Pith. sign in

REVIEW 3 major objections 5 minor 50 references

POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning

T0 review · 3 major / 5 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 POI-Enhancer claims that LLM-extracted text about a POI's visit pattern, address, and surroundings, fused into any classic POI embedding through cross-attention and contrastive learning, improves recommendation, sequence classification…

desk verdict Useful, portable LLM enhancer for POI embeddings with a plausible leakage problem in the attribute pipeline and one contradictory ablation row; worth refereeing, but the numbers need a closer look. read the letter →

arxiv 2502.10038 v2 pith:P7X6JFZZ submitted 2025-02-14 cs.AI

Jiawei Cheng , Jingyuan Wang , Yichuan Zhang , Jiahao Ji , Yuanshao Zhu , Zhibo Zhang , Xiangyu Zhao This is my paper
classification cs.AI
keywords POIrepresentationlearninglargelanguagemodelssemanticfeaturefusioncross-attentioncontrastiverecommendationcheck-insequenceclassificationvisitorflowprediction
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

POI-Enhancer is a portable framework that takes the vector representation of a point of interest produced by any existing POI embedding model and enriches it with textual knowledge extracted from a frozen large language model. The framework generates three specialized prompts—visit pattern, address, and surrounding environment—and uses the LLM's last hidden states as semantic features. Those features are aligned, fused, and injected into the original embedding through attention-based modules, then refined by contrastive learning that samples positives by time, geography, and function. On Foursquare check-in data from New York, Singapore, and Tokyo, the paper reports that every one of six classic POI embedding methods improves across POI recommendation, check-in sequence classification, and visitor flow prediction. The paper's central claim is that this semantic enrichment is the missing modality that lets mobility embeddings capture human-understandable meaning beyond coordinates and co-occurrence.

What carries the argument

The load-bearing object is the semantic feature vector $E_{LLM}$, the attention-weighted fusion of two aligned LLM outputs, and the cross-attention module that injects it into the original POI embedding. The paper generates three prompts per POI—a visit-pattern prompt, an address prompt, and a surrounding prompt—and takes the frozen LLM's final hidden state for each. Dual Feature Alignment runs a multi-head Transformer encoder between the address feature and each of the other two features, so address semantics refine visit and surrounding semantics; Semantic Feature Fusion then merges the two refined vectors with learned attention scores. Cross Attention Fusion, built from multi-query attention and parallel attention and feed-forward layers, treats the semantic vector as queries and the original embedding as keys and values, producing the enhanced vector $E_{FUSE}$. Multi-View Contrastive Learning selects positive samples by same-date sequence adjacency, same-category proximity within a 0.5 km square, and shared category plus visit pattern, and trains with InfoNCE plus a cosine-similarity preservation loss.

What would settle it

Recompute the visit-pattern and surrounding-category attributes using only training-split check-ins and POI records, re-run the full POI-Enhancer pipeline on the same six baselines and three datasets, and compare Hit@1, Accuracy, and MAE with the paper's tables; if the gains shrink to near the no-text baseline, the reported improvement depends on leakage rather than on genuine semantic enrichment.

Watch

Extended reading notes

Core claim

The paper's central claim is that a point of interest's ordinary embedding—trained from check-in sequences and geography—leaves out exactly the kind of semantic knowledge about the place that a large language model already has, and that adding this knowledge on top is enough to lift every tested embedding method on every downstream task. Concretely, POI-Enhancer extracts three views of each POI (when people visit, where it is, what surrounds it) as LLM hidden states, aligns the address view with the other two via a Transformer encoder, fuses them with attention-weighted summation, and then uses a cross-attention Transformer to adaptively combine the semantic vector with the original POI embedding. A multi-view contrastive loss then pulls together POIs that share visit times, nearby same-category locations, and functional patterns. The paper reports consistent improvements on Hit@1 and Hit@5 for recommendation, Accuracy and Macro-F1 for sequence classification, and MAE and RMSE for visitor-flow prediction, with some classification metrics more than doubling. The strongest gains appear for models like POI2Vec that previously ignored semantics; the framework is model-agnostic because the base embedding network is left frozen.

Load-bearing premise

The weakest premise is that the visit-pattern and surrounding-category attributes fed into the LLM prompts are informative without leakage: the appendix computes them from the full collection of check-in records and POIs, and the paper never states that this preprocessing is restricted to the training split, so test-period information could enter the text features and inflate the reported improvements.

Editorial extensions

If this is right

  • Any existing POI embedding method can be upgraded without retraining the base model; the framework only trains the fusion and contrastive modules.
  • The biggest reported gains come in check-in sequence classification, suggesting that LLM text adds discriminative signals that trajectory co-occurrence alone cannot provide.
  • With the LLM frozen, the semantic features for all POIs can be computed once and reused, so the per-training-step cost comes from the fusion modules rather than from LLM inference.
  • The enhanced representations separate same-category, nearby POIs that differ in ambiance and function, as the case study's increased Euclidean distance between two French restaurants shows.
  • The framework is portable across cities: the same prompts and modules improve embeddings on New York, Singapore, and Tokyo without city-specific tuning.

Reading between the lines

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

  • The preprocessing that computes visit patterns and surrounding categories is described using the full check-in records and POI set; if it is not restricted to the training split, test-period records could shape the text features and make the reported gains partly a leakage effect. Restricting that preprocessing to training data is a direct falsification test.
  • The same three-prompt extraction and cross-attention fusion could be applied to other text encoders or to region-level embeddings, not just POI vectors, whenever a frozen language model can describe a spatial unit.
  • The framework's contrastive positives encode explicit human views (time, space, function); a testable extension is whether removing any one view changes downstream gains in a way that reveals which semantic dimension carries the improvement.
  • Because the LLM output is free-form hidden states rather than parsed text, replacing Llama-2-7B with a larger or smaller model would test how much of the gain comes from parametric world knowledge versus the prompt structure.
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

3 major / 5 minor

Summary. The paper proposes POI-Enhancer, a modular framework that enriches POI representations from classical embedding models (Skip-Gram, POI2Vec, Geo-Teaser, TALE, Hier, CTLE) with textual features extracted from a frozen LLM (Llama-2-7B). Three specialized prompts are designed to capture visit pattern, address, and surrounding-environment semantics; a Dual Feature Alignment and Semantic Feature Fusion module refine the LLM hidden states, a Cross Attention Fusion module injects them into the original embeddings, and Multi-View Contrastive Learning (sequence-time, geography, functional) further shapes the representation space. Experiments on three Foursquare datasets claim significant improvements across POI recommendation, check-in sequence classification, and visitor flow prediction. An ablation study, parameter analysis, clustering quality analysis, and case study are also provided.

Significance. If the reported gains are robust, the paper makes a useful contribution: it is a portable, model-agnostic enhancement layer that can be stacked on existing POI embedding models without modifying their training procedures, and it introduces a systematic way to turn LLM knowledge into vector-form features for location-based tasks. The experiments cover three datasets, six classical baselines plus three LLM baselines, and three downstream tasks, which is a reasonably broad evaluation. The authors also release code, which supports reproducibility. The main scientific value lies in demonstrating that frozen LLM hidden states can serve as a general-purpose semantic side-information source for POI representations, rather than in any single model component.

major comments (3)
  1. [Technical Appendix: POI Attributes Preprocess] The visit pattern and surrounding attributes are computed by aggregating the entire check-in dataset ('We conduct a statistical analysis of the check-in data', 'We categorize each POI's check-in records into the corresponding time slot') with no explicit statement that this preprocessing is restricted to the training split. These attributes are inserted into the LLM prompts (Fig. 5), become LLM hidden states E^V and E^S, and flow through Dual Feature Alignment, Semantic Feature Fusion, and Cross Attention Fusion into the final E_FUSE that is evaluated on test sets. If any test-period check-in contributes to the modal visit pattern or surrounding categories, the framework leaks test information into the enhanced representations, and the improvements in Table 2 would not be attributable to semantic enrichment. The authors must either state and enforce a train-only computation of these attributes or re-run the experiments with such a restriction; this is load-bearing for the central claim.
  2. [Table 9] Table 9 reports that Hier+ (with contrastive learning and the full fusion process) and Hier+(w/o CL) (which replaces the fusion process with vector addition and removes contrastive learning) achieve identical Hit@1 = 8.009 and Hit@5 = 19.197. This is inconsistent with the text in Sections 'Further Analysis on POI-Enhancer' and the appendix, which claim that both the fusion process and contrastive learning contribute to performance. Identical results to four decimal places are not expected unless the modification has no effect on the downstream task, which contradicts the stated ablation conclusion. The authors need to explain this result, correct the experimental setup or reporting, and re-run the ablation. Also, the caption of Table 9 reads 'The attributes of two similar POIs', which is a copy-paste error; the caption should describe the ablation results.
  3. [Across all experiments (Tables 2, 6, 7, 8; Fig. 2)] The abstract and conclusion claim that the framework 'significantly improves' performance, but no error bars, standard deviations, or significance tests are reported. Table 2 shows a single run per condition, and several improvements are small in relative terms (e.g., Skip-Gram+ Hit@1 in TKY improves by 3.46%, and Skip-Gram+ MAE in NY improves by 9.43% on a metric where lower is better, leaving unclear whether the difference is within run-to-run noise). The word 'significant' therefore lacks statistical support. The authors should provide results over multiple random seeds (at least 3, preferably 5) and either report confidence intervals or perform appropriate significance tests (e.g., paired t-test or Wilcoxon) on the downstream metrics. This is necessary to support the central claim of consistent improvement.
minor comments (5)
  1. [Section 4, Experiment Setup] The split ratio is reported as '2:1:7 for the test set, validation set, and training set' and then the training set is stated to be used for contrastive sampling. It is not clear whether the validation set is used for early stopping or hyperparameter selection, and whether the attribute preprocessing (visit pattern, surrounding) is computed before or after this split. Please clarify the exact data flow, including which parts of the data are used for the LLM feature extraction and which for contrastive sampling.
  2. [Equation (13)] The InfoNCE loss in Eq. (13) uses a sum with index 'i=0' but the numerator involves a single positive sample and the denominator over the batch; the index notation is inconsistent with the preceding text and should be rewritten for clarity.
  3. [Technical Appendix, Baseline Descriptions] The appendix describes the baselines but does not state the source of the hyperparameters for each baseline (e.g., embedding dimension, number of negative samples, iterations). Since the comparison is central to the claims, a table of hyperparameters or a statement that default settings from the original papers were used would improve reproducibility.
  4. [Table 2, Improvement rows] For MAE and RMSE, lower is better, yet the 'Impr.' rows list positive percentages for all metrics. The sign convention (e.g., a 9.43% improvement in MAE corresponds to a 9.43% decrease) should be stated explicitly to avoid confusion.
  5. [Section 3, Prompt Generation] The prompt examples in Fig. 5 include the phrase 'Time Information: People usually visit the POI between 6 am and 9 am. And people usually come to the POI on weekdays.' This text is derived from the data-derived visit pattern, which highlights the point that the 'semantic knowledge' partly originates from the same check-in statistics rather than purely from external LLM knowledge. The paper should discuss this distinction and its implications for the claimed contribution.

Circularity Check

2 steps flagged · score 5.0 of 10

Two supporting analyses are partly circular: the NMI quality metric is aligned with the category-based contrastive supervision, and the visit-pattern statistic injected into the LLM prompt may encode test-period check-in information; the core Table 2 downstream comparisons remain independent.

  1. fitted input called prediction [Section 3 'Multi-View Contrastive Learning' (Functional Contrastive Learning) and Section 4 'Quality Analysis' (Fig. 4)]
    "only POIs that share the same category and visit pattern as the anchor sample are regarded as positive samples. ... The number of clusters was set to match the number of POI categories in each dataset. We then assessed the clustering performance using the Normalized Mutual Information (NMI) metric."

    The NMI clustering metric is evaluated against POI category labels, which are exactly the labels used to define positive and negative samples in Geography Contrastive Learning ('POIs of the same category in that area as positive samples') and Functional Contrastive Learning. The InfoNCE loss in Eq. (13) directly optimizes the enhanced representations to pull same-category POIs together and push different-category POIs apart. Reporting the post-enhancement NMI increase in Fig. 4 as evidence of 'successfully extracted high-quality textual features' therefore measures the training objective itself through a K-means wrapper; the improvement is substantially forced by construction rather than independent evidence of semantic quality.

  2. fitted input called prediction [Technical Appendix 'POI Attributes Preprocess' (Visit Pattern of POI); Section 3 Eq. (1); Section 4 'POI Visitor Flow Prediction']
    "We conduct a statistical analysis of the check-in data to identify the visit patterns of a POI... Afterward, we categorize each POI’s check-in records into the corresponding time slot. Finally, the time slot with the highest number of check-in records will represent the daily visit pattern."

    The visit-pattern statistic is inserted into the POI Visit Pattern Prompt (Fig. 5 states 'People usually visit the POI between 6 am and 9 am... on weekdays') and becomes part of the semantic feature through Eq. (1), E^V_pi = H(T^V_pi), which flows through Dual Feature Alignment, Semantic Feature Fusion, and Cross Attention Fusion into E_FUSE. The appendix never states that this aggregation is restricted to the training split; the only split statement in Section 4 concerns the dataset for downstream tasks and contrastive sampling. Since POI Visitor Flow Prediction builds its labels from check-in counts in the same dataset, a visit pattern computed over all check-ins would leak aggregate test-period information into the representation.

full rationale

The paper's primary empirical claim—that POI-Enhancer improves six POI embedding methods on three downstream tasks (Table 2)—is evaluated on held-out test splits and is not forced by construction; the downstream models are trained separately and the enhanced representations are not fitted to those test labels. I therefore do not regard the core Table 2 comparison as circular. However, two supporting elements are partially circular. First, the 'quality analysis' in Section 4 uses NMI against POI category labels as its ground truth, while the Geography and Functional Contrastive Learning objectives (and the InfoNCE loss, Eq. 13) explicitly use the same category labels to pull same-category POIs together and push different-category POIs apart. The improved NMI in Fig. 4 is thus largely the training objective read back through K-means, not independent evidence of semantic quality. Second, the 'visit pattern' attribute inserted into the LLM prompt is a statistical mode computed from 'the check-in data' with no stated train/test restriction; since the POI Visit Pattern Prompt becomes E^V via Eq. (1) and flows into E_FUSE, any use of test-period check-ins in that aggregation would let the enhanced representation encode aggregate test information for tasks such as visitor flow prediction, whose labels are derived from the same check-in records. The paper should state and enforce train-only preprocessing. I also note that Table 9 reports identical Hit@1/Hit@5 for Hier+ and Hier+(w/o CL), which contradicts the text's claim that contrastive learning contributes; this is a reporting inconsistency to verify, though it is not itself a circularity. No load-bearing self-citation chain or uniqueness argument appears in the paper.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical and does not depend on a small set of deriving constants. The main assumptions are about the informativeness of LLM text features and the validity of the hand-crafted attributes; these are domain assumptions rather than fitted parameters. The key hyperparameter choices (L1, L2, tau, lambda, gamma) are tuned or hand-set and could affect the magnitude of gains.

free parameters (6)
  • L1: number of Dual Feature Alignment layers = 4
    Selected via parameter analysis on the NY dataset with the Hier model; Fig. 3 shows an optimum at 4.
  • L2: number of Cross Attention Fusion layers = 2
    Selected via parameter analysis; performance degrades beyond 2 on the NY dataset.
  • tau: sampling distance for surrounding and Geography Contrastive Learning = 0.5 km
    Hand-set for both the surrounding attributes and the geography contrastive sampler; no sensitivity analysis is reported.
  • lambda: threshold for Sequence-Time contrastive neighbors = 2
    Hand-set; no sensitivity analysis is reported.
  • gamma: InfoNCE temperature = 0.1
    Hand-set; no sensitivity analysis is reported.
  • Prompt template components (role-playing, attribute headers, question) = hand-designed
    The ablation shows removing these components hurts performance, but the design space is not explored, so the exact template is a hand-chosen input.
assumptions (3)
  • domain assumption Llama-2-7B's frozen last hidden state contains useful POI-related semantic knowledge
    The entire method rests on the premise that LLM text features are informative for POI representations; no validation of this assumption independent of the downstream tasks is provided.
  • domain assumption Visit pattern and surrounding attributes computed from check-in data accurately represent POI semantics
    These hand-crafted attributes are inputs to the prompts; the paper does not prove they are representative and does not state that they are computed only on training data.
  • standard math Transformer attention, InfoNCE loss, and standard backpropagation assumptions hold
    Standard machine learning machinery is used without modification, so these are accepted background assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning." pith.science (2026). https://pith.science/paper/P7X6JFZZ

@misc{pith2026250210038,
  author       = {Pith},
  title        = {Pith review of: POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P7X6JFZZ}},
  note         = {Machine review of arXiv:2502.10038}
}
read the original abstract

POI representation learning plays a crucial role in handling tasks related to user mobility data. Recent studies have shown that enriching POI representations with multimodal information can significantly enhance their task performance. Previously, the textual information incorporated into POI representations typically involved only POI categories or check-in content, leading to relatively weak textual features in existing methods. In contrast, large language models (LLMs) trained on extensive text data have been found to possess rich textual knowledge. However leveraging such knowledge to enhance POI representation learning presents two key challenges: first, how to extract POI-related knowledge from LLMs effectively, and second, how to integrate the extracted information to enhance POI representations. To address these challenges, we propose POI-Enhancer, a portable framework that leverages LLMs to improve POI representations produced by classic POI learning models. We first design three specialized prompts to extract semantic information from LLMs efficiently. Then, the Dual Feature Alignment module enhances the quality of the extracted information, while the Semantic Feature Fusion module preserves its integrity. The Cross Attention Fusion module then fully adaptively integrates such high-quality information into POI representations and Multi-View Contrastive Learning further injects human-understandable semantic information into these representations. Extensive experiments on three real-world datasets demonstrate the effectiveness of our framework, showing significant improvements across all baseline representations.

Figures

Figures reproduced from arXiv: 2502.10038 by the authors.

Figure 1
Figure 1. (a): Prompt Generation and Feature Extraction are used to obtain prompts and get textual features from the LLM. (b): [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The result of ablation experiment. (A) is for POI [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The effect of L1 and L2. Skip-Gram POI2Vec Geo-Teaser TALE Hier CTLE Skip-Gram+ POI2Vec+ Geo-Teaser+ TALE+ Hier+ CTLE+ 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 NMI New York Skip-Gram POI2Vec Geo-Teaser TALE Hier CTLE Skip-Gram+ POI2Vec+ Geo-Teaser+ TALE+ Hier+ CTLE+ Singapore Skip-Gram POI2Vec Geo-Teaser TALE Hier CTLE Skip-Gram+ POI2Vec+ Geo-Teaser+ TALE+ Hier+ CTLE+ Tokyo [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The result of POI cluster task. Quality Analysis To further evaluate the quality of the en￾hanced vectors produced by POI-Enhancer, we conducted clustering tasks using the K-means algorithm on three datasets. We applied this algorithm to all types of repre￾sentation ve…
Figure 5
Figure 5. Figure 5: The examples of special prompt we design [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 32 canonical work pages

  1. [1]

    Balsebre, P.; Huang, W.; Cong, G.; and Li, Y. 2023. Cityfm: City foundation models to solve urban challenges. arXiv preprint arXiv:2310.00583

    arXiv 2023

  2. [2]

    Chang, B.; Park, Y.; Park, D.; Kim, S.; and Kang, J. 2018. Content-aware hierarchical point-of-interest embedding model for successive poi recommendation. In IJCAI, volume 20, 3301--3307

    work page 2018

  3. [3]

    Chen, Y.; Wang, X.; and Xu, G. 2023. Gatgpt: A pre-trained large language model with graph attention network for spatiotemporal imputation. arXiv preprint arXiv:2311.14332

    arXiv 2023

  4. [4]

    Dai, S.; Yu, Y.; Fan, H.; and Dong, J. 2022. Spatio-temporal representation learning with social tie for personalized POI recommendation. Data Science and Engineering, 7(1): 44--56

    work page 2022

  5. [5]

    S.; and Bao, H

    Ding, X.; Chen, L.; Gao, Y.; Jensen, C. S.; and Bao, H. 2018. UlTraMan: A unified platform for big trajectory data management and analytics. Proceedings of the VLDB Endowment, 11(7): 787--799

    work page 2018

  6. [6]

    Feng, S.; Cong, G.; An, B.; and Chee, Y. M. 2017. Poi2vec: Geographical latent representation for predicting future visitors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31

    work page 2017

  7. [7]

    GLM, T.; Zeng, A.; Xu, B.; Wang, B.; Zhang, C.; Yin, D.; Zhang, D.; Rojas, D.; Feng, G.; Zhao, H.; et al. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793

    arXiv 2024

  8. [8]

    Golkar, S.; Pettee, M.; Eickenberg, M.; Bietti, A.; Cranmer, M.; Krawezik, G.; Lanusse, F.; McCabe, M.; Ohana, R.; Parker, L.; et al. 2023. xval: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989

    arXiv 2023

Show all 50 references
  1. [9]

    Gurnee, W.; and Tegmark, M. 2024. Language Models Represent Space and Time. arXiv:2310.02207

    2024 arXiv

  2. [10]

    Han, X.; Zhou, D.-X.; Shen, G.; Kong, X.; and Zhao, Y. 2024. Deep Trajectory Recovery Approach of Offline Vehicles in the Internet of Vehicles. IEEE Transactions on Vehicular Technology, 73(11): 16051--16062

    2024

  3. [11]

    Ji, J.; Wang, J.; Huang, C.; Wu, J.; Xu, B.; Wu, Z.; Zhang, J.; and Zheng, Y. 2023. Spatio-temporal self-supervised learning for traffic flow prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 4356--4364

    2023

  4. [12]

    Ji, J.; Wang, J.; Jiang, Z.; Jiang, J.; and Zhang, H. 2022 a . STDEN: Towards physics-guided neural networks for traffic flow prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 4048--4056

    2022

  5. [13]

    Ji, J.; Wang, J.; Wu, J.; Han, B.; Zhang, J.; and Zheng, Y. 2022 b . Precision CityShield against hazardous chemicals threats via location mining and self-supervised learning. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 3072--3080

    2022

  6. [14]

    X.; and Wang, J

    Jiang, J.; Han, C.; Zhao, W. X.; and Wang, J. 2023 a . Pdformer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 4365--4373

    2023

  7. [15]

    Jiang, J.; Pan, D.; Ren, H.; Jiang, X.; Li, C.; and Wang, J. 2023 b . Self-supervised trajectory representation learning with temporal regularities and travel semantics. In 2023 IEEE 39th international conference on data engineering (ICDE), 843--855. IEEE

    2023

  8. [16]

    Jiang, J.; Pan, D.; Ren, H.; Jiang, X.; Li, C.; and Wang, J. 2023 c . Self-supervised trajectory representation learning with temporal regularities and travel semantics. In 2023 IEEE 39th international conference on data engineering (ICDE), 843--855. IEEE

    2023

  9. [17]

    X.; Wang, J.; and Jiang, J

    Jiang, W.; Zhao, W. X.; Wang, J.; and Jiang, J. 2023 d . Continuous trajectory generation based on two-stage GAN. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 4374--4382

    2023

  10. [18]

    Li, Y.; Huang, W.; Cong, G.; Wang, H.; and Wang, Z. 2023. Urban region representation learning with openstreetmap building footprints. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 1363--1373

    2023

  11. [19]

    Li, Z.; Xia, L.; Tang, J.; Xu, Y.; Shi, L.; Xia, L.; Yin, D.; and Huang, C. 2024. Urbangpt: Spatio-temporal large language models. arXiv preprint arXiv:2403.00813

    2024 arXiv

  12. [20]

    Lin, Y.; Wan, H.; Guo, S.; and Lin, Y. 2021. 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, 4241--4248

    2021

  13. [21]

    Liu, C.; Yang, S.; Xu, Q.; Li, Z.; Long, C.; Li, Z.; and Zhao, R. 2024 a . Spatial-temporal large language model for traffic prediction. arXiv preprint arXiv:2401.10134

    2024 arXiv

  14. [22]

    Liu, Z.; Wang, J.; Li, Z.; and He, Y. 2024 b . Full Bayesian Significance Testing for Neural Networks in Traffic Forecasting. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI)

    2024

  15. [23]

    Manvi, R.; Khanna, S.; Mai, G.; Burke, M.; Lobell, D.; and Ermon, S. 2023. Geollm: Extracting geospatial knowledge from large language models. arXiv preprint arXiv:2310.06213

    2023 arXiv

  16. [24]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; and Sutskever, I. 2019. Language Models are Unsupervised Multitask Learners

    2019

  17. [25]

    Shazeer, N. 2019. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150

    2019 arXiv

  18. [26]

    Shimizu, T.; Yabe, T.; and Tsubouchi, K. 2020. Enabling finer grained place embeddings using spatial hierarchy from human mobility trajectories. In Proceedings of the 28th International Conference on Advances in Geographic Information Systems, 187--190

    2020

  19. [27]

    Song, C.; Lin, Y.; Guo, S.; and Wan, H. 2020. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 34, 914--921

    2020

  20. [28]

    Sonkar, S.; and Baraniuk, R. G. 2023. Investigating the Role of Feed-Forward Networks in Transformers Using Parallel Attention and Feed-Forward Net Design. arXiv preprint arXiv:2305.13297

    2023 arXiv

  21. [29]

    Sun, F.; Qi, J.; Chang, Y.; Fan, X.; Karunasekera, S.; and Tanin, E. 2024. Urban region representation learning with attentive fusion. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), 4409--4421. IEEE

    2024

  22. [30]

    Tomas, M.; Kai, C.; Greg, C.; and Jeffrey, D. 2013. Efficient estimation of word representations in vector space. arXiv:1301.3781v3

    2013 arXiv

  23. [31]

    Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.; Babaei, Y.; Bashlykov, N.; Batra, S.; Bhargava, P.; Bhosale, S.; et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  24. [32]

    Wan, H.; Lin, Y.; Guo, S.; and Lin, Y. 2021. Pre-training time-aware location embeddings from spatial-temporal trajectories. IEEE Transactions on Knowledge and Data Engineering, 34(11): 5510--5523

    2021

  25. [33]

    J.; Xie, X.; and Xiong, Z

    Wang, J.; He, X.; Wang, Z.; Wu, J.; Yuan, N. J.; Xie, X.; and Xiong, Z. 2018. CD-CNN: a partially supervised cross-domain deep learning model for urban resident recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32

    2018

  26. [34]

    Wang, J.; Ji, J.; Jiang, Z.; and Sun, L. 2022. Traffic flow prediction based on spatiotemporal potential energy fields. IEEE Transactions on Knowledge and Data Engineering, 35(9): 9073--9087

    2022

  27. [35]

    Wang, J.; Jiang, J.; Jiang, W.; Li, C.; and Zhao, W. X. 2021 a . Libcity: An open library for traffic prediction. In Proceedings of the 29th international conference on advances in geographic information systems, 145--148

    2021

  28. [36]

    Wang, J.; Lin, X.; Zuo, Y.; and Wu, J. 2021 b . DGeye: Probabilistic risk perception and prediction for urban dangerous goods management. ACM Transactions on Information Systems (TOIS), 39(3): 1--30

    2021

  29. [37]

    Wang, J.; Wang, X.; and Wu, J. 2018. Inferring metapopulation propagation network for intra-city epidemic control and prevention. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 830--838

    2018

  30. [38]

    X.; and Jin, Y

    Wu, N.; Wang, J.; Zhao, W. X.; and Jin, Y. 2019. Learning to effectively estimate the travel time for fastest route recommendation. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, 1923--1932

    2019

  31. [39]

    Xie, M.; Yin, H.; Wang, H.; Xu, F.; Chen, W.; and Wang, S. 2016. Learning graph-based poi embedding for location-based recommendation. In Proceedings of the 25th ACM international on conference on information and knowledge management, 15--24

    2016

  32. [40]

    Yan, Y.; Wen, H.; Zhong, S.; Chen, W.; Chen, H.; Wen, Q.; Zimmermann, R.; and Liang, Y. 2024. Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web. In Proceedings of the ACM on Web Conference 2024, 4006--4017

    2024

  33. [41]

    W.; and Yu, Z

    Yang, D.; Zhang, D.; Zheng, V. W.; and Yu, Z. 2014. Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 45(1): 129--142

    2014

  34. [42]

    Yu, X.; Wang, J.; Yang, Y.; Huang, Q.; and Qu, K. 2024. BIGCity: A Universal Spatiotemporal Model for Unified Trajectory and Traffic State Data Analysis. arXiv preprint arXiv:2412.00953

    2024 arXiv

  35. [43]

    Zhang, Y.; Wei, C.; Wu, S.; He, Z.; and Yu, W. 2023 a . GeoGPT: understanding and processing geospatial tasks through an autonomous GPT. arXiv preprint arXiv:2307.07930

    2023 arXiv

  36. [44]

    J.; and Zhao, H

    Zhang, Z.; Huang, Z.; Hu, Z.; Zhao, X.; Wang, W.; Liu, Z.; Zhang, J.; Qin, S. J.; and Zhao, H. 2023 b . MLPST: MLP is All You Need for Spatio-Temporal Prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 3381--3390

    2023

  37. [45]

    S.; and Zhou, X

    Zhao, P.; Luo, A.; Liu, Y.; Xu, J.; Li, Z.; Zhuang, F.; Sheng, V. S.; and Zhou, X. 2020. Where to go next: A spatio-temporal gated network for next poi recommendation. IEEE Transactions on Knowledge and Data Engineering, 34(5): 2512--2524

    2020

  38. [46]

    Zhao, S.; Zhao, T.; King, I.; and Lyu, M. R. 2017. Geo-teaser: Geo-temporal sequential embedding rank for point-of-interest recommendation. In Proceedings of the 26th international conference on world wide web companion, 153--162

    2017

  39. [47]

    Zhu, Y.; Ye, Y.; Wu, Y.; Zhao, X.; and Yu, J. 2023. Synmob: Creating high-fidelity synthetic gps trajectory dataset for urban mobility analysis. Advances in Neural Information Processing Systems, 36: 22961--22977

    2023

  40. [48]

    J.; Zhao, X.; Liu, Q.; Ye, Y.; Chen, W.; Zhang, Z.; Wei, X.; and Liang, Y

    Zhu, Y.; Yu, J. J.; Zhao, X.; Liu, Q.; Ye, Y.; Chen, W.; Zhang, Z.; Wei, X.; and Liang, Y. 2024. Controltraj: Controllable trajectory generation with topology-constrained diffusion model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ...

    2024

  41. [49]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  42. [50]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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