Pith. sign in

REVIEW 4 major objections 3 minor 84 references

Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning

T0 review · 4 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Trajectory embeddings improve when routes are read as decisions shaped by the surrounding city, not as passive GPS traces.

desk verdict Nice framework with genuine novelty; headline claim overreaches given one negative metric and missing error bars. read the letter →

arxiv 2510.14819 v3 pith:CVDBQK7Z submitted 2025-10-16 cs.CV cs.LG

Ji Cao , Yu Wang , Tongya Zheng , Jie Song , Qinghong Guo , Zujie Ren , Canghong Jin , Gang Chen
show 1 more author
Mingli Song
This is my paper
classification cs.CVcs.LG
keywords trajectoryrepresentationlearningroutechoicemodelingpointofinterestlargelanguagemodelsmixture-of-expertscontrastiveroadnetworkembeddingurbanmobility
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 argues that a GPS trace is not just a sequence of positions but the visible outcome of a series of route choices, each conditioned on the surrounding urban environment. Most existing trajectory representation learning treats trajectories as passive spatiotemporal sequences, missing that behavioral layer. CORE instead builds a context-enriched road network by using a large language model to distill semantics from nearby point-of-interest distributions, then encodes each step as a choice between the taken next road segment and adjacent segments not taken, mediated by a mixture-of-experts router and navigational cues. On four real-world datasets and six downstream tasks, the paper reports consistent wins over twelve existing methods, with an average gain over the best baseline of roughly 10%. If this holds, trajectory learning should model decision context, not just geometry.

What carries the argument

Context-aware route choice semantics: each step of a trajectory is treated as a choice among adjacent road segments conditioned on the local urban environment. The load-bearing machinery is (1) an Environment Perception Module that uses a large language model to turn nearby POI distributions into textual descriptions, embeds them, propagates them across the road network with a graph attention network, and fuses them with coarse-grained hotspot semantics via gating; (2) a Route Choice Encoder that, at each segment, encodes the chosen next segment and the rejected adjacent segments, with a mixture-of-experts router selecting specialized subnetworks based on the current state plus navigational

What would settle it

Compare a CORE variant trained with real POI data against one trained with the same POI labels randomly shuffled across road segments. If shuffling does not degrade downstream metrics, the claimed environmental-semantics mechanism is not the source of the gains. A more localized test: hold out all POIs for a random subset of road segments and check whether road-label prediction on those segments drops relative to covered segments.

Watch

Extended reading notes

Core claim

CORE's central claim is that injecting context-aware route choice semantics as an inductive bias improves trajectory representations at every evaluated level. The framework first constructs a context-enriched road network: fine-grained road-segment semantics come from LLM-generated descriptions of nearby POIs, propagated from the top 20% busiest segments across a graph attention network, and coarse-grained urban functional structure comes from LLM descriptions of POI hotspots on a 1 km grid, both fused with road attributes through gating. A Route Choice Encoder then represents each decision step by comparing the selected next road segment against unselected neighbors, using a mixture-of-expe

Load-bearing premise

POI distributions are an adequate observable proxy for the urban decision-making context; if the places near a road do not reflect what actually drives drivers' choices, the context-enriched network loses its value.

Editorial extensions

If this is right

  • Trajectory representation learning should adopt behavioral modeling rather than purely geometric sequence encoding if these results hold.
  • The largest reported gains appear on road segment classification and similar-trajectory retrieval, suggesting the embeddings carry fine-grained route-choice distinctions that geometry alone misses.
  • LLM-generated semantic maps are described as cheap enough for city scale, with a reported cost of about $2 and two hours for Beijing, making the context-enriched network practical.
  • The model retains strong performance under small training-data fractions, indicating the behavioral inductive bias reduces data hunger.
  • Trajectory generation also improves substantially, so the same semantics used for retrieval and prediction can produce realistic movements.

Reading between the lines

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

  • The authors leave implicit that the mechanism can be tested against other context sources such as congestion, tolls, or road quality; where those factors dominate route choice, POI-derived semantics alone would likely underperform and should be enriched.
  • A direct extension of their logic is cross-city transfer: if route-choice semantics are universal, an embedding pretrained on one city should transfer to another without retraining; the paper currently reports single-city results and flags this as future work.
  • The reported sensitivity to POI completeness suggests a testable limit: near-random performance under heavily shuffled or sparse POI data would confirm that environmental semantics, not architecture, drive the gains.
  • The mixture-of-experts gating visualizations imply that experts specialize by functional region and road type; quantifying this specialization could yield interpretable route-choice summaries for planning applications.
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 / 3 minor

Summary. The paper proposes CORE, a trajectory representation learning framework that treats a GPS/map-matched trajectory as a sequence of context-aware route choices rather than a passive spatiotemporal sequence. CORE first constructs a context-enriched road network by using an LLM to summarize POI distributions at fine (road-segment) and coarse (grid/functional-hotspot) granularities, then encodes route-choice behavior with an MoE-based Route Choice Encoder that uses selected and unselected adjacent segments, historical transition likelihoods, and directional deviation. A Transformer encoder and contrastive pretraining produce the final trajectory embedding. The authors evaluate on four real-world datasets across six downstream tasks (road label prediction, travel time estimation, similar trajectory retrieval, destination prediction, path ranking, and trajectory generation), reporting large gains on several tasks and claiming that CORE "consistently outperforms" 12 state-of-the-art baselines with an average improvement of 9.79%.

Significance. If the reported gains are reliable, the core idea is a useful and timely contribution: injecting behavioral route-choice semantics and LLM-distilled POI context into trajectory embeddings is a natural inductive bias that appears to help most downstream tasks, especially road-label prediction and similar-trajectory retrieval. The paper is also commendable for releasing code, for including detailed ablations and sensitivity analyses, and for transparently reporting the API cost and wall-clock time of the LLM preprocessing stage. However, the central "consistently outperforms" claim is not supported by the paper's own numbers, and the absence of error bars or significance tests makes it difficult to judge the robustness of the many positive differences, several of which are small. The contribution is substantive enough to warrant a major revision rather than rejection, but the overclaim and missing statistical grounding must be fixed.

major comments (4)
  1. [Table VII (Path Ranking, Chengdu row)] CORE's MAE is 0.0934 while TRACK achieves 0.0927, and the reported Improvement is -0.76%. This directly contradicts the abstract, introduction, and conclusion claim that CORE "consistently outperforms" all 12 state-of-the-art baselines. Even if every other table entry favors CORE, one counterexample falsifies the strict claim. The claim should be weakened to "outperforms the baselines on most datasets and tasks" or otherwise qualified, with a concrete explanation for the Chengdu path-ranking degradation. This is not a wording issue; the consistency claim is the paper's headline conclusion.
  2. [Section V-A3 / Section V-B] The paper states that all results are averaged over five independent trials, but no standard deviations, confidence intervals, or significance tests are reported anywhere. Several of the favorable differences are small (e.g., Table VI Chengdu Acc@1: 0.4296 vs. 0.4150 for JGRM; Table IV Porto TTE MAE: 1.3519 vs. 1.4001 for TRACK; Table XI Chengdu MAE: 1.3681 vs. 1.3675 for DeepSeek-V3.2-Exp). Without variability information, the reader cannot determine whether these differences are reliable or within trial-to-trial noise. Please add error bars or standard deviations for at least the headline tables and report paired significance tests for the main claim on each dataset/task.
  3. [Abstract vs. Introduction] The abstract reports "15 state-of-the-art TRL methods" and a "9.20%" average improvement, while the title page, introduction, and conclusion report "12 state-of-the-art TRL methods" and a "9.79%" average improvement. This numerical inconsistency in the central claim must be corrected, and the aggregation rule (how the average improvement is computed across tasks, metrics, and datasets) should be stated explicitly.
  4. [Section IV-B, Eq. (13)] The historical transition likelihood P(r_c|r_i) is computed from counts N(r_i->r_c), but the paper does not state whether these counts are restricted to the training split or are computed over the entire trajectory corpus before the chronological 7:1:2 split. If the latter, transition probabilities would include test-set route frequencies, leaking future information into CORE's inputs and giving it an unfair advantage over baselines that do not use this signal. Please clarify that the counts are computed from the training split only (or from an independent historical period), and confirm that the ablation "w/o historical transition" remains valid under that setting.
minor comments (3)
  1. [Section V.B.5 (Path Ranking)] The task description is vague: path labels are said to be the IoU of each candidate with the "optimal" path, but the ground-truth path is the observed trajectory. It would help to define how the candidate set is generated and how the ground-truth path is chosen, so that the reader can interpret what the ranking score is measuring.
  2. [Table IX / Figure 6] The efficiency and data-efficiency experiments compare CORE with only a subset of baselines (JGRM, TRACK, GREEN, HOSER, STEGA). Please justify the selection of these baselines for these analyses, since the full comparison set is larger.
  3. [Figure 3] The labels in Figure 3 (visit count distributions) contain garbled glyph sequences in the version I reviewed, making the figure difficult to interpret. Please ensure the final PDF renders the axis labels and legend correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; independent empirical evaluation with minor reporting inconsistencies.

full rationale

CORE's derivation chain is self-contained: the Environment Perception Module builds road-segment representations from POI distributions via external LLMs (Qwen3), the Route Choice Encoder uses these plus navigational factors (Eqs. 10-20), and pretraining uses contrastive NT-Xent on augmented trajectories (Eq. 23). No downstream target (lanes, destination, travel time, retrieval detours, path-ranking targets, generated trajectories) enters these equations as a fitted input; the six evaluation tasks are held out and external to the pretraining objective. The historical transition likelihood P(r_c|r_i) in Eq. 13 is computed from the same trajectory corpus, but it is an input feature, not a predicted label, so it does not constitute fitting a parameter and then predicting it. The only self-citations are two baselines (STEGA [63], HOSER [64]) used for comparison; the central outperformance claim is an empirical result, not a consequence of those citations. Reporting issues exist: the abstract states '15 state-of-the-art TRL methods' and '9.20%' while the main text says '12' and '9.79%'; Section V-A3 says 'All reported results are averaged over five independent trials' but no error bars or significance tests are shown; and Table VII lists CORE's Chengdu path-ranking MAE as 0.0934 vs TRACK's 0.0927 with Improvement(%) -0.76, which contradicts the word 'consistently outperforms.' These are correctness/evidence concerns, not circularity, so the circularity score is 0.

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

The central result is a configuration of known components (LLM text embeddings, GAT, CNN pooling, MoE, Transformer, NT-Xent) plus a route-choice inductive bias. There are no new physical entities. All hand-set spatial scales and top-N rules are listed above; none is derived from first principles. The main uncharged dependencies are POI availability/quality and LLM fidelity.

free parameters (7)
  • Fine-grained POI perception radius δ = 100 m
    POIs within radius δ around each road segment are summarized by the LLM; δ is a hand-set scale, though sensitivity-tested (Fig. 10).
  • Coarse-grained grid side length L = 1000 m
    Grid size for functional hotspots; hand-set, sensitivity-tested (Fig. 10).
  • Critical segment selection ratio η = 20% (top 20% by traffic volume)
    Only the top 20% of segments receive LLM fine-grained semantics; η is empirically chosen and affects performance (Fig. 11).
  • Functional hotspot threshold = top 10% grids per POI category
    Grids with the highest POI counts per category are summarized by the LLM; hand-set threshold without dedicated sensitivity analysis.
  • MoE configuration = n_e=8 experts, k=2 activated
    Architecture hyperparameters; no sensitivity analysis is reported for these choices.
  • Contrastive augmentation parameters = crop ratio ~ U(0.05,0.15); 15% of segments perturbed with r ~ U(0.15,0.30)
    Augmentation hyperparameters chosen without sensitivity analysis.
  • Embedding dimension and Transformer depth = d=128, depth=6
    Fixed implementation choices shared with baselines.
assumptions (5)
  • domain assumption POI distributions are observable proxies for urban environmental semantics
    Section IV-A: "Because POIs provide a rich characterization of urban functions, we use them as observable proxies for environmental semantics." If POIs are incomplete or misaligned with actual route-choice drivers, the context-enriched network degrades; the paper's POI-completeness experiments confirm this dependence.
  • domain assumption LLM (Qwen3-8B) distills faithful environmental semantics from POI lists
    The fine- and coarse-grained modules treat LLM-generated descriptions as ground-truth semantic summaries (Appendix A). LLM hallucination or bias would inject wrong semantics into road segment representations.
  • domain assumption Historical transition counts P(r_c|r_i) from the training corpus are stationary and representative
    Eq. 13 computes transition likelihoods from historical trajectory counts; if route-choice behavior shifts over time or across cities, these features become stale. Used as a navigational factor in every route-choice step.
  • domain assumption Map matching (FMM) and OpenStreetMap networks are accurate enough
    All trajectories are map-matched with FMM to OSMnx-extracted networks (Section V-A1); matching errors propagate into segment sequences, POI assignments, and the selected/unselected contrast.
  • domain assumption Contrastive pretraining with cropping/temporal perturbation transfers to all six downstream tasks
    Section IV-C adopts the standard NT-Xent objective; the paper relies on the assumption that this objective shapes an embedding space useful for RLP, DP, TTE, STR, PR, and TG.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning." pith.science (2026). https://pith.science/paper/CVDBQK7Z

@misc{pith2026251014819,
  author       = {Pith},
  title        = {Pith review of: Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVDBQK7Z}},
  note         = {Machine review of arXiv:2510.14819}
}
read the original abstract

Trajectory representation learning (TRL) aims to encode raw trajectory data into low-dimensional embeddings for downstream tasks such as travel time estimation, mobility prediction, and trajectory similarity analysis. From a behavioral perspective, a trajectory reflects a sequence of route choices within an urban environment. However, most existing TRL methods ignore this underlying decision-making process and instead treat trajectories as static, passive spatiotemporal sequences, thereby limiting the semantic richness of the learned representations. To bridge this gap, we propose CORE, a TRL framework that integrates context-aware route choice semantics into trajectory embeddings. CORE first incorporates a multi-granular Environment Perception Module, which leverages large language models (LLMs) to distill environmental semantics from point of interest (POI) distributions, thereby constructing a context-enriched road network. Building upon this backbone, CORE employs a Route Choice Encoder with a mixture-of-experts (MoE) architecture, which captures route choice patterns by jointly leveraging the context-enriched road network and navigational factors. Finally, a Transformer encoder aggregates the route-choice-aware representations into a global trajectory embedding. Extensive experiments on 4 real-world datasets across 6 downstream tasks demonstrate that CORE consistently outperforms 15 state-of-the-art TRL methods, achieving an average improvement of 9.20\% over the best-performing baseline. Our code is available at https://github.com/caoji2001/CORE.

Figures

Figures reproduced from arXiv: 2510.14819 by the authors.

Figure 1
Figure 1. Even with the same OD pair, different drivers may choose substantially [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of CORE. The Environment Perception Module first constructs a context-aware road network based on the environmental semantics implied [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visit count distributions over road segments in Beijing and Chengdu. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (22 more)
Figure 4
Figure 4. Figure 4: Spatial distributions of the four trajectory datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of CORE and its ablated variants on the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Results of using varying proportions of training data on the Beijing [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 9
Figure 9. Figure 9: t-SNE visualization of trajectory embeddings for trajectories sharing [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 8
Figure 8. Figure 8: Visualization of MoE gating distributions. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: Sensitivity of CORE to spatial perception scale hyperparameters ( [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Sensitivity of CORE to POI completeness on the Beijing dataset [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Performance comparison of CORE and its ablated variants on the Chengdu dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Performance comparison of CORE and its ablated variants on the Xi’an dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Performance comparison of CORE and its ablated variants on the Porto dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 16
Figure 16. Figure 16: Results of using varying proportions of training data on the Chengdu dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]
Figure 17
Figure 17. Figure 17: Results of using varying proportions of training data on the Xi’an dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]
Figure 18
Figure 18. Figure 18: Results of using varying proportions of training data on the Porto dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_18.png]
Figure 19
Figure 19. Figure 19: Sensitivity of CORE to spatial perception scale hyperparameters ( [PITH_FULL_IMAGE:figures/full_fig_p019_19.png]
Figure 20
Figure 20. Figure 20: Sensitivity of CORE to spatial perception scale hyperparameters ( [PITH_FULL_IMAGE:figures/full_fig_p019_20.png]
Figure 21
Figure 21. Figure 21: Sensitivity of CORE to spatial perception scale hyperparameters ( [PITH_FULL_IMAGE:figures/full_fig_p019_21.png]
Figure 22
Figure 22. Figure 22: Impact of the critical segment selection ratio [PITH_FULL_IMAGE:figures/full_fig_p019_22.png]
Figure 23
Figure 23. Figure 23: Impact of the critical segment selection ratio [PITH_FULL_IMAGE:figures/full_fig_p020_23.png]
Figure 24
Figure 24. Figure 24: Impact of the critical segment selection ratio [PITH_FULL_IMAGE:figures/full_fig_p020_24.png]
Figure 25
Figure 25. Figure 25: Sensitivity of CORE to POI completeness on the Chengdu dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_25.png]
Figure 26
Figure 26. Figure 26: Sensitivity of CORE to POI completeness on the Xi’an dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_26.png]
Figure 27
Figure 27. Figure 27: Sensitivity of CORE to POI completeness on the Porto dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_27.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

84 extracted references · 6 linked inside Pith

  1. [1]

    Trajectory data mining: an overview,

    Y . Zheng, “Trajectory data mining: an overview,”ACM Trans. Intell. Syst. Technol., 2015

    2015

  2. [2]

    A survey on trajectory data management, analytics, and learning,

    S. Wang, Z. Bao, J. S. Culpepper, and G. Cong, “A survey on trajectory data management, analytics, and learning,”ACM Comput. Surv., 2021. 13

    2021

  3. [3]

    Deep learning for spatio-temporal data mining: A survey,

    S. Wang, J. Cao, and P. S. Yu, “Deep learning for spatio-temporal data mining: A survey,”IEEE Trans. Knowl. Data Eng., 2022

    2022

  4. [4]

    Citywide estimation of travel time distributions with bayesian deep graph learning,

    J. J. Yu, “Citywide estimation of travel time distributions with bayesian deep graph learning,”IEEE Trans. Knowl. Data Eng., 2023

    2023

  5. [5]

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

    C. Zhang, K. Zhao, and M. Chen, “Beyond the limits of predictability in human mobility prediction: Context-transition predictability,”IEEE Trans. Knowl. Data Eng., 2023

    2023

  6. [6]

    Spatio-temporal trajectory similarity measures: A comprehensive survey and quantitative study,

    D. Hu, L. Chen, H. Fang, Z. Fang, T. Li, and Y . Gao, “Spatio-temporal trajectory similarity measures: A comprehensive survey and quantitative study,”IEEE Trans. Knowl. Data Eng., 2024

    2024

  7. [7]

    Jointly contrastive representation learning on road network and trajectory,

    Z. Mao, Z. Li, D. Li, L. Bai, and R. Zhao, “Jointly contrastive representation learning on road network and trajectory,” inCIKM, 2022

    2022

  8. [8]

    Lightpath: Lightweight and scalable path representation learning,

    S. B. Yang, J. Hu, C. Guo, B. Yang, and C. S. Jensen, “Lightpath: Lightweight and scalable path representation learning,” inSIGKDD, 2023

    2023

Show all 84 references
  1. [9]

    Self-supervised trajectory representation learning with temporal regularities and travel semantics,

    J. Jiang, D. Pan, H. Ren, X. Jiang, C. Li, and J. Wang, “Self-supervised trajectory representation learning with temporal regularities and travel semantics,” inICDE, 2023

    2023

  2. [10]

    Trembr: Exploring road networks for trajectory representation learning,

    T.-Y . Fu and W.-C. Lee, “Trembr: Exploring road networks for trajectory representation learning,”ACM Trans. Intell. Syst. Technol., 2020

    2020

  3. [11]

    Unsupervised path representation learning with curriculum negative sampling,

    S. B. Yang, C. Guo, J. Hu, J. Tang, and B. Yang, “Unsupervised path representation learning with curriculum negative sampling,” inIJCAI, 2021

    2021

  4. [12]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Comput., 1997

    1997

  5. [13]

    Learning phrase representations using rnn encoder- decoder for statistical machine translation,

    K. Choet al., “Learning phrase representations using rnn encoder- decoder for statistical machine translation,” inEMNLP, 2014

    2014

  6. [14]

    Attention is all you need,

    A. Vaswaniet al., “Attention is all you need,” inNeurIPS, 2017

    2017

  7. [15]

    Robust road network representation learning: When traffic patterns meet traveling semantics,

    Y . Chenet al., “Robust road network representation learning: When traffic patterns meet traveling semantics,” inCIKM, 2021

    2021

  8. [16]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,” inICLR, 2018

    2018

  9. [17]

    Grid and road expressions are complementary for trajectory representation learning,

    S. Zhou, S. Shang, L. Chen, P. Han, and C. S. Jensen, “Grid and road expressions are complementary for trajectory representation learning,” inSIGKDD, 2025

    2025

  10. [18]

    Route choice modeling: past, present and future research directions,

    C. G. Prato, “Route choice modeling: past, present and future research directions,”Journal of Choice Modelling, 2009

    2009

  11. [19]

    Understanding individual routing behaviour,

    A. Lima, R. Stanojevic, D. Papagiannaki, P. Rodriguez, and M. C. Gonz´alez, “Understanding individual routing behaviour,”Journal of The Royal Society Interface, 2016

    2016

  12. [20]

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

    D. Daiet al., “Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models,” inACL, 2024

    2024

  13. [21]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inICML, 2020

    2020

  14. [22]

    Deep representation learning for trajectory similarity computation,

    X. Li, K. Zhao, G. Cong, C. S. Jensen, and W. Wei, “Deep representation learning for trajectory similarity computation,” inICDE, 2018

    2018

  15. [23]

    Computing trajectory similarity in linear time: A generic seed-guided neural metric learning approach,

    D. Yao, G. Cong, C. Zhang, and J. Bi, “Computing trajectory similarity in linear time: A generic seed-guided neural metric learning approach,” inICDE, 2019

    2019

  16. [24]

    A graph-based approach for trajectory similarity computation in spatial networks,

    P. Han, J. Wang, D. Yao, S. Shang, and X. Zhang, “A graph-based approach for trajectory similarity computation in spatial networks,” in SIGKDD, 2021

    2021

  17. [25]

    When will you arrive? estimating travel time based on deep neural networks,

    D. Wang, J. Zhang, W. Cao, J. Li, and Y . Zheng, “When will you arrive? estimating travel time based on deep neural networks,” inAAAI, 2018

    2018

  18. [26]

    Learning to estimate the travel time,

    Z. Wang, K. Fu, and J. Ye, “Learning to estimate the travel time,” in SIGKDD, 2018

    2018

  19. [27]

    Constgat: Contextual spatial-temporal graph attention network for travel time estimation at baidu maps,

    X. Fang, J. Huang, F. Wang, L. Zeng, H. Liang, and H. Wang, “Constgat: Contextual spatial-temporal graph attention network for travel time estimation at baidu maps,” inSIGKDD, 2020

    2020

  20. [28]

    Red: Effective trajectory representation learning with comprehensive information,

    S. Zhou, S. Shang, L. Chen, C. S. Jensen, Jensen, and P. Kalnis, “Red: Effective trajectory representation learning with comprehensive information,” inVLDB, 2025

    2025

  21. [29]

    More than routing: Joint gps and route modeling for refine trajectory representation learning,

    Z. Maet al., “More than routing: Joint gps and route modeling for refine trajectory representation learning,” inWWW, 2024

    2024

  22. [30]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inNAACL, 2019

    2019

  23. [31]

    Bridging traffic state and trajectory for dynamic road network and trajectory representation learning,

    C. Hanet al., “Bridging traffic state and trajectory for dynamic road network and trajectory representation learning,” inAAAI, 2025

    2025

  24. [32]

    Learning embeddings of intersections on road networks,

    M.-x. Wang, W.-C. Lee, T.-y. Fu, and G. Yu, “Learning embeddings of intersections on road networks,” inSIGSPATIAL, 2019

    2019

  25. [33]

    On representation learning for road networks,

    M.-X. Wang, W.-C. Lee, T.-Y . Fu, and G. Yu, “On representation learning for road networks,”ACM Trans. Intell. Syst. Technol., 2020

    2020

  26. [34]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” inSIGKDD, 2016

    2016

  27. [35]

    Semantic-enhanced representation learning for road networks with temporal dynamics,

    Y . Chen, X. Li, G. Cong, Z. Bao, and C. Long, “Semantic-enhanced representation learning for road networks with temporal dynamics,” arXiv preprint arXiv:2403.11495, 2024

    2024 arXiv

  28. [36]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,”IEEE Trans. Neural Netw. Learn. Syst., 2020

    2020

  29. [37]

    Relational fusion networks: Graph convolutional networks for road networks,

    T. S. Jepsen, C. S. Jensen, and T. D. Nielsen, “Relational fusion networks: Graph convolutional networks for road networks,”IEEE Trans. Intell. Transp. Syst., 2020

    2020

  30. [38]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inICLR, 2017

    2017

  31. [39]

    Learning effective road network representation with hierarchical graph neural networks,

    N. Wu, X. W. Zhao, J. Wang, and D. Pan, “Learning effective road network representation with hierarchical graph neural networks,” in SIGKDD, 2020

    2020

  32. [40]

    Road network representation learning: A dual graph-based approach,

    L. Zhang and C. Long, “Road network representation learning: A dual graph-based approach,”ACM Trans. Knowl. Discov. Data, 2023

    2023

  33. [41]

    A survey on hypergraph representation learning,

    A. Antelmi, G. Cordasco, M. Polato, V . Scarano, C. Spagnuolo, and D. Yang, “A survey on hypergraph representation learning,”ACM Comput. Surv., 2023

    2023

  34. [42]

    Spatial structure-aware road network embedding via graph contrastive learning,

    Y . Chang, E. Tanin, X. Cao, and J. Qi, “Spatial structure-aware road network embedding via graph contrastive learning,” inEDBT, 2023

    2023

  35. [43]

    Graph contrastive learning with adaptive augmentation,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” inWWW, 2021

    2021

  36. [44]

    Training language models to follow instructions with human feedback,

    L. Ouyanget al., “Training language models to follow instructions with human feedback,” inNeurIPS, 2022

    2022

  37. [45]

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

    J. Wanget al., “Large language models as urban residents: An llm agent framework for personal mobility generation,” inNeurIPS, 2024

    2024

  38. [46]

    Chain-of-planned-behaviour workflow elicits few-shot mobility generation in llms,

    C. Shaoet al., “Chain-of-planned-behaviour workflow elicits few-shot mobility generation in llms,”arXiv preprint arXiv:2402.09836, 2024

    2024 arXiv

  39. [47]

    Agentmove: Predicting human mo- bility anywhere using large language model based agentic framework,

    J. Feng, Y . Du, J. Zhao, and Y . Li, “Agentmove: Predicting human mo- bility anywhere using large language model based agentic framework,” arXiv preprint arXiv:2408.13986, 2024

    2024 arXiv

  40. [48]

    Mobility-llm: Learning visiting intentions and travel preference from human mobility data with large language models,

    L. Gonget al., “Mobility-llm: Learning visiting intentions and travel preference from human mobility data with large language models,” in NeurIPS, 2024

    2024

  41. [49]

    Large language models for next point-of-interest recommendation,

    P. Li, M. de Rijke, H. Xue, S. Ao, Y . Song, and F. D. Salim, “Large language models for next point-of-interest recommendation,” inSIGIR, 2024

    2024

  42. [50]

    Poi-enhancer: An llm-based semantic enhancement framework for poi representation learning,

    J. Chenget al., “Poi-enhancer: An llm-based semantic enhancement framework for poi representation learning,” inAAAI, 2025

    2025

  43. [51]

    Promptst: Prompt-enhanced spatio-temporal multi- attribute prediction,

    Z. Zhanget al., “Promptst: Prompt-enhanced spatio-temporal multi- attribute prediction,” inCIKM, 2023

    2023

  44. [52]

    Urbangpt: Spatio-temporal large language models,

    Z. Liet al., “Urbangpt: Spatio-temporal large language models,” in SIGKDD, 2024

    2024

  45. [53]

    Unist: A prompt-empowered universal model for urban spatio-temporal prediction,

    Y . Yuan, J. Ding, J. Feng, D. Jin, and Y . Li, “Unist: A prompt-empowered universal model for urban spatio-temporal prediction,” inSIGKDD, 2024

    2024

  46. [54]

    Trajcogn: Leveraging llms for cognizing movement patterns and travel purposes from trajectories,

    Z. Zhouet al., “Trajcogn: Leveraging llms for cognizing movement patterns and travel purposes from trajectories,” inIJCAI, 2025

    2025

  47. [55]

    Hillier and J

    B. Hillier and J. Hanson,The social logic of space. Cambridge University Press, 1989

    1989

  48. [56]

    Qwen3 technical report,

    A. Yanget al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  49. [57]

    Qwen3 embedding: Advancing text embedding and reranking through foundation models,

    Y . Zhanget al., “Qwen3 embedding: Advancing text embedding and reranking through foundation models,”arXiv preprint arXiv:2506.05176, 2025

    2025 arXiv

  50. [58]

    The nature of cities,

    C. D. Harris and E. L. Ullman, “The nature of cities,”The annals of the American academy of political and social science, 1945

    1945

  51. [59]

    Discovering regions of different functions in a city using human mobility and pois,

    J. Yuan, Y . Zheng, and X. Xie, “Discovering regions of different functions in a city using human mobility and pois,” inSIGKDD, 2012

    2012

  52. [60]

    Modeling and analyzing urban networks and amenities with osmnx,

    G. Boeing, “Modeling and analyzing urban networks and amenities with osmnx,”Geographical Analysis, 2025

    2025

  53. [61]

    Fast map matching, an algorithm integrating hidden markov model with precomputation,

    C. Yang and G. Gidofalvi, “Fast map matching, an algorithm integrating hidden markov model with precomputation,”Int. J. Geogr. Inf. Sci., 2018

    2018

  54. [62]

    Continuous trajectory generation based on two-stage gan,

    W. Jiang, W. X. Zhao, J. Wang, and J. Jiang, “Continuous trajectory generation based on two-stage gan,” inAAAI, 2023

    2023

  55. [63]

    Spa- tiotemporal gated traffic trajectory simulation with semantic-aware graph learning,

    Y . Wang, J. Cao, W. Huang, Z. Liu, T. Zheng, and M. Song, “Spa- tiotemporal gated traffic trajectory simulation with semantic-aware graph learning,”Inf. Fusion, 2024

    2024

  56. [64]

    Holistic semantic representation for navigational trajectory generation,

    J. Caoet al., “Holistic semantic representation for navigational trajectory generation,” inAAAI, 2025

    2025

  57. [65]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inICLR, 2019

    2019

  58. [66]

    Auxiliary-loss- free load balancing strategy for mixture-of-experts,

    L. Wang, H. Gao, C. Zhao, X. Sun, and D. Dai, “Auxiliary-loss- free load balancing strategy for mixture-of-experts,”arXiv preprint arXiv:2408.15664, 2024. 14

    2024 arXiv

  59. [67]

    Context-aware path ranking in road networks,

    S. B. Yang, C. Guo, and B. Yang, “Context-aware path ranking in road networks,”IEEE Trans. Knowl. Data Eng., 2022

    2022

  60. [68]

    A new measure of rank correlation,

    M. G. Kendall, “A new measure of rank correlation,”Biometrika, 1938

    1938

  61. [69]

    The proof and measurement of association between two things,

    C. Spearman, “The proof and measurement of association between two things,”Am. J. Psychol, 1904

    1904

  62. [70]

    Distributed trajectory similarity search,

    D. Xie, F. Li, and J. M. Phillips, “Distributed trajectory similarity search,” inVLDB, 2017

    2017

  63. [71]

    Exact indexing of dynamic time warping,

    E. Keogh, “Exact indexing of dynamic time warping,” inVLDB, 2002

    2002

  64. [72]

    Robust and fast similarity search for moving object trajectories,

    L. Chen, M. T. ¨Ozsu, and V . Oria, “Robust and fast similarity search for moving object trajectories,” inSIGMOD, 2005

    2005

  65. [73]

    Visualizing data using t-sne,

    L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,”J. Mach. Learn. Res., 2008. 15 SUPPLEMENTARYMATERIAL CAPTURINGCONTEXT-AWAREROUTECHOICESEMANTICS FORTRAJECTORYREPRESENTATIONLEARNING APPENDIXA DETAILS OFENVIRONMENTPERCEPTIONMODULE A. Prompts for Fine-Grained Semanti...

    2008

  66. [74]

    Vibrant Commerce and Diverse Consumption Scenes •Dense Shopping Services: Six shopping-related POIs form a commercial belt focused on clothing, shoes, specialty retail, and convenience stores. •Blend of Fashion and Tradition: Stores such as ‘‘Fashion Frontline (Xinjiekou Stree...

  67. [75]

    •Public Security: The ‘‘Dashilan Police Patrol Station’’ reflects strict public safety management, enhancing residents’ and visitors’ sense of security

    Transportation Hub and High Foot Traffic •Direct Metro Access: ‘‘Qianmen Subway Station (Southwest Exit C)’’ links subway and ground transport, generating heavy flows during rush hours and serving both commercial and commuting needs. •Public Security: The ‘‘Dashilan Police Pat...

  68. [76]

    Prominent Tourism and Accommodation Functions •Comprehensive Lodging Options: Hotels such as ‘‘Qianmen Guanqi Hotel (Qianmen East Branch)’’ and ‘‘Shunchaolong Hotel’’ confirm the area as a tourist hotspot, likely near historical and cultural attractions like Qianmen Street. •C...

  69. [77]

    vibrant commerce and diverse consumption scenes

    Healthcare and Community Services •Blood Donation Cabin: The blood donation cabin provides basic medical support and public welfare services for residents and tourists, strengthening community functions. Summary: This road section lies in thecore Qianmen business district, int...

  70. [78]

    •Supporting Daily Services: [convenience stores] (52) and [supermarkets] (28) enable one-stop shopping for everyday needs

    Highly Concentrated, Diverse Business Mix •Core Business Types: Mainly [specialty stores] (790) and [clothing & shoes stores] (708), indicating strong fashion and personalized shopping, likely with brand flagships and boutiques. •Supporting Daily Services: [convenience stores]...

  71. [79]

    Likely Location: Beijing’s Core Commercial Districts •POI Name Clues: Frequent ‘‘Guomao’’ in POI names suggests proximity to the Guomao CBD or a similar prime business district. •Exceptionally High POI Density: Nearly 2,000 shopping-related POIs within one square kilometer, fa...

  72. [80]

    Multi-Level Consumption Structure •Mass Market Consumption: Supermarkets and convenience stores address the daily needs of nearby residents. •High-End and Niche Consumption: Stores like ‘‘Dionysus Wine Cellar’’, ‘‘Montblanc’’, and ‘‘Beijing Jianya Century Auction Co., Ltd.’’ r...

  73. [81]

    ‘‘Guomao’’ further hints at a metro hub or major retail cluster

    Commercial Vitality Driven by Transit and Foot Traffic •Density Backed by Accessibility: Such POI concentration implies strong transport links (e.g., metro, bus hubs) and high population density (e.g., office towers, housing). ‘‘Guomao’’ further hints at a metro hub or major r...

  74. [82]

    •Traffic Pressure: Intensive commercial activity can strain surrounding roads, requiring effective traffic planning and management

    Potential Challenges •Oversaturation Risk: High POI density may cause homogeneous competition and pressure on smaller retailers. •Traffic Pressure: Intensive commercial activity can strain surrounding roads, requiring effective traffic planning and management. Conclusion:The a...

  75. [83]

    Within the examined ranges, performance remains stable, suggesting that CORE does not critically depend on a specific choice of spatial scale

    Hyperparameter Sensitivity:Figures 19 to 21 show the sensitivity of CORE to the spatial perception hyperparameters (i.e., the POI perception radiusδand the grid side lengthL) on the Chengdu, Xi’an, and Porto datasets. Within the examined ranges, performance remains stable, sug...

  76. [84]

    Sensitivity to POI Completeness:Figures 25 to 27 report the impact of POI completeness on CORE across the Chengdu, Xi’an, and Porto datasets. The observed trends align perfectly with those in the main text, confirming that CORE consistently maintains robustness against sparse ...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.