama-Mob on the same test set) is a notable practical strength, as is the explicit attempt to model multi-functional location semantics and behavioral heterogeneity without relying on user IDs. However, two load-bearing evaluation issues currently prevent the results from being accepted at face value: the zero-shot comparison in Table 2 appears to reuse fully-supervised baseline numbers, and the coordinate-regression-plus-KD-tree retrieval protocol is not shown to be comparable to the ID-classification protocol used by all baselines. The paper also omits reproducibility-critical hyperparameters for two of three datasets and reports no variance or significance information.

major comments (4)
  1. [§5.2.2, Table 2] The zero-shot comparison is internally inconsistent as printed. The rows for Llama-Mob, LLMMob, and ZS-NL in Table 2 (Hit@1/Hit@5/Hit@10 = 15.78/33.55/43.42, 10.95/25.54/35.77, and 8.811/22.97/31.76 on Kumamoto) are numerically identical to the fully-supervised results for the same models in Table 1. A zero-shot evaluation tests models trained on other cities with no Kumamoto fine-tuning, so these numbers cannot equal fully-supervised scores. The only zero-shot entries with distinct transfer-source labels are NextlocLLM and NextLocMoE. As printed, Table 2 does not establish zero-shot superiority because it compares transferred NextLocMoE against baseline numbers that appear to be supervised. The authors must either correct the baseline rows with actual zero-shot evaluations or clearly state a different protocol; without this fix the zero-shot claim is unsupported.
  2. [§4.7, §5.2] The evaluation protocol is not commensurable between NextLocMoE and the baselines. Baselines are trained and evaluated as ID classifiers, producing a top-k over location IDs, whereas NextLocMoE regresses a continuous spatial coordinate and then retrieves the k nearest locations via KD-Tree. Hit@k under these two protocols can differ systematically; for example, coordinate regression can benefit from spatial locality even when the exact ID is wrong, while ID classification does not. The paper does not justify that the retrieval step is equivalent to the baseline protocol, nor does it report an ID-classification variant of NextLocMoE or a coordinate-retrieval variant of the baselines. This affects the central fully-supervised SOTA claim in Tables 1 and 2 and should be resolved with an ablation or a clear argument that the two protocols are interchangeable.
  3. [App. G, §5.1] Reproducibility of the main empirical claim is incomplete. Table 7 reports hyperparameters only for Kumamoto; no corresponding settings are given for Shanghai and Singapore, even though the fully-supervised results in Table 1 and the zero-shot results in Table 2 depend on those datasets. Additionally, no error bars, number of runs, or statistical significance tests are reported, so it is unclear whether the reported SOTA margins (e.g., 17.77% vs. 15.78% Hit@1 on Kumamoto) are stable. The authors should provide full hyperparameter tables for all datasets and report variance or significance information at least for the headline comparisons.
  4. [§4.3, §4.4, App. B, App. C] The five hand-defined location functions and eleven hand-defined user personas are treated as sufficient semantic priors, but the paper does not test sensitivity to these choices. Since expert initialization and routing are core to the claimed improvements, the authors should report at least a small ablation varying the number of location-function categories or personas, or otherwise justify that the specific hand-defined sets are not a critical factor in the results. This is not a circularity issue, but it is a sensitivity question for a load-bearing design component.
minor comments (7)
  1. [Table 1] The column header "Sinapore" is a typo for "Singapore".
  2. [Table 2] The spelling "NextlocLLM" is inconsistent with "NextLocLLM" used elsewhere, including in the same table for the transfer-source labels.
  3. [§5.2.4] The case study refers to "Fig. ??", which is unresolved in the manuscript.
  4. [§4.1] The reference "App. 4" for the task-specific prompt should be "App. D", since Appendix D is titled "Prompt Prefix".
  5. [App. E] The first sentence of Appendix E says "NextLocLoE" instead of "NextLocMoE".
  6. [App. F] The entry for LLMMob is written as "LLmMob", which is inconsistent with the capitalization used throughout the paper.
  7. [Fig. 3] The axes and exact metric plotted in Fig. 3(a) are not labeled; adding axis labels and a legend would make the hyperparameter sensitivity result interpretable.

Circularity Check

1 steps flagged · score 6.0 of 10

Zero-shot SOTA comparison reuses fully-supervised baseline numbers, so the zero-shot 'prediction' reduces to the paper's own supervised table.

  1. fitted input called prediction [Section 5.2.2 (Zero-shot Prediction Comparison), Table 2 vs Section 5.2.1, Table 1]
    "Table 1 (fully-supervised, Kumamoto): Llama-Mob 15.78% 33.55% 43.42%; LLMMob 10.95% 25.54% 35.77%. Table 2 (zero-shot, Kumamoto): Llama-Mob 15.78% 33.55% 43.42%; LLMMob 10.95% 25.54% 35.77%."

    Section 5.2.2 defines the zero-shot protocol as testing models 'directly tested using parameters trained on other cities without any fine-tuning,' yet the zero-shot rows for Llama-Mob and LLMMob are numerically identical to their fully-supervised rows in Table 1. A model trained on Kumamoto (Table 1) and a model trained on other cities (Table 2) cannot produce identical Hit@1/Hit@5/Hit@10 unless the Table 2 entries were copied from Table 1.

full rationale

The paper is an empirical systems paper: NextLocMoE is trained on real mobility data with a coordinate-regression objective and evaluated with KD-Tree nearest-neighbor retrieval; none of the architectural components (Location Semantics MoE, Personalized MoE, history-aware router) is defined in terms of the Hit@k metric, and no fitted hyperparameter is renamed as a prediction. The only reduction I can exhibit is in the zero-shot comparison: Table 2's Llama-Mob and LLMMob rows are numerically identical to their fully-supervised rows in Table 1, even though Section 5.2.2 defines zero-shot as testing parameters trained on other cities without fine-tuning. As printed, those baseline rows cannot be genuine zero-shot measurements; they are the supervised numbers reused as the zero-shot comparison, so the claim that NextLocMoE 'achieves the best across all metrics' in zero-shot is not established by an independent prediction. This is a concrete, quotable reduction. The self-citation [12] is used only as a baseline and does not carry a load-bearing argument, and the LLM-initialized expert categories are priors rather than recycled predictions. If the duplicated entries are corrected, the residual circularity would be 0-2.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the hand-defined category sets, the comparability of coordinate-regression evaluation, and a set of hyperparameters reported only for one dataset. No new physical entities are introduced; the main epistemic burden is the unvalidated evaluation protocol.

free parameters (7)
  • Kf (number of location function experts) = 5
    Number of location function categories in App. B, chosen by hand without data-driven validation.
  • Kp (number of user group experts) = 11
    Number of user personas in App. C, chosen by hand without validation against real user roles.
  • tau (expert activation confidence threshold) = 0.8
    Reported in Table 7 for Kumamoto; presumably tuned, not reported for Shanghai/Singapore.
  • lambda (entropy regularization weight) = 300
    Reported in Table 7 for Kumamoto; other datasets' values not given.
  • top-k for Location Semantics MoE
    Number of active function experts is not specified in the paper.
  • Historical/current trajectory lengths M, N = M=40, N=5
    Set following prior work; no sensitivity analysis for these values.
  • LLM layer split L1, L2 = L1=8, L2=4
    Selected via hyperparameter sensitivity in Fig. 3(a); not validated across datasets.
assumptions (5)
  • domain assumption Coordinate regression with KD-tree retrieval is comparable to direct location-ID classification.
    The paper does not discuss whether Hit@k under nearest-coordinate retrieval yields the same evaluation as ID classification used by all baselines.
  • ad hoc to paper The five hand-defined location function categories are sufficient to capture location semantics.
    Categories in App. B are not validated against ground-truth functions or data.
  • ad hoc to paper The eleven hand-defined user personas are sufficient to model behavioral heterogeneity.
    Personas in App. C are not validated against actual user roles or clusters.
  • domain assumption LLM-encoded natural language descriptions provide useful inductive bias for expert initialization.
    No analysis of sensitivity to description wording or of the transfer from language embeddings to spatial semantics.
  • domain assumption A 40-record historical trajectory and a 5-record current trajectory are sufficient for prediction.
    No sensitivity analysis is reported for M and N.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixture-of-Experts for Personalized and Semantic-Aware Next Location Prediction." pith.science (2026). https://pith.science/paper/LJ725DSH

@misc{pith2026250524597,
  author       = {Pith},
  title        = {Pith review of: Mixture-of-Experts for Personalized and Semantic-Aware Next Location Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJ725DSH}},
  note         = {Machine review of arXiv:2505.24597}
}
read the original abstract

Next location prediction plays a critical role in understanding human mobility patterns. However, existing approaches face two core limitations: (1) they fall short in capturing the complex, multi-functional semantics of real-world locations; and (2) they lack the capacity to model heterogeneous behavioral dynamics across diverse user groups. To tackle these challenges, we introduce NextLocMoE, a novel framework built upon large language models (LLMs) and structured around a dual-level Mixture-of-Experts (MoE) design. Our architecture comprises two specialized modules: a Location Semantics MoE that operates at the embedding level to encode rich functional semantics of locations, and a Personalized MoE embedded within the Transformer backbone to dynamically adapt to individual user mobility patterns. In addition, we incorporate a history-aware routing mechanism that leverages long-term trajectory data to enhance expert selection and ensure prediction stability. Empirical evaluations across several real-world urban datasets show that NextLocMoE achieves superior performance in terms of predictive accuracy, cross-domain generalization, and interpretability

Figures

Figures reproduced from arXiv: 2505.24597 by the authors.

Figure 1
Figure 1. Overall architecture of NextLocMoE, a Mixture-of-Experts enhanced LLM framework for [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the two expert modules. (a) Location Semantics MoE, (b) Personalized MoE. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) Hyperparameter sensitivity; (b) Personalized expert activation analysis [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Prompt prefix used in NextLocMoE. E Dataset Description We use three real-world mobility datasets to validate the effectiveness of NextLocLoE, and the detailed descrip￾tions of these datasets are as follows: Kumamoto 1 This is an open-source and anonymized dataset of h…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 47 canonical work pages

  1. [1]

    Real-time transportation prediction correction using reconstruction error in deep learning

    Shuai Liu, Guojie Song, and Wenhao Huang. Real-time transportation prediction correction using reconstruction error in deep learning. ACM Transactions on Knowledge Discovery from Data (TKDD), 14(2):1–20, 2020

    work page 2020

  2. [2]

    Mcn4rec: Multi-level collaborative neural network for next location recommendation

    Shuzhe Li, Wei Chen, Bin Wang, Chao Huang, Yanwei Yu, and Junyu Dong. Mcn4rec: Multi-level collaborative neural network for next location recommendation. ACM Transactions on Information Systems, 42(4):1–26, 2024

    work page 2024

  3. [3]

    Cdrp3: Cascade deep reinforcement learning for urban driving safety with joint perception, prediction, and planning

    Yuxiang Yang, Fenglong Ge, Jinlong Fan, Jufeng Zhao, and Zhekang Dong. Cdrp3: Cascade deep reinforcement learning for urban driving safety with joint perception, prediction, and planning. IEEE Transactions on Intelligent Transportation Systems, 2024

    work page 2024

  4. [4]

    Empirical evaluation of gated recurrent neural networks on sequence modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014

    arXiv 2014

  5. [5]

    Long short-term memory

    Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012

    work page 2012

  6. [6]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  7. [7]

    Context-aware multi-head self-attentional neural network model for next location prediction.Transportation Research Part C: Emerging Technologies, 156:104315, 2023

    Ye Hong, Yatao Zhang, Konrad Schindler, and Martin Raubal. Context-aware multi-head self-attentional neural network model for next location prediction.Transportation Research Part C: Emerging Technologies, 156:104315, 2023

    work page 2023

  8. [8]

    Cllp: Contrastive learning framework based on latent preferences for next poi recommendation

    Hongli Zhou, Zhihao Jia, Haiyang Zhu, and Zhizheng Zhang. Cllp: Contrastive learning framework based on latent preferences for next poi recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1473–1482, 2024

    work page 2024

Show all 65 references
  1. [9]

    Getnext: trajectory flow map enhanced transformer for next poi recommendation

    Song Yang, Jiamou Liu, and Kaiqi Zhao. Getnext: trajectory flow map enhanced transformer for next poi recommendation. In Proceedings of the 45th International ACM SIGIR Conference on research and development in information retrieval, pages 1144–1153, 2022

    2022

  2. [10]

    Instruction- tuning llama-3-8b excels in city-scale mobility prediction

    Peizhi Tang, Chuang Yang, Tong Xing, Xiaohang Xu, Renhe Jiang, and Kaoru Sezaki. Instruction- tuning llama-3-8b excels in city-scale mobility prediction. In Proceedings of the 2nd ACM SIGSPATIAL International Workshop on Human Mobility Prediction Challenge, pages 1–4, 2024

    2024

  3. [11]

    Where would i go next? large language models as human mobility predictors

    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

    2023 arXiv

  4. [12]

    nextlocllm: next location prediction using llms

    Shuai Liu, Ning Cao, Yile Chen, Yue Jiang, and Gao Cong. nextlocllm: next location prediction using llms. arXiv preprint arXiv:2410.09129, 2024

    2024

  5. [13]

    Harder tasks need more experts: Dynamic routing in moe models

    Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. arXiv preprint arXiv:2403.07652, 2024

    2024

  6. [14]

    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. In Proceedings of the 2018 world wide web conference, pages 1459–1468, 2018

    2018

  7. [15]

    Seaget: Seasonal and active hours guided graph enhanced transformer for the next poi recommendation

    Alif Al Hasan and Md Musfique Anwar. Seaget: Seasonal and active hours guided graph enhanced transformer for the next poi recommendation. Array, page 100385, 2025

    2025

  8. [16]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  9. [17]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  10. [18]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  11. [19]

    Gshard: Scaling giant models with conditional computation and automatic sharding

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020

    2006 arXiv

  12. [20]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1–39, 2022. 10

    2022

  13. [21]

    Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models

    Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024

    2024 arXiv

  14. [22]

    Time-moe: Billion-scale time series foundation models with mixture of experts

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024

    2024 arXiv

  15. [23]

    Moirai-moe: Empowering time series foundation models with sparse mixture of experts

    Xu Liu, Juncheng Liu, Gerald Woo, Taha Aksu, Yuxuan Liang, Roger Zimmermann, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. Moirai-moe: Empowering time series foundation models with sparse mixture of experts. arXiv preprint arXiv:2410.10469, 2024

    arXiv

  16. [24]

    Graph mixture of experts and memory- augmented routers for multivariate time series anomaly detection

    Xiaoyu Huang, Weidong Chen, Bo Hu, and Zhendong Mao. Graph mixture of experts and memory- augmented routers for multivariate time series anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 17476–17484, 2025

    2025

  17. [25]

    Temporal convolutional networks for action segmentation and detection

    Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 156–165, 2017

    2017

  18. [26]

    Layer by layer: Uncovering hidden representations in language models

    Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. arXiv preprint arXiv:2502.02013, 2025

    2025 arXiv

  19. [27]

    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. In Proceedings of the web conference 2021, pages 2177–2185, 2021

    2021

  20. [28]

    Mfnp: A meta- optimized model for few-shot next poi recommendation

    Huimin Sun, Jiajie Xu, Kai Zheng, Pengpeng Zhao, Pingfu Chao, and Xiaofang Zhou. Mfnp: A meta- optimized model for few-shot next poi recommendation. In IJCAI, volume 2021, pages 3017–3023, 2021

    2021

  21. [29]

    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. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 759–770, 2024

    2024

  22. [30]

    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. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1463–1472, 2024

    2024

  23. [31]

    Large language models are zero-shot next location predictors

    Ciro Beneduce, Bruno Lepri, and Massimiliano Luca. Large language models are zero-shot next location predictors. IEEE Access, 2025

    2025

  24. [32]

    Llama-moe: Building mixture-of-experts from llama with continual pre-training

    Tong Zhu, Xiaoye Qu, Daize Dong, Jiacheng Ruan, Jingqi Tong, Conghui He, and Yu Cheng. Llama-moe: Building mixture-of-experts from llama with continual pre-training. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15913–15923, 2024

    2024

  25. [33]

    A survey on next location prediction techniques, applications, and challenges

    Ayele Gobezie Chekol and Marta Sintayehu Fufa. A survey on next location prediction techniques, applications, and challenges. EURASIP Journal on Wireless Communications and Networking, 2022(1):29, 2022

    2022

  26. [34]

    Mobility prediction in cellular networks: A survey

    Nilakshee Rajule, Mithra Venkatesan, Radhika Menon, and Anju Kulkarni. Mobility prediction in cellular networks: A survey. In 2023 International Conference on Recent Trends in Electronics and Communication (ICRTEC), pages 1–9. IEEE, 2023

    2023

  27. [35]

    Mobility prediction: A survey on state-of-the-art schemes and future applications

    Hongtao Zhang and Lingcheng Dai. Mobility prediction: A survey on state-of-the-art schemes and future applications. IEEE access, 7:802–822, 2018

    2018

  28. [36]

    Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network

    Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena, 404:132306, 2020

    2020

  29. [37]

    Sasrm: a semantic and attention spatio-temporal recurrent model for next location prediction

    Xu Zhang, Boming Li, Chao Song, Zhengwen Huang, and Yan Li. Sasrm: a semantic and attention spatio-temporal recurrent model for next location prediction. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2020

    2020

  30. [38]

    Beyond the limits of predictability in human mobility prediction: Context-transition predictability

    Chao Zhang, Kai Zhao, and Meng Chen. Beyond the limits of predictability in human mobility prediction: Context-transition predictability. IEEE Transactions on Knowledge and Data Engineering, 35(5):4514– 4526, 2022

    2022

  31. [39]

    Secdm: A secure and lossless human mobility prediction system

    Lin Liu, Shaojing Fu, Xuelun Huang, Yuchuan Luo, Xuyun Zhang, and Kim-Kwang Raymond Choo. Secdm: A secure and lossless human mobility prediction system. IEEE Transactions on Services Comput- ing, 17(4):1793–1805, 2024

    2024

  32. [40]

    Mobility prediction via sequential trajectory disentanglement (student abstract)

    Jinyu Hong, Fan Zhou, Qiang Gao, Ping Kuang, and Kunpeng Zhang. Mobility prediction via sequential trajectory disentanglement (student abstract). In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 16230–16231, 2023. 11

    2023

  33. [41]

    An lstm-based mobility prediction mechanism in the icn-based vehicular networks

    Kamrul Hasan and Seong-Ho Jeong. An lstm-based mobility prediction mechanism in the icn-based vehicular networks. In 2022 27th Asia Pacific Conference on Communications (APCC), pages 244–246. IEEE, 2022

    2022

  34. [42]

    Multiwave covid-19 prediction from social awareness using web search and mobility data

    Jiawei Xue, Takahiro Yabe, Kota Tsubouchi, Jianzhu Ma, and Satish Ukkusuri. Multiwave covid-19 prediction from social awareness using web search and mobility data. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4279–4289, 2022

    2022

  35. [43]

    Trajgraph: A dual-view graph transformer model for effective next location recommendation

    Jiafeng Zhao, Hao Ni, Canghong Jin, Tongya Zheng, Longxiang Shi, and Xiaoliang Wang. Trajgraph: A dual-view graph transformer model for effective next location recommendation. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2024

    2024

  36. [44]

    Next point-of- interest recommendation with adaptive graph contrastive learning

    Xuan Rao, Renhe Jiang, Shuo Shang, Lisi Chen, Peng Han, Bin Yao, and Panos Kalnis. Next point-of- interest recommendation with adaptive graph contrastive learning. IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  37. [45]

    Integrating personalized spatio-temporal clustering for next poi recommendation

    Chao Song, Zheng Ren, and Li Lu. Integrating personalized spatio-temporal clustering for next poi recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 12550–12558, 2025

    2025

  38. [46]

    Pre-training context and time aware location embeddings from spatial-temporal trajectories for user next location prediction

    Yan Lin, Huaiyu Wan, Shengnan Guo, and Youfang Lin. Pre-training context and time aware location embeddings from spatial-temporal trajectories for user next location prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 4241–4248, 2021

    2021

  39. [47]

    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. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mini...

    2024

  40. [48]

    Enhancing long-and short- term representations for next poi recommendations via frequency and hierarchical contrastive learning

    Jiajie Chen, Yu Sang, Peng-Fei Zhang, Jiaan Wang, Jianfeng Qu, and Zhixu Li. Enhancing long-and short- term representations for next poi recommendations via frequency and hierarchical contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volum...

    2025

  41. [49]

    Improving next location recommendation services with spatial-temporal multi-group contrastive learning

    Zhixuan Jia, Yushun Fan, Jia Zhang, Chunyu Wei, Ruyu Yan, and Xing Wu. Improving next location recommendation services with spatial-temporal multi-group contrastive learning. IEEE Transactions on Services Computing, 16(5):3467–3478, 2023

    2023

  42. [50]

    Deep sequential multi-task modeling for next check-in time and location prediction

    Wenwei Liang, Wei Zhang, and Xiaoling Wang. Deep sequential multi-task modeling for next check-in time and location prediction. In International Conference on Database Systems for Advanced Applications, pages 353–357. Springer, 2019

    2019

  43. [51]

    Sanmove: next location recommendation via self-attention network

    Bin Wang, Huifeng Li, Le Tong, Qian Zhang, Sulei Zhu, and Tao Yang. Sanmove: next location recommendation via self-attention network. Data Technologies and Applications, 57(3):330–343, 2023

    2023

  44. [52]

    Tcsa-net: A temporal-context-based self- attention network for next location prediction

    Guiming Sun, Heng Qi, Yanming Shen, and Baocai Yin. Tcsa-net: A temporal-context-based self- attention network for next location prediction. IEEE Transactions on Intelligent Transportation Systems, 23(11):20735–20745, 2022

    2022

  45. [53]

    Taming the long tail in human mobility prediction

    Xiaohang Xu, Renhe Jiang, Chuang Yang, Kaoru Sezaki, et al. Taming the long tail in human mobility prediction. Advances in Neural Information Processing Systems, 37:54748–54771, 2024

    2024

  46. [54]

    Physics-informed spatio-temporal model for human mobility prediction

    Quanyan Gao, Chao Li, and Qinmin Yang. Physics-informed spatio-temporal model for human mobility prediction. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 409–425. Springer, 2024

    2024

  47. [55]

    Towards an event-aware urban mobility prediction system

    Zhaonan Wang, Renhe Jiang, Zipei Fan, Xuan Song, and Ryosuke Shibasaki. Towards an event-aware urban mobility prediction system. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 1303–1304, 2023

    2023

  48. [56]

    Human mobility prediction challenge: Next location prediction using spatiotemporal bert

    Haru Terashima, Naoki Tamura, Kazuyuki Shoji, Shin Katayama, Kenta Urano, Takuro Yonezawa, and Nobuo Kawaguchi. Human mobility prediction challenge: Next location prediction using spatiotemporal bert. In Proceedings of the 1st International Workshop on the Human Mobility Predi...

    2023

  49. [57]

    Exploring the potential of deep regression model for next-location prediction

    Pushpak Shukla and Shailendra Shukla. Exploring the potential of deep regression model for next-location prediction. Knowledge and Information Systems, 66(7):4093–4124, 2024

    2024

  50. [58]

    Agentmove: A large language model based agentic framework for zero-shot next location prediction

    Jie Feng, Yuwei Du, Jie Zhao, and Yong Li. Agentmove: A large language model based agentic framework for zero-shot next location prediction. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Lang...

    2025

  51. [59]

    Causalmob: Causal human mobility prediction with llms-derived human intentions toward public events

    Xiaojie Yang, Hangli Ge, Jiawei Wang, Zipei Fan, Renhe Jiang, Ryosuke Shibasaki, and Noboru Koshizuka. Causalmob: Causal human mobility prediction with llms-derived human intentions toward public events. arXiv preprint arXiv:2412.02155, 2024. 12

    2024 arXiv

  52. [60]

    A closer look into mixture-of-experts in large language models

    Ka Man Lo, Zeyu Huang, Zihan Qiu, Zili Wang, and Jie Fu. A closer look into mixture-of-experts in large language models. arXiv preprint arXiv:2406.18219, 2024

    2024 arXiv

  53. [61]

    Pmoe: Progressive mixture of experts with asymmetric transformer for continual learning

    Min Jae Jung and JooHee Kim. Pmoe: Progressive mixture of experts with asymmetric transformer for continual learning. arXiv preprint arXiv:2407.21571, 2024

    2024 arXiv

  54. [62]

    Moe jetpack: From dense checkpoints to adaptive mixture of experts for vision tasks

    Xingkui Zhu, Yiran Guan, Dingkang Liang, Yuchao Chen, Yuliang Liu, and Xiang Bai. Moe jetpack: From dense checkpoints to adaptive mixture of experts for vision tasks. Advances in Neural Information Processing Systems, 37:12094–12118, 2024

    2024

  55. [63]

    Not all experts are equal: Efficient expert pruning and skipping for mixture-of-experts large language models

    Xudong Lu, Qi Liu, Yuhui Xu, Aojun Zhou, Siyuan Huang, Bo Zhang, Junchi Yan, and Hongsheng Li. Not all experts are equal: Efficient expert pruning and skipping for mixture-of-experts large language models. arXiv preprint arXiv:2402.14800, 2024

    2024 arXiv

  56. [64]

    Multimodal instruction tuning with conditional mixture of lora

    Ying Shen, Zhiyang Xu, Qifan Wang, Yu Cheng, Wenpeng Yin, and Lifu Huang. Multimodal instruction tuning with conditional mixture of lora. arXiv preprint arXiv:2402.15896, 2024

    2024 arXiv

  57. [65]

    Large language models as urban residents: An llm agent framework for personal mobility generation

    W ANG JIAWEI, Renhe Jiang, Chuang Yang, Zengqing Wu, Ryosuke Shibasaki, Noboru Koshizuka, Chuan Xiao, et al. Large language models as urban residents: An llm agent framework for personal mobility generation. Advances in Neural Information Processing Systems, 37:124547–124574, ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.