Pith. sign in

REVIEW 4 major objections 7 minor 49 references

TrajDiff: Diffusion Bridge Network with Semantic Alignment for Trajectory Similarity Computation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read TrajDiff argues that learned trajectory similarity needs three fixes: aligning GPS and grid semantics, denoising via a diffusion bridge, and ranking-aware training.

desk verdict Plausible engineering paper with large reported gains, but the bridge-specific pretraining contribution is not isolated and the manuscript is not yet reproducible. read the letter →

arxiv 2506.15898 v1 pith:JEWDLSCU submitted 2025-06-18 cs.LG

classification cs.LG
keywords trajectorysimilaritycomputationdenoisingdiffusionbridgemodelssemanticalignmentcross-attentionranking-awareregularizationembeddingGPSandgridfeaturefusiontop-kretrieval
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

TrajDiff claims that three obstacles prevent learned trajectory similarity from matching geometric measures: GPS and grid features live on different semantic scales, trajectory noise is amplified by grid discretization, and pointwise or pairwise losses ignore global ranking. The paper's answer is a single framework that aligns the two feature scales with a cross-attention module, pretrains the encoder with a denoising diffusion bridge between arbitrary trajectory pairs, and adds a listwise ranking regularizer. On Porto, GeoLife, and T-Drive it reports an average HR@1 gain of 33.38% over eight baselines across SSPD, discrete Fréchet, and Hausdorff, with a maximum gain of 92.99% on T-Drive under SSPD. The practical upshot, if correct, is a single model that approximates all three heuristic distances at near-constant inference cost.

What carries the argument

The load-bearing machinery is the dual semantic alignment attention inside stacked SALayers: each layer computes cross-attention from one feature scale to the other, symmetrically reverses the roles, and fuses the two attention maps with learnable scaling factors $\lambda_{\text{self}}$ and $\lambda_{\text{cross}}$. Pretraining uses a denoising diffusion bridge model (DDBM), a diffusion process pinned at two endpoints via Doob's $h$-transform; here the endpoints are two arbitrary trajectories, and the intermediate state is a Gaussian interpolation of their GPS features whose mean is given by Eq. (17). The model learns to reconstruct this interpolated state in embedding space. Fine-tuning then combines MSE with two listwise objectives, ListNet and a rank-decay weighted version, so the model is supervised by both absolute similarity magnitudes and the global ordering of candidates.

What would settle it

On a data split where training pairs are forced to be spatially disjoint or semantically unrelated, remove DDBM pretraining and compare HR@1; if the gap disappears, the bridge-interpolation assumption is doing the work attributed to it. A second check is to replace the interpolated bridge state with pure Gaussian noise and see whether the pretraining gain survives.

Watch

Extended reading notes

Core claim

The central claim is that trajectory similarity is best learned not from one feature scale or one loss, but from a joint objective: a semantic alignment module that fuses coarse grid and fine GPS representations, a DDBM pretraining step that learns stochastic transitions between trajectory pairs as a noise-robustness signal, and overall ranking-aware regularization that supervises the global order of candidate trajectories. The paper reports that this combination consistently beats eight baselines across three datasets and three heuristic similarity measures, with the largest single improvement on the noisier T-Drive dataset. It further claims that the ranking regularizer alone lifts HR@1 by 23.52%, 55.69%, and 21.24% when added to TrajCL and T3S on the three datasets, which it reads as evidence that each component is independently useful.

Load-bearing premise

The DDBM pretraining assumes that the linear interpolation between the GPS features of two arbitrary trajectories is a meaningful intermediate trajectory state, so that denoising this interpolated state teaches representations useful for similarity; if arbitrary trajectory pairs do not lie on a shared semantic manifold, this signal may just teach the encoder to reconstruct meaningless averages.

Editorial extensions

If this is right

  • A single TrajDiff model can serve as a fast proxy for SSPD, discrete Fréchet, and Hausdorff similarity in trajectory retrieval.
  • Adding the ranking-aware regularizer to existing trajectory encoders improves their HR@1 and speeds convergence, making the regularizer a portable training component.
  • DDBM pretraining matters most on noisy datasets; on relatively clean Porto the gain is smaller, indicating that noise robustness is the pretraining's main contribution.
  • The ablated single-scale and naive-fusion variants perform substantially worse, so cross-scale semantic alignment is the largest single contributor to the reported accuracy.

Reading between the lines

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

  • If the bridge interpolation assumption is valid, restricting pretraining pairs to trajectories that share spatial context could sharpen the learned manifold; if it is not, the gains attributed to DDBM may come from the contrastive endpoint structure rather than from denoising.
  • The listwise regularization is a generic recipe: it can be added to any embedding-based similarity model whose training has access to a full ranking of candidates.
  • Because the model is pretrained on Porto and fine-tuned on GeoLife and T-Drive, the reported gains on those datasets also test cross-city transfer; a same-city pretraining comparison would isolate that effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes TrajDiff, a learnable trajectory similarity computation framework combining three components: a Semantic Alignment Module (SAM) that fuses GPS and grid features through dual cross-attention with adaptive masking, a DDBM-based pretraining stage that denoises interpolated states between pairs of trajectories, and an overall ranking-aware regularization based on ListNet and Rank-Decay ListNet losses. The method is evaluated on Porto, Geolife, and T-Drive against several baselines under SSPD, Discrete Fréchet, and Hausdorff metrics, reporting consistent HR@1 and Recall gains, with an average HR@1 improvement of 33.38% claimed in the abstract. The authors also provide ablations, hyperparameter sensitivity, efficiency, and convergence analyses, and release code.

Significance. If substantiated, TrajDiff would be a useful contribution to trajectory similarity learning: the dual-scale semantic alignment addresses a real limitation of existing single-scale or additive fusion methods, the listwise ranking objective is a sensible fit for the retrieval-oriented evaluation metrics, and the pretraining idea is novel in this application area. The paper also has notable strengths: experiments span three public datasets and three different heuristic similarity labels, include significance tests, and the authors provide code. However, the central causal claim about the DDBM bridge pretraining currently rests on an unvalidated interpolation assumption, and several reporting gaps prevent full verification of the claimed improvements. The significance of the result is therefore conditional on resolving these issues.

major comments (4)
  1. [Section IV-B2, Eq. (17)] The DDBM bridge is defined between two arbitrary trajectories T_0 and T_T, but Eq. (17) requires the intermediate state e_t_gps to be a convex combination of e_0_gps and e_T_gps at corresponding sequence positions. Table II reports trajectory lengths ranging from 20 to 300, yet the paper never states how sequences are padded, resampled, or aligned before the bridge is applied. If sequences are padded, the interpolated state is partly padding and is not a meaningful trajectory state; if they are not padded, Eq. (17) is undefined for unequal lengths. This is a load-bearing assumption for the DDBM pretraining contribution, and the manuscript must specify the preprocessing and provide evidence that the interpolated states are semantically reasonable (for example, visualizations or a quantitative check of the interpolated states against real sub-trajectories).
  2. [Section V-C, Fig. 4] The 'w/o Bridge' ablation removes the entire pretraining stage, which simultaneously changes the pretraining objective, the input distribution, and the presence of any denoising signal. As a result, the observed drop in HR@1 cannot be attributed specifically to the diffusion bridge or to the trajectory-to-trajectory interpolation in Eq. (17); a generic denoising autoencoder or a single-trajectory diffusion pretraining could plausibly produce a similar effect. The authors should add control ablations, such as standard Gaussian denoising, interpolation without a diffusion process, or scrambled endpoint pairs, and report the numerical HR@1 differences, since this is the main evidence for the second stated contribution.
  3. [Section V-A3 and Table III] The paper states that each experiment was run three times with random seeds 18, 66, and 108 and that the average is reported, but Table III gives no standard deviations. The significance asterisks in Table III are also not accompanied by a description of the test procedure (e.g., paired vs. unpaired, what is being compared, and how the multiple metrics are handled). Because the abstract's central claim is that TrajDiff 'consistently outperforms' baselines, the absence of variance information and test specification prevents verification of that claim. The authors should report mean±std for all metrics and specify the statistical test used.
  4. [Section IV-C and V-A3] The construction of the candidate lists for ListNet and RD-ListNet is unspecified: the list size k, the sampling strategy for candidates per query trajectory, and the mix of ranking examples with MSE examples in each batch are not given. Similarly, the DDBM pretraining section does not state how trajectory pairs are sampled, how many diffusion timesteps are used, or the values of beta_min and beta_max in the linear schedule of Eq. (18). These are required to reproduce the main experiments and to interpret the contribution of the ranking regularization and the pretraining component.
minor comments (7)
  1. [Eq. (13) and Fig. 5(b)] Equation (13) introduces the fusion weight as epsilon but the surrounding text says 'where µ∈(0,1)'; Fig. 5(b) uses epsilon. Please make the notation consistent.
  2. [Section V-A1] The dataset name is misspelled as 'Proto' in the paragraph and as 'T-Driver' in Table II; the paper should use 'Porto' and 'T-Drive' consistently.
  3. [Section V-A4] The evaluation metrics paragraph mentions 'H5@20', but Table III reports the same metric as 'R5@20'; please unify the name.
  4. [Section V-A2] NeuTraj is listed as a baseline in Section V-A2 but does not appear in Table III or anywhere in the experimental results; the authors should either include it or remove it from the baseline list.
  5. [Section V-A2] The description of KGTS appears to describe a traffic forecasting graph neural network rather than the cited trajectory similarity paper [20]; please correct the baseline description.
  6. [Figures 3 and 7] The captions of Figures 3 and 7 contain Chinese text (e.g., 'TrajCL 添加LIST'), and the subfigure labels in Figure 3 are unclear; please translate the captions and clarify the subfigure contents.
  7. [Algorithm 1] The input line says 'per-encoding module PE' but the body and Section IV use 'pre-encoding'; please fix the typo.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported gains come from held-out supervised approximation of heuristic distances, and DDBM pretraining is self-supervised; self-citations are background only.

full rationale

The derivation chain is self-contained. The problem statement (Eq. 1) asks f_theta to approximate a heuristic d(Ti,Tj), and fine-tuning uses a 7:1:2 split, so HR@k on test queries is not a fitted-input-as-prediction loop. DDBM pretraining (Eqs. 17-21) generates intermediate states from paired trajectories without using heuristic labels and reconstructs the noiseless interpolation in embedding space; this is self-supervised and does not encode the evaluation metric. The ranking losses (Eqs. 22-27) are supervised objectives on heuristic labels, which is standard metric learning rather than circular bootstrapping. No equation defines the predicted quantity in terms of the very input it is supposed to predict. Self-citations to prior trajectory work (e.g., HHL-Traj [31], Traj2SimVec [30]) appear only in related work and are not load-bearing for the central claims. The skeptical concern that Eq. 17's coordinate-wise interpolation may not be semantically meaningful, and that the 'w/o Bridge' ablation does not isolate DDBM-specific gains, is an experimental validity issue, not a circularity issue.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The central claim rests on established diffusion and ranking theory, dataset-specific hyperparameter choices, and the unvalidated semantic-bridge assumption. No new physical entities or external postulates are introduced. The most consequential uncharged input is the assumption that interpolating two arbitrary trajectories in feature space is a meaningful learning signal.

free parameters (10)
  • gamma_1 (ListNet weight) = 0.1
    Balances the ListNet loss in Eq. (27); tuned via sensitivity analysis on Porto.
  • gamma_2 (RD-ListNet weight) = 0.001
    Balances the rank-decay ListNet loss in Eq. (27); tuned via sensitivity analysis on Porto.
  • epsilon (GPS-grid fusion weight) = 0.5
    Controls the weighted combination of Z_gps and Z_grid in Eq. (13); selected on Porto after sensitivity analysis in Fig. 5(b).
  • grid_cell_size = 100
    Determines the uniform grid discretization; selected on Porto after sensitivity analysis in Fig. 5(c) and reused for Beijing datasets.
  • num_SAM_layers = 1
    Depth of the semantic alignment module; set to 1 in the experiments without a reported study.
  • batch_size = 128
    Training batch size for pretraining and fine-tuning; reported in Section V-A3.
  • learning_rate = 0.001
    Optimizer learning rate reported in Section V-A3.
  • attention_heads = 16
    Number of attention heads reported in Section V-A3.
  • DDBM noise schedule beta_min and beta_max
    The linear noise schedule is introduced in Eq. (18), but the numerical values of beta_min and beta_max are not reported.
  • ranking_candidate_list_size_k
    The size of the candidate list for ListNet and RD-ListNet in Eqs. (22)-(26) is never specified in Section IV-C or the experiments.
assumptions (5)
  • domain assumption Supervised heuristic distances are valid ground truth for trajectory similarity
    Defined in the Problem statement and used to generate labels H for training and evaluation; if SSPD, Fréchet, or Hausdorff do not capture application-relevant similarity, the model's ranking is only as good as the metric.
  • ad hoc to paper DDBM bridge between arbitrary trajectory pairs is semantically meaningful
    Section IV-B2 assumes interpolating between GPS features of two arbitrary trajectories with Eq. (17) produces a useful signal for learning state transitions; this is not validated outside the paper.
  • standard math Standard diffusion bridge theory imports to embedding space
    Equations (14)-(16) rely on Doob h-transform and Gaussian bridge marginals from DDBM [46]; no re-derivation or domain-specific validation is provided, but the mathematics is standard.
  • domain assumption Grid cell size 100 generalizes across cities
    Cell size is selected on Porto via the sensitivity study in Fig. 5(c) and applied without re-tuning to Beijing datasets, where geographic scale and sampling density differ.
  • domain assumption Mean pooling of fused embeddings preserves similarity ranking
    Fine-tuning Algorithm 1 pools token embeddings to a single vector and uses vector distance as similarity; this assumes information needed for ranking is preserved by averaging, which is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TrajDiff: Diffusion Bridge Network with Semantic Alignment for Trajectory Similarity Computation." pith.science (2026). https://pith.science/paper/JEWDLSCU

@misc{pith2026250615898,
  author       = {Pith},
  title        = {Pith review of: TrajDiff: Diffusion Bridge Network with Semantic Alignment for Trajectory Similarity Computation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEWDLSCU}},
  note         = {Machine review of arXiv:2506.15898}
}
read the original abstract

With the proliferation of location-tracking technologies, massive volumes of trajectory data are continuously being collected. As a fundamental task in trajectory data mining, trajectory similarity computation plays a critical role in a wide range of real-world applications. However, existing learning-based methods face three challenges: First, they ignore the semantic gap between GPS and grid features in trajectories, making it difficult to obtain meaningful trajectory embeddings. Second, the noise inherent in the trajectories, as well as the noise introduced during grid discretization, obscures the true motion patterns of the trajectories. Third, existing methods focus solely on point-wise and pair-wise losses, without utilizing the global ranking information obtained by sorting all trajectories according to their similarity to a given trajectory. To address the aforementioned challenges, we propose a novel trajectory similarity computation framework, named TrajDiff. Specifically, the semantic alignment module relies on cross-attention and an attention score mask mechanism with adaptive fusion, effectively eliminating semantic discrepancies between data at two scales and generating a unified representation. Additionally, the DDBM-based Noise-robust Pre-Training introduces the transfer patterns between any two trajectories into the model training process, enhancing the model's noise robustness. Finally, the overall ranking-aware regularization shifts the model's focus from a local to a global perspective, enabling it to capture the holistic ordering information among trajectories. Extensive experiments on three publicly available datasets show that TrajDiff consistently outperforms state-of-the-art baselines. In particular, it achieves an average HR@1 gain of 33.38% across all three evaluation metrics and datasets.

Figures

Figures reproduced from arXiv: 2506.15898 by the authors.

Figure 1
Figure 1. Many-to-one mapping issues caused by feature mis [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DDBM-based Noise-robust Pre-Training. Definition 3 (Trajectory Grid Sequence). A trajectory grid sequence, denoted as Tg, is the sequence of grid cells obtained by mapping each point in a trajectory T = [p1, p2, . . . , pn] to its corresponding grid cell in the grid set G: Tg = [gi1,j1 , gi2,j2 , . . . , gin,jn ] ∈ R n×2 , where gin,jn is the grid cell containing pn. Definition 4 (Heuristic Similarity). Heuristic me… view at source ↗
Figure 3
Figure 3. Performance improvement T3S/TrajCL with our [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: The impact of hyperparameters on the performance of [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Convergence speed of T3S/TrajCL with our [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Convergence speed of different methods. 3) Convergence Speed Comparison: We evaluate and com￾pare the convergence speeds of various representative trajec￾tory learning methods on the Porto dataset, utilizing three widely adopted evaluation metrics to ensure a comprehen…
Figure 6
Figure 6. Figure 6: Comparison of inference speed across diverse models. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 32 canonical work pages

  1. [1]

    Subtrajectory clustering: Models and algorithms,

    P. K. Agarwal, K. Fox, K. Munagala, A. Nath, J. Pan, and E. Taylor, “Subtrajectory clustering: Models and algorithms,” inProceedings of the 37th ACM SIGMOD-SIGACT-SIGAI symposium on principles of database systems, 2018, pp. 75–87

  2. [2]

    Fast large-scale trajectory clustering,

    S. Wang, Z. Bao, J. S. Culpepper, T. Sellis, and X. Qin, “Fast large-scale trajectory clustering,”Proceedings of the VLDB Endowment, vol. 13, no. 1, pp. 29–42, 2019

  3. [3]

    Clustering and aggregating clues of trajectories for mining trajectory patterns and routes,

    C.-C. Hung, W.-C. Peng, and W.-C. Lee, “Clustering and aggregating clues of trajectories for mining trajectory patterns and routes,”The VLDB Journal, vol. 24, pp. 169–192, 2015

  4. [4]

    A survey of trajectory distance measures and performance evaluation,

    H. Su, S. Liu, B. Zheng, X. Zhou, and K. Zheng, “A survey of trajectory distance measures and performance evaluation,”The VLDB Journal, vol. 29, pp. 3–32, 2020

  5. [5]

    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,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

  6. [6]

    Spatial and temporal characterization of travel patterns in a traffic network using vehicle trajectories,

    J. Kim and H. S. Mahmassani, “Spatial and temporal characterization of travel patterns in a traffic network using vehicle trajectories,”Trans- portation Research Procedia, vol. 9, pp. 164–184, 2015

  7. [7]

    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,” inProceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, 2012, pp. 186–194

  8. [8]

    Epimob: Interactive visual analytics of citywide human mobility re- strictions for epidemic control,

    C. Yang, Z. Zhang, Z. Fan, R. Jiang, Q. Chen, X. Song, and R. Shibasaki, “Epimob: Interactive visual analytics of citywide human mobility re- strictions for epidemic control,”IEEE Transactions on Visualization and Computer Graphics, vol. 29, no. 8, pp. 3586–3601, 2022

Show all 49 references
  1. [9]

    Trajectory clustering via deep representation learning,

    D. Yao, C. Zhang, Z. Zhu, J. Huang, and J. Bi, “Trajectory clustering via deep representation learning,” in2017 international joint conference on neural networks (IJCNN). IEEE, 2017, pp. 3880–3887

  2. [10]

    Learning deep representation for trajectory clustering,

    D. Yao, C. Zhang, Z. Zhu, Q. Hu, Z. Wang, J. Huang, and J. Bi, “Learning deep representation for trajectory clustering,”Expert Systems, vol. 35, no. 2, p. e12252, 2018

  3. [11]

    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 Computing Surveys (CSUR), vol. 54, no. 2, pp. 1–36, 2021

  4. [12]

    Tmn: trajectory matching networks for predicting similarity,

    P. Yang, H. Wang, D. Lian, Y . Zhang, L. Qin, and W. Zhang, “Tmn: trajectory matching networks for predicting similarity,” in2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022, pp. 1700–1713

  5. [13]

    Trajectory outlier detection: Algorithms, taxonomies, evaluation, and open challenges,

    A. Belhadi, Y . Djenouri, J. C.-W. Lin, and A. Cano, “Trajectory outlier detection: Algorithms, taxonomies, evaluation, and open challenges,” ACM Transactions on Management Information Systems (TMIS), vol. 11, no. 3, pp. 1–29, 2020

  6. [14]

    An overview on trajectory outlier detection,

    F. Meng, G. Yuan, S. Lv, Z. Wang, and S. Xia, “An overview on trajectory outlier detection,”Artificial Intelligence Review, vol. 52, pp. 2437–2456, 2019

  7. [15]

    Contrastive trajectory similarity learning with dual-feature attention,

    Y . Chang, J. Qi, Y . Liang, and E. Tanin, “Contrastive trajectory similarity learning with dual-feature attention,” in2023 IEEE 39th International conference on data engineering (ICDE). IEEE, 2023, pp. 2933–2945

  8. [17]

    Computing discrete fr ´echet distance,

    T. Eiter and H. Mannila, “Computing discrete fr ´echet distance,” 1994

  9. [18]

    A comparative analysis of trajectory similarity measures,

    Y . Tao, A. Both, R. I. Silveira, K. Buchin, S. Sijben, R. S. Purves, P. Laube, D. Peng, K. Toohey, and M. Duckham, “A comparative analysis of trajectory similarity measures,”GIScience & Remote Sensing, vol. 58, no. 5, pp. 643–669, 2021

  10. [19]

    A survey of traffic prediction: from spatio-temporal data to intelligent transportation,

    H. Yuan and G. Li, “A survey of traffic prediction: from spatio-temporal data to intelligent transportation,”Data Science and Engineering, vol. 6, no. 1, pp. 63–85, 2021

  11. [20]

    Kgts: contrastive trajectory similarity learning over prompt knowledge graph embedding,

    Z. Chen, D. Zhang, S. Feng, K. Chen, L. Chen, P. Han, and S. Shang, “Kgts: contrastive trajectory similarity learning over prompt knowledge graph embedding,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8311–8319

  12. [21]

    Trajgat: A graph- based long-term dependency modeling approach for trajectory similarity computation,

    D. Yao, H. Hu, L. Du, G. Cong, S. Han, and J. Bi, “Trajgat: A graph- based long-term dependency modeling approach for trajectory similarity computation,” inProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 2275–2285

  13. [22]

    node2vec: Scalable feature learning for networks,

    A. Grover and J. Leskovec, “node2vec: Scalable feature learning for networks,” inProceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, 2016, pp. 855– 864

  14. [23]

    An overview of quadtrees, octrees, and related hierarchical data structures,

    H. Samet, “An overview of quadtrees, octrees, and related hierarchical data structures,”Theoretical Foundations of Computer Graphics and CAD, pp. 51–68, 1988

  15. [24]

    Rotate: Knowledge graph embedding by relational rotation in complex space,

    Z. Sun, Z.-H. Deng, J.-Y . Nie, and J. Tang, “Rotate: Knowledge graph embedding by relational rotation in complex space,”arXiv preprint arXiv:1902.10197, 2019

  16. [25]

    Grlstm: trajectory similarity computation with graph-based residual lstm,

    S. Zhou, J. Li, H. Wang, S. Shang, and P. Han, “Grlstm: trajectory similarity computation with graph-based residual lstm,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 4, 2023, pp. 4972–4980

  17. [26]

    Knowledge graph embedding by translating on hyperplanes,

    Z. Wang, J. Zhang, J. Feng, and Z. Chen, “Knowledge graph embedding by translating on hyperplanes,” inProceedings of the AAAI conference on artificial intelligence, vol. 28, no. 1, 2014

  18. [27]

    T3s: Effective representation learning for trajectory similarity computation,

    P. Yang, H. Wang, Y . Zhang, L. Qin, W. Zhang, and X. Lin, “T3s: Effective representation learning for trajectory similarity computation,” in2021 IEEE 37th International Conference on Data Engineering (ICDE). IEEE, 2021, pp. 2183–2188

  19. [28]

    Long short-term memory,

    S. Hochreiter, “Long short-term memory,”Neural Computation MIT- Press, 1997

  20. [29]

    Learning repre- sentations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- sentations by back-propagating errors,”nature, vol. 323, no. 6088, pp. 533–536, 1986

  21. [30]

    Trajectory similarity learning with auxiliary supervision and optimal matching,

    H. Zhang, X. Zhang, Q. Jiang, B. Zheng, Z. Sun, W. Sun, and C. Wang, “Trajectory similarity learning with auxiliary supervision and optimal matching,” 2020

  22. [31]

    Hypergraph hash learning for efficient trajectory similarity computation,

    Y . Cao, L. Li, X. Chen, X. Xu, Z. Huang, and Y . Yu, “Hypergraph hash learning for efficient trajectory similarity computation,” inProceedings of the 33rd ACM International Conference on Information and Knowl- edge Management, 2024, pp. 175–186

  23. [32]

    Hypergraph neural net- works,

    Y . Feng, H. You, Z. Zhang, R. Ji, and Y . Gao, “Hypergraph neural net- works,” inProceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01, 2019, pp. 3558–3565

  24. [33]

    Bootstrap your own latent-a new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azaret al., “Bootstrap your own latent-a new approach to self-supervised learning,” Advances in neural information processing systems, vol. 33, pp. 21...

  25. [34]

    Pixel recurrent neural networks,

    A. Van Den Oord, N. Kalchbrenner, and K. Kavukcuoglu, “Pixel recurrent neural networks,” inInternational conference on machine learning. PMLR, 2016, pp. 1747–1756

  26. [35]

    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,” in2018 IEEE 34th international conference on data engineering (ICDE). IEEE, 2018, pp. 617–628

  27. [36]

    Sequence to sequence learning with neural networks,

    I. Sutskever, “Sequence to sequence learning with neural networks,” arXiv preprint arXiv:1409.3215, 2014

  28. [37]

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

    K. Cho, B. Van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoder-decoder for statistical machine translation,”arXiv preprint arXiv:1406.1078, 2014

  29. [38]

    Efficient trajec- tory similarity computation with contrastive learning,

    L. Deng, Y . Zhao, Z. Fu, H. Sun, S. Liu, and K. Zheng, “Efficient trajec- tory similarity computation with contrastive learning,” inProceedings of the 31st ACM International Conference on Information & Knowledge Management, 2022, pp. 365–374

  30. [39]

    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,” inInternational conference on machine learning. PMLR, 2020, pp. 1597–1607

  31. [40]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

  32. [41]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

  33. [42]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in9th International Conference on Learning Representations, ICLR 2021

  34. [43]

    J. L. Doob and J. Doob,Classical potential theory and its probabilistic counterpart. Springer, 1984, vol. 262

  35. [44]

    L. C. G. Rogers and D. Williams,Diffusions, Markov processes, and martingales: It ˆo calculus. Cambridge university press, 2000, vol. 2

  36. [45]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,”Advances in neural infor- mation processing systems, vol. 35, pp. 26 565–26 577, 2022

  37. [46]

    Denoising diffusion bridge models,

    L. Zhou, A. Lou, S. Khanna, and S. Ermon, “Denoising diffusion bridge models,”arXiv preprint arXiv:2309.16948, 2023

  38. [47]

    Learning to rank: from pairwise approach to listwise approach,

    Z. Cao, T. Qin, T.-Y . Liu, M.-F. Tsai, and H. Li, “Learning to rank: from pairwise approach to listwise approach,” inProceedings of the 24th international conference on Machine learning, 2007, pp. 129–136

  39. [48]

    Geolife: A collaborative social networking service among user, location and trajectory

    Y . Zheng, X. Xie, W.-Y . Maet al., “Geolife: A collaborative social networking service among user, location and trajectory.”IEEE Data Eng. Bull., vol. 33, no. 2, pp. 32–39, 2010

  40. [49]

    T- drive: driving directions based on taxi trajectories,

    J. Yuan, Y . Zheng, C. Zhang, W. Xie, X. Xie, G. Sun, and Y . Huang, “T- drive: driving directions based on taxi trajectories,” inProceedings of the 18th SIGSPATIAL International conference on advances in geographic information systems, 2010, pp. 99–108

  41. [50]

    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,” in2019 IEEE 35th international conference on data engineering (ICDE). IEEE, 2019, pp. 1358–1369

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.