Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Grid and Road Expressions Are Complementary for Trajectory Representation Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read GREEN claims that jointly encoding grid and road trajectory views yields trajectory representations that consistently beat single-view methods on three downstream tasks.

desk verdict A genuinely novel combination of grid and road trajectory views with solid TTE/TC gains, but the headline similarity-search numbers rest on a self-built benchmark that overstates the method. read the letter →

arxiv 2411.14768 v1 pith:B27ASAK5 submitted 2024-11-22 cs.LG cs.AI

classification cs.LGcs.AI
keywords TrajectoryrepresentationlearningGridRoadMultimodalContrastiveMaskedreconstructionTraveltimeestimationsimilarity
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

GREEN is built on the observation that a raw GPS trajectory can be expressed in two complementary forms: a grid trajectory, which records the regions and locations it passes through in free space, and a road trajectory, which records the sequence of road segments in the underlying network. The paper argues that existing trajectory representation learning (TRL) methods exploit only one of these forms and therefore throw away information the other form provides. Its central claim is that jointly encoding both expressions, aligning them with a contrastive loss, and using the grid expression to help reconstruct masked road segments yields trajectory vectors that are more accurate for downstream tasks than vectors learned from either expression alone. Across travel time estimation, trajectory classification, and most similar trajectory search on the Porto and Chengdu datasets, the method reports consistent gains and an average 15.99 percent improvement over the best baseline.

What carries the argument

The central machinery is the two complementary trajectory encoders plus the two alignment losses. Each raw GPS trajectory is converted into a grid trajectory and a road trajectory; the grid encoder uses a CNN on a three-channel grid map (coordinates and traffic flow) followed by a Transformer, and the road encoder uses a graph attention network on road-segment attributes followed by a Transformer whose attention is adjusted by road-type continuity. A contrastive loss encourages the grid and road embeddings of the same trajectory to land near each other, and a masked-language-model loss makes the grid representation help the model recover segments masked from the road representation; this second loss is what forces the grid view to supply information the road view lacks. The dual-modal interactor, a cross-attention layer with queries from the road view and keys and values from the grid view, fuses both views into the final vector.

What would settle it

Re-run the most similar trajectory search experiments with an independent ground truth, for example human-annotated similar trajectory pairs or a classic GPS-level similarity measure such as dynamic time warping, and check whether GREEN's mean-rank reduction over JGRM persists; if the ranking advantage shrinks or reverses, the key claim about similarity search would not survive.

Watch

Extended reading notes

Core claim

The paper's discovery is that the two trajectory expressions are not redundant but complementary, and that this complementarity can be exploited by a dual-encoder architecture. The grid encoder applies a CNN over cell traffic-flow and coordinate channels, then feeds a Transformer with trajectory-specific GPS and time features; the road encoder applies a graph attention network and a Transformer with coarse-grained day-of-week and minute-of-day encodings, and injects road-type continuity into the attention matrix. The two encoders are aligned by a contrastive loss and by a masked-language-model loss in which grid representations assist in reconstructing masked stretches of road trajectories. A dual-modal interactor then fuses the two views through cross-attention to produce the final trajectory representation. The author would state the result as: for travel time estimation, trajectory classification, and most similar trajectory search, GREEN consistently outperforms seven strong TRL baselines on two real-world datasets, with larger gains on tasks that depend directly on the learned representations.

Load-bearing premise

The load-bearing premise is that the author-built ground truth for most similar trajectory search, which downsamples GPS points by 0.5, map matches them again, and keeps only road trajectories with a change rate between 0.3 and 0.5 as positives, measures real trajectory similarity; if this heuristic favors methods that tolerate downsampling, the large reported similarity-search improvements would not reflect general accuracy.

Editorial extensions

If this is right

  • Travel time estimation improves by 12.26% to 26.79% in MAE over the best baseline on the two datasets, because the fused representation carries both region-level and road-level timing information.
  • Trajectory classification improves by smaller margins (1.26% to 3.11%), which the paper attributes to classification being an easier task, but GREEN still leads across both datasets.
  • Most similar trajectory search shows the largest relative gains on the author-built benchmark: mean rank drops by 46.99% on Porto and 55.68% on Chengdu compared with the best baseline.
  • Pre-training with the two losses matches the accuracy of training from scratch on the full Chengdu dataset using only 30% of the training data.
  • Training time is shorter than the strongest road-based baselines and inference time is comparable, so the accuracy gain does not come with an obvious computational penalty.

Reading between the lines

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

  • The grid-helps-road masking idea is a template for any pair of trajectory expressions with unequal reconstruction difficulty: the easier-to-reconstruct view (here, road segments with network continuity) should be the reconstruction target, and the harder view should be the prompt; the paper argues the opposite direction fails because discrete grids are hard to recover.
  • If the most-similar-trajectory ground truth (downsampling by 0.5 and keeping road trajectories with change rate 0.3-0.5) is replaced by an established similarity measure, GREEN's large reported improvement in similarity search would need to be rechecked, because that improvement may partly reward robustness to downsampling rather than general similarity accuracy.
  • The two-view pretraining objective could transfer to other mobility representations, such as semantic maps or point-of-interest sequences, whenever two views of the same movement carry complementary information; this extension is not tested in the paper.
  • Because the ablations show grid-only and road-only variants trade off differently across tasks (grid helps classification and travel time, road helps similarity search), a practitioner could tune the fusion weight per task rather than using fixed cross-attention fusion.
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 / 4 minor

Summary. The paper proposes GREEN, a self-supervised trajectory representation learning method that jointly uses grid and road trajectory expressions. It builds a CNN-plus-Transformer grid encoder and a GNN-plus-Transformer road encoder, trains them with a contrastive loss and a masked-language-model-style loss in which grid trajectories assist the reconstruction of masked road segments, and fuses the two views with a cross-attention interactor. The evaluation on Porto and Chengdu compares GREEN with seven baselines on travel time estimation, trajectory classification, and most similar trajectory search, and reports consistent improvements, together with ablations, a pre-training data-efficiency study, an efficiency comparison, and a transferability analysis.

Significance. The core observation that grid and road expressions carry complementary information is well motivated, and the model is described in sufficient detail for reproduction. The self-supervised losses are not fitted to downstream labels, and the pre-training data-efficiency experiment in Figure 5 is a useful sanity check. If the empirical claims are validated, the work would be a solid contribution to trajectory representation learning, and the public code and data release are a clear strength. However, the evaluation currently contains a load-bearing caveat around the most-similar-trajectory-search benchmark and lacks statistical grounding, so the magnitude and generality of the claimed improvements are not yet firmly established.

major comments (4)
  1. [Appendix A.3; Tables 2-3] The 'most similar trajectory search' benchmark is constructed by taking each query's own raw GPS trajectory, keeping the endpoints, downsampling the remaining points by 0.5, re-map-matching, and retaining candidates whose road change rate is between 0.3 and 0.5. As constructed, this task measures robustness to a specific downsampling-and-map-matching perturbation rather than general trajectory similarity, and the ground truth is not validated against any external similarity measure. Because GREEN is explicitly trained to align the grid and road expressions of the same raw trajectory, it has a structural advantage on exactly this kind of perturbation. The MR reductions of 46.99% and 55.68% therefore cannot be interpreted as evidence that GREEN retrieves more semantically similar trajectories in general. Moreover, the abstract's headline average improvement of 15.99% is dominated by these two MR entries; excluding them leaves roughly 11-12%. Please either reposition the MSTS claim as a downsampled-trajectory retrieval result or validate the benchmark against independent ground truth, for example human judgments or a standard trajectory similarity measure such as DTW/LCSS.
  2. [Section 5.1 and Appendix A.4; Tables 2-4] All accuracy numbers are single runs with no standard deviations, confidence intervals, or significance tests, and the hyperparameter study in Appendix A.4 appears to select values (grid cell size, mask ratio and length, dimensions, layer counts) using the same evaluation metrics reported in the main tables, without any statement that a separate validation split was used. With improvements as small as 1.26% and 1.79% on Chengdu trajectory classification, it is difficult to tell whether the reported margins are genuine or are artifacts of run-to-run variance and test-set selection. Please report means and standard deviations over multiple seeds, and state explicitly whether any hyperparameter was chosen on the test set.
  3. [Section 5.3, Table 5] The transferability section claims that the transferred model 'still outperforms most baselines,' but Table 5 only compares the w/ Transfer and w/o Transfer variants and contains no baseline rows. Either add the corresponding baseline results to the table or soften the claim, because the current statement is not supported by the reported evidence.
  4. [Section 5.3, Table 4] The 'w/o MLM loss' variant removes both the MLM loss and the dual-modal interactor, replacing the learned fusion with a simple average of the two encoder outputs. The observed degradation therefore conflates the effect of the MLM objective with the effect of the cross-attention fusion module, so the text's conclusion that 'the MLM loss is more important for trajectory classification than the contrastive loss' is not supported by this ablation. A cleaner ablation would remove only the MLM loss while retaining the interactor, or otherwise separate the two design choices.
minor comments (4)
  1. [Section 4.4, Eq. (14)] Equation (14) appears to compute the cross-entropy over all positions in the road trajectory, while the text states that only masked road segments are predicted. Please clarify whether unmasked positions contribute to the loss and how masked positions are represented in the output.
  2. [Section 4.4, Figure 3] The mask-recovery curves in Figure 3 are used to justify the asymmetric MLM design, but the experimental setup for this figure is not described anywhere (dataset split, number of runs, definition of accuracy on masked tokens). Please add this information so the figure can be interpreted.
  3. [Section 4.1] The statement that 'less than 1% of the grids under 100m×100m in the dataset have more than one GPS point' is presented without a derivation or a specified dataset; please provide the computation or a citation backing this statistic.
  4. [Section 5.2, Tables 2-3] The term 'accuracy' in the abstract is used as an umbrella for MAE/MAPE/RMSE, F1 scores, and retrieval metrics; consider replacing it with a more neutral phrase such as 'performance' to avoid implying a single accuracy measure.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; self-supervised losses are not fitted to downstream labels, and the self-constructed MSTS benchmark is an external-validity caveat rather than a circular derivation.

full rationale

GREEN’s derivation is self-contained: the grid and road encoders are defined on the two expressions of the same GPS trajectories (Eqs. 1–10), the contrastive and MLM losses (Eqs. 12, 14) are optimized over raw trajectories without downstream label information, and the TTE/TC results are obtained by fine-tuning on held-out splits while MSTS uses the pre-trained representations directly. The one notable caveat is the MSTS benchmark in Appendix A.3, where the paper itself states "the lack of ground truth makes it difficult to evaluate the accuracy" and then generates positives by keeping the query’s start/end points, downsampling the other GPS points by 0.5, re-map-matching, and filtering to change rates 0.3–0.5. This makes the large MR reductions (46.99% Porto, 55.68% Chengdu) a measure of robustness to that specific downsampling-and-map-matching perturbation rather than of general semantic similarity; that is a validity limitation, not a circular derivation, because GREEN is not fitted to the benchmark and the ranking is evaluated on held-out queries. The authors’ prior works ([16], [42]) are cited only as related work and are not load-bearing for the model’s design or for any uniqueness claim. I find no equation or fitted parameter that, by the paper’s own construction, is equivalent to the reported prediction; the score of 2 reflects the minor self-citations and the benchmark caveat rather than a true circular step.

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

The paper is an empirical systems paper; the central claim rests on dataset representativeness, map-matching correctness, and a hand-constructed similarity benchmark. The only tuned quantities are hyperparameters, and no new physical or conceptual entities are introduced.

free parameters (6)
  • grid_cell_size = 100m x 100m
    Appendix A.4 tests 50/100/200m on Porto and selects 100m as default; affects the TTE/MSTS tradeoff.
  • mask_ratio = 0.2
    Appendix A.4 tests 0.1/0.2/0.3 and selects 0.2; optimal at intermediate difficulty.
  • mask_length = 2
    Appendix A.4 tests 1/2/3 and selects 2; continuous masking of length at least 2.
  • embedding_dimension = 128
    Appendix A.4 tests 16-256 and selects 128; larger dimension overfits TTE.
  • contrastive_temperature = 0.07
    Set in Appendix A.2; standard CLIP-style temperature.
  • msts_downsample_ratio_and_change_rate = downsample 0.5, change rate [0.3, 0.5]
    Appendix A.3; hand-chosen threshold to define positive matches in the similarity-search benchmark.
assumptions (6)
  • domain assumption The Porto and Chengdu datasets and their OSMnx-derived road networks are representative of real-world urban trajectory data.
    The paper generalizes from these two taxi datasets to claim consistent improvements; no other domains are tested.
  • domain assumption The map matching algorithm [38] produces road trajectories that correctly reflect the actual routes taken.
    Road trajectories are obtained by map matching; errors in matching would propagate to the road encoder and all comparisons.
  • domain assumption Adjacent road segments in a trajectory usually have the same road type, so injecting road type into attention is beneficial.
    Section 4.2, Eq. (9) and footnote 3; this design choice is not ablated.
  • domain assumption The constructed similarity-search ground truth (downsampling by 0.5, change-rate filter 0.3-0.5) identifies genuinely similar trajectories.
    Appendix A.3; the threshold is arbitrary and unvalidated.
  • domain assumption Pre-training is performed on the training split only, so no test information leaks into the learned representations.
    The paper does not explicitly state the pre-training data split; if pre-training uses the full dataset, the downstream results would be optimistic.
  • standard math Standard stochastic training (Adam, dropout, learned temperature) converges to a good optimum for the self-supervised objectives.
    Relies on standard ML practice; no convergence guarantees.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Grid and Road Expressions Are Complementary for Trajectory Representation Learning." pith.science (2026). https://pith.science/paper/B27ASAK5

@misc{pith2026241114768,
  author       = {Pith},
  title        = {Pith review of: Grid and Road Expressions Are Complementary for Trajectory Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B27ASAK5}},
  note         = {Machine review of arXiv:2411.14768}
}
read the original abstract

Trajectory representation learning (TRL) maps trajectories to vectors that can be used for many downstream tasks. Existing TRL methods use either grid trajectories, capturing movement in free space, or road trajectories, capturing movement in a road network, as input. We observe that the two types of trajectories are complementary, providing either region and location information or providing road structure and movement regularity. Therefore, we propose a novel multimodal TRL method, dubbed GREEN, to jointly utilize Grid and Road trajectory Expressions for Effective representatioN learning. In particular, we transform raw GPS trajectories into both grid and road trajectories and tailor two encoders to capture their respective information. To align the two encoders such that they complement each other, we adopt a contrastive loss to encourage them to produce similar embeddings for the same raw trajectory and design a mask language model (MLM) loss to use grid trajectories to help reconstruct masked road trajectories. To learn the final trajectory representation, a dual-modal interactor is used to fuse the outputs of the two encoders via cross-attention. We compare GREEN with 7 state-of-the-art TRL methods for 3 downstream tasks, finding that GREEN consistently outperforms all baselines and improves the accuracy of the best-performing baseline by an average of 15.99\%.

Figures

Figures reproduced from arXiv: 2411.14768 by the authors.

Figure 1
Figure 1. Comparison of three types of trajectories. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of GREEN. where 𝜎 is the relu [1] activation function, Conv1 and Conv2 are convolution kernels of size 3 × 3. Note that X 𝑔 2 ∈ R 𝐻 ×𝑊 ×𝐶 ′ , where 𝐶 ′ is the number of output channels. To learn grid embeddings with hidden dimension ℎ, we reshape X 𝑔 2 into a 2D-tensor, and use a 2-layer MLP to map it: X 𝑔 = MLP(Reshape(X 𝑔 2 )), (2) where X 𝑔 ∈ R (𝐻 ·𝑊 )×ℎ . We take X 𝑔 as the grid embedding table and tran… view at source ↗
Figure 3
Figure 3. Comparison of grid and road mask recovery. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of the top-3 similar trajectories retrieved by GREEN and JGRM from Porto dataset for a query trajectory. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The effect of GREEN pre-training on Chengdu. X [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Effect of embedding dimension in Porto. 0.1 0.2 0.3 Mask ratio 1 2 3 Mask length 1.419 1.494 1.398 1.415 1.352 1.372 1.390 1.436 1.384 (a) Mask on travel time 2 4 6 Layer 1.34 1.36 1.38 1.40 1.42 MAE (b) Layer on travel time Road Grid Interactor [PITH_FULL_IMAGE:figur…
Figure 8
Figure 8. Figure 8: Effect of mask and layer in Porto. representations, and over-fitting happens when there are too many parameters at large embedding dimension. Mask. We vary the mask length among [1, 2, 3] and the mask ratio among [0.1, 0.2, 0.3] when masking the road trajectory for the…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Learning Generalized and Flexible Trajectory Models from Omni-Semantic Supervision

    cs.CV 2025-05 conditional novelty 6.0 of 10

    OmniTraj aligns four trajectory modalities (raw path, topology, road segments, regions) in a shared embedding space, enabling flexible condition-based retrieval that outperforms similarity-only baselines on Chengdu an...

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abien Fred Agarap. 2018. Deep Learning using Rectified Linear Units (ReLU). In arXiv preprint. http://arxiv.org/abs/1803.08375

  2. [2]

    Berndt and James Clifford

    Donald J. Berndt and James Clifford. 1994. Using Dynamic Time Warping to Find Patterns in Time Series. In AAAI. 359–370

  3. [3]

    Geoff Boeing. 2017. OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks. Computers, environment and urban systems 65 (2017), 126–139

  4. [4]

    Yanchuan Chang, Jianzhong Qi, Yuxuan Liang, and Egemen Tanin. 2023. Con- trastive Trajectory Similarity Learning with Dual-Feature Attention. In ICDE. 2933–2945

  5. [5]

    Lei Chen and Raymond T. Ng. 2004. On The Marriage of Lp-norms and Edit Distance. In VLDB. 792–803

  6. [6]

    Tamer Özsu, and Vincent Oria

    Lei Chen, M. Tamer Özsu, and Vincent Oria. 2005. Robust and Fast Similarity Search for Moving Object Trajectories. In SIGMOD. 491–502

  7. [7]

    Wei Chen, Yuxuan Liang, Yuanshao Zhu, Yanchuan Chang, Kang Luo, Haomin Wen, Lei Li, Yanwei Yu, Qingsong Wen, Chao Chen, Kai Zheng, Yunjun Gao, Xiaofang Zhou, and Yu Zheng. 2024. Deep Learning for Trajectory Data Management and Mining: A Survey and Beyond. arXiv preprint (2024), https://arxiv.org/abs/2403.14151

  8. [8]

    Yile Chen, Xiucheng Li, Gao Cong, Zhifeng Bao, Cheng Long, Yiding Liu, Arun Ku- mar Chandran, and Richard Ellison. 2021. Robust Road Network Representation Learning: When Traffic Patterns Meet Traveling Semantics. In CIKM. 211–220

Show all 44 references
  1. [9]

    Yuqi Chen, Hanyuan Zhang, Weiwei Sun, and Baihua Zheng. 2023. RNTrajRec: Road Network Enhanced Trajectory Recovery with Spatial-Temporal Trans- former. In ICDE. 829–842

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL. 4171–4186

  3. [11]

    Ziquan Fang, Yuntao Du, Lu Chen, Yujia Hu, Yunjun Gao, and Gang Chen. 2021. E2DTC: An End to End Deep Trajectory Clustering Framework via Self-Training. In ICDE. 696–707

  4. [12]

    Ziquan Fang, Yuntao Du, Xinjun Zhu, Danlei Hu, Lu Chen, Yunjun Gao, and Christian S. Jensen. 2022. Spatio-Temporal Trajectory Similarity Learning in Road Networks. In KDD, Aidong Zhang and Huzefa Rangwala (Eds.). 347–356

  5. [13]

    Tao-Yang Fu and Wang-Chien Lee. 2020. Trembr: Exploring road networks for trajectory representation learning. ACM Transactions on Intelligent Systems and Technology 11, 1 (2020), 1–25

  6. [14]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for Networks. In KDD. 855–864

  7. [15]

    Jindong Han, Hao Liu, Shui Liu, Xi Chen, Naiqiang Tan, Hua Chai, and Hui Xiong. 2023. iETA: A Robust and Scalable Incremental Learning Framework for Time-of-Arrival Estimation. In KDD. 4100–4111

  8. [16]

    Peng Han, Jin Wang, Di Yao, Shuo Shang, and Xiangliang Zhang. 2021. A Graph- based Approach for Trajectory Similarity Computation in Spatial Networks. In KDD. 556–564

  9. [17]

    Jiawei Jiang, Dayan Pan, Houxing Ren, Xiaohan Jiang, Chao Li, and Jingyuan Wang. 2023. Self-supervised Trajectory Representation Learning with Temporal Regularities and Travel Semantics. In ICDE. 843–855

  10. [18]

    Brubaker

    Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus A. Brubaker. 2019. Time2vec: Learning a vector representation of time. In arXiv preprint. https://arxiv.org/abs/1907.05321

  11. [19]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. In ICLR. https://arxiv.org/abs/1412.6980

  12. [20]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR. https://arxiv.org/abs/1609.02907

  13. [21]

    Tian Lan, Ziyue Li, Zhishuai Li, Lei Bai, Man Li, Fugee Tsung, Wolfgang Ketter, Rui Zhao, and Chen Zhang. 2023. MM-DAG: Multi-task DAG Learning for Multi-modal Data - with Application for Traffic Congestion Analysis. In KDD. 1188–1199

  14. [22]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient- based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278– 2324

  15. [23]

    Selvaraju, Akhilesh Gotmare, Shafiq R

    Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. 2021. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. In NeurIPS. 9694–9705

  16. [24]

    Jensen, and Wei Wei

    Xiucheng Li, Kaiqi Zhao, Gao Cong, Christian S. Jensen, and Wei Wei. 2018. Deep Representation Learning for Trajectory Similarity Computation. In ICDE. 617–628

  17. [25]

    Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He

  18. [26]

    Yuxuan Liang, Kun Ouyang, Hanshu Yan, Yiwei Wang, Zekun Tong, and Roger Zimmermann. 2021. Modeling Trajectories with Neural Ordinary Differential Equations. In IJCAI. 1498–1504

  19. [27]

    Xiang Liu, Xiaoying Tan, Yuchun Guo, Yishuai Chen, and Zhe Zhang. 2022. CSTRM: Contrastive Self-Supervised Trajectory Representation Model for trajec- tory similarity computation. Computer Communications 185 (2022), 159–167

  20. [28]

    Yiding Liu, Kaiqi Zhao, Gao Cong, and Zhifeng Bao. 2020. Online Anomalous Trajectory Detection with Deep Generative Sequence Modeling. In ICDE. 949– 960

  21. [29]

    Yin Lou, Chengyang Zhang, Yu Zheng, Xing Xie, Wei Wang, and Yan Huang

  22. [30]

    Zhipeng Ma, Zheyan Tu, Xinhai Chen, Yan Zhang, Deguo Xia, Guyue Zhou, Yilun Chen, Yu Zheng, and Jiangtao Gong. 2024. More Than Routing: Joint GPS and Route Modeling for Refine Trajectory Representation Learning. In WWW. 3064–3075

  23. [31]

    Zhenyu Mao, Ziyue Li, Dedong Li, Lei Bai, and Rui Zhao. 2022. Jointly contrastive representation learning on road network and trajectory. In CIKM. 1501–1510

  24. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In ICML. 8748–8763

  25. [33]

    Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to Sequence Learning with Neural Networks. In NIPS. 3104–3112

  26. [34]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In NIPS. 5998–6008

  27. [35]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In ICLR. https://arxiv.org/abs/1710.10903

  28. [36]

    Yu Wang, Tongya Zheng, Yuxuan Liang, Shunyu Liu, and Mingli Song. 2024. COLA: Cross-city Mobility Transformer for Human Trajectory Simulation. In WWW. 3509–3520

  29. [37]

    Chengcheng Yang, Lisi Chen, Hao Wang, and Shuo Shang. 2021. Towards Efficient Selection of Activity Trajectories based on Diversity and Coverage. In AAAI. 689–696

  30. [38]

    Can Yang and Gyözö Gidófalvi. 2018. Fast map matching, an algorithm integrating hidden Markov model with precomputation.International Journal of Geographical Information Science 32, 3 (2018), 547–570

  31. [39]

    Sean Bin Yang, Chenjuan Guo, Jilin Hu, Jian Tang, and Bin Yang. 2021. Unsu- pervised Path Representation Learning with Curriculum Negative Sampling. In IJCAI. 3286–3292

  32. [40]

    Di Yao, Haonan Hu, Lun Du, Gao Cong, Shi Han, and Jingping Bi. 2022. Traj- GAT: A Graph-based Long-term Dependency Modeling Approach for Trajectory Similarity Computation. In KDD. 2275–2285

  33. [41]

    Di Yao, Chao Zhang, Zhihua Zhu, Jian-Hui Huang, and Jingping Bi. 2017. Trajec- tory clustering via deep representation learning. In IJCNN. 3880–3887

  34. [42]

    Silin Zhou, Jing Li, Hao Wang, Shuo Shang, and Peng Han. 2023. GRLSTM: Trajectory Similarity Computation with Graph-Based Residual LSTM. In AAAI. 4972–4980. A Appendix A.1 Baselines GPS-based trajectory representation learning: • Traj2vec7: An RNN-based seq2seq model converts ...

  35. [2009]

    In SIGSPATIAL

    Map-matching for low-sampling-rate GPS trajectories. In SIGSPATIAL. 352–361

  36. [2023]

    Scaling Language-Image Pre-Training via Masking. In CVPR. 23390–23400

Pith tools

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