Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion

Baoshen Guo; Jinhua Zhao; Junyi Li; Shenhao Wang; Zhiqing Hong

REVIEW 4 major objections 6 minor 79 references

Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Coarse-to-fine cascaded diffusion — a coarse road-segment latent stage followed by a fine-grained GPS stage — generates synthetic urban trajectories that match real mobility distributions more closely than VAE, GAN, and single-stage…

desk verdict A genuinely new cascade of latent-diffusion and U-Net for trajectory synthesis, whose reported wins over SOTA are plausible but not yet established because the closest baselines are missing and no variance is reported. read the letter →

arxiv 2507.13366 v1 pith:HMTXFAAN submitted 2025-07-08 cs.SI cs.CV

Baoshen Guo , Zhiqing Hong , Junyi Li , Shenhao Wang , Jinhua Zhao This is my paper · ORCID

classification cs.SIcs.CV

keywords trajectorysynthesisdiffusionmodelscascadedlatenturbanmobilitygenerationroadnetworkconstraintsGPSprivacy-preserving

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

The paper is trying to establish that the spatial hierarchy of urban trips can be exploited for generation: instead of synthesizing a fine GPS trajectory in one shot, a model should first generate the coarse sequence of road segments a trip follows, then generate the detailed GPS points conditioned on that segment sequence. If this two-stage (cascaded) design works, it matters because fine-grained mobility data is expensive and privacy-sensitive, and a generator that reproduces real aggregate movement patterns could stand in for withheld data in smart-city applications. The paper's evidence is a battery of distribution-similarity measurements — divergences of spatial point distributions, trip-length distributions, and origin–destination patterns — computed on three large city datasets, on which the cascaded model reports the lowest values among the methods compared. It further claims the two-level design creates a flexible privacy–utility switch, because differential-privacy noise can be injected into either stage alone, and reports that protecting only the coarse stage costs little utility.

What carries the argument

The carrying object is the cascaded hybrid diffusion pipeline. An autoregressive trajectory autoencoder — a BART-style transformer encoder with a Perceiver-inspired cross-attention compression module and a GPT-style autoregressive decoder — maps variable-length road-segment sequences onto fixed-length continuous latent vectors, so the discrete space is never diffused directly. The segment-level latent denoising network is built from Diffusion Transformer (DiT) blocks with adaptive layer-normalization conditioning on time and trip attributes, and is trained with an adaptive physics-informed validity loss that penalizes consecutive segments that are not adjacent in the road network, applied only in the later denoising stages where structure is discernible. The fine-grained GPS-level network is a U-Net with spatial-transformer cross-attention layers in which the noisy GPS representation is the query and the segment-level latent provides keys and values. A noise-augmentation mechanism corrupts the clean segment-level latent at a random diffusion timestep during training but uses the clean latent at inference, teaching the second stage to follow imperfect coarse guidance. The forward noising processes of both stages run in parallel and the reverse denoising runs sequentially, coarse first and fine second.

What would settle it

Re-run the Table 3 protocol with multiple independent sampling rounds per method and report the spread of each divergence, while adding the two structure-aware diffusion baselines the paper cites but does not benchmark (refs [47] and [62]); if an omitted baseline matches or beats Cardiff on spatial or origin–destination divergence on any dataset, or if Cardiff's margin over its own ablations (for instance 0.0043 versus 0.0066 in Singapore spatial divergence) reverses across sampling rounds, the paper's superiority claim is falsified.

Watch

Extended reading notes

Core claim

The framework's central claim is that a cascaded hybrid diffusion pipeline — a first stage that denoises a compressed latent of the discrete road-segment sequence, followed by a second stage that denoises the continuous GPS sequence conditioned on that segment-level latent — produces higher-fidelity synthetic trajectories than any single-stage baseline it is compared with. Concretely, on the Singapore dataset the framework reports a spatial-distribution divergence of 0.0043, a length-distribution divergence of 0.0089, and an origin–destination divergence of 0.0164, and it reports the lowest spatial and origin–destination divergences on Chengdu and Porto as well. The paper attributes the gains to three design choices: operating the coarse diffusion in a low-dimensional latent space of segments, a physics-informed validity loss applied only in the later denoising steps, and a noise-augmentation conditioning mechanism that trains the fine stage against corrupted versions of the coarse condition so that inference is robust. The authors also claim the cascade removes the need for per-trajectory road-segment supervision at generation time and enables a stage-wise privacy–utility trade-off.

Load-bearing premise

The claim rests on the assumption that the distribution-similarity scores in Table 3 — each measured once on 2,000 sampled trajectories with no variance reported, against a baseline list that omits the structure-aware diffusion methods the paper cites — are a fair and stable basis for concluding the cascade is better, so if a single run is unrepresentative or an omitted baseline scores lower, the superiority claim is not established.

Editorial extensions

If this is right

The coarse stage generates its own segment-level conditioning from noise, so fine-grained trajectories can be synthesized without per-trajectory road-segment supervision at inference time, removing the main scalability obstacle of the ControlTraj-style approach.
Because differential privacy can be applied to either stage alone, a deployment could protect fine-grained locations such as home and workplace by privatizing the GPS stage while keeping the coarse stage intact, with the reported cost on Singapore being a modest rise from 0.0043 to 0.0052 in spatial divergence.
Sampling with an interval of 50 (20 denoising steps instead of 1000) cuts generation time by nearly sixfold with minimal quality loss on Singapore, which is what makes the two-stage pipeline practical at city scale.
Training a next-location predictor on the synthetic trajectories alone reaches accuracy comparable to training on real data (0.8882 versus 0.9101 on Singapore), indicating that the generated data carries usable signal for downstream mobility tasks.
Because the segment-level diffusion operates on length-32 latent vectors rather than raw segment sequences, the coarse-to-fine scheme stays computationally tractable on datasets of hundreds of thousands to millions of trips.

Reading between the lines

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

A testable consequence the paper leaves implicit is that the same cascade should transfer to any sequence-generation task with a discrete high-level structure constraining a continuous low-level output, such as route-then-waypoint or block-then-address generation, and the quality gain over single-stage models should grow with the strength of the topological constraint.
The noise-augmentation conditioning is effectively training-time augmentation of the error the first stage will make; an extension would be to measure how gracefully the fine stage degrades as the coarse stage's error is artificially inflated, which would directly predict the cascade's robustness on unseen cities.
The privacy results measure non-replication via minimum Hausdorff distance rather than a formal differential-privacy guarantee for the whole cascade, so a formal DP accounting and a membership-inference evaluation would be needed before the privacy claims could support a public data release.
Behind the divergence comparisons there is an entropy-partition argument the paper only gestures at: if most of a trip's distributional information lives in which segments it visits, a two-stage model that pins topology first should beat a one-stage model whenever the road network is dense and the GPS detail within a segment is relatively predictable, a hypothesis testable by varying road-network

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.

Desk Editor's Note

read the letter

Cardiff is a real, workable new architecture, not a repackaging: a BART-style autoencoder compresses road-segment sequences into a compact latent, a DiT-based diffusion generates segment-level latents, and a U-Net with cross-attention conditions GPS-level generation on those latents with a noise-augmentation trick. That combination is absent from DiffTraj and ControlTraj, and the coarse-to-fine decomposition is well motivated by the hierarchical structure of mobility data. The conditional generation demos (Fig. 13) and the downstream next-segment prediction table are good additions.

The load-bearing empirical claim rests on Table 3, which reports single JSD point estimates with no variance, no seeds, and no significance testing. Differences of 0.001 in JSD-trip between Cardiff and its ablations are almost certainly inside the sampling noise of 2,000 trajectories over a 30×30 grid. More importantly, the two closest structure-aware diffusion baselines, Diff-RNTraj and SEED, are cited but not compared. That makes 'outperforms state-of-the-art' a stretch; at best it is 'outperforms the four baselines we selected.' Singapore is proprietary Veraset data, so the most favorable results cannot be independently regenerated. The privacy section reports DP-SGD training options but no epsilon, so the word 'guarantee' is not supported. The paper should report a privacy budget or soften the language. There is also a small train/sample mismatch: Sec 3.4.2 trains the GPS denoiser on noise-augmented z_hat, while Algorithm 2 samples with clean z0. That might be okay—the model sees a range of noise levels and inference uses the zero-noise case—but it needs a sentence of justification.

None of this is fatal to the architecture; the core idea is sound and the paper is honest about some limitations. But the empirical case needs hardening before the headline claim can be taken at face value. I would send this to peer review with an emphasis on re-running the comparison with variance, adding SEED and Diff-RNTraj, and fixing the privacy accounting. A serious referee could turn this into a solid paper.

Referee Report

4 major / 6 minor

Summary. The paper proposes Cardiff, a two-stage cascaded diffusion framework for fine-grained trajectory synthesis. The first stage encodes discrete road-segment sequences into a compact latent space with an autoregressive transformer autoencoder and performs latent diffusion with a DiT-based denoiser plus an adaptive physics-informed adjacency loss. The second stage trains a conditional U-Net that generates continuous GPS trajectories, conditioned on the first-stage segment latent through noise-augmented cross-attention and on trajectory-level attributes. Experiments on Singapore, Chengdu, and Porto compare Cardiff with VAE, GAN, DiffTraj, and ControlTraj baselines using JSD-SD, JSD-LD, and JSD-trip, along with ablations, conditional generation visualizations, downstream utility, and a DP-SGD privacy-utility analysis. The central claim is that Cardiff outperforms state-of-the-art baselines in various realism metrics while enabling a tunable privacy-utility trade-off.

Significance. If the empirical claims hold, Cardiff is a useful architectural contribution: the two-level coarse-to-fine decomposition is well motivated by the spatial hierarchy of trajectories, the latent compression and physical validity loss are reasonable design choices, the code is released, and the reported JSD gains over VAE/GAN/single-stage diffusion are substantial (e.g., Singapore JSD-SD 0.0043 vs. 0.0285 for DiffTraj). The stage-selective DP-SGD analysis is a nice demonstration of how cascaded design can localize privacy intervention. However, the headline 'outperforms state-of-the-art' claim is not fully established by the current evaluation: the closest structure-aware diffusion baselines are absent, the main table reports a single draw of 2,000 generated trajectories with no variance or significance information, the strongest result uses proprietary Veraset data, and the privacy 'guarantee' is stated without a formal privacy budget.

major comments (4)

[Sec. 4.1.3, Table 3] The central empirical claim rests on single point estimates of JSD-SD, JSD-LD, and JSD-trip computed from 2,000 sampled trajectories per method, with no number of seeds, no variance, no confidence intervals, and no significance tests. Several decisive margins are very small, for example Chengdu JSD-trip 0.0320 (Cardiff) vs. 0.0326 (Cardiff w/o aug) and Porto JSD-LD 0.0155 (Cardiff) vs. 0.0214 (Cardiff w/o p), and a 30x30 grid histogram over 2,000 trips has non-negligible sampling noise. Please report mean and standard deviation over at least 3-5 independent seeds, and where possible bootstrap confidence intervals or paired significance tests, for Cardiff and all baselines.
[Sec. 4.1.4, Sec. 5.1] The baseline set omits the two closest structure-aware diffusion methods that the paper itself cites: Diff-RNTraj [62] and SEED [47]. Diff-RNTraj is a road-network-constrained diffusion generator and SEED bridges sequence and diffusion models for road trajectory generation, so both are arguably stronger competitors than the included DiffTraj and ControlTraj. Without them, the 'state-of-the-art' label used in the abstract and Sec. 4.2.1 is not established. In addition, the best result is reported on proprietary Veraset data, so the headline numbers cannot be independently regenerated; please include the omitted baselines on at least the two public datasets, or provide a public processed version of the Singapore data.
[Sec. 4.2.7, Sec. 6] The paper repeatedly states that Cardiff provides 'privacy guarantees' and a 'tunable balance between privacy preservation and utility,' but the evaluation reports only the CDF of minimum Hausdorff distance and utility numbers, with no privacy budget epsilon, no DP-SGD clipping or noise parameters, and no formal differential privacy accounting. The Limitations section itself concedes that the current approach does not fully exploit diffusion noise for coordination with DP mechanisms. Please either remove the word 'guarantee' and describe the results as empirical privacy-leakage measurements, or provide the missing DP accounting and epsilon values.
[Sec. 3.4.2, Sec. 3.5.2, Algorithm 2] There is a train/sampling mismatch in the noise-augmentation mechanism: Eq. (17) trains the GPS-level denoiser on noise-augmented latents z_hat with a sampled augmentation step t_r, but Algorithm 2 and Sec. 3.5.2 sample with the clean latent z0 and do not pass any t_r to the denoiser. As a result, the second-stage network is evaluated on an input distribution it did not see during training, and the Cardiff vs. Cardiff w/o aug ablation does not cleanly isolate the proposed mechanism. Please justify the omission, for example as a deliberate zero-noise inference rule, or apply the same augmentation schedule at inference.

minor comments (6)

[Algorithm 1, line 13] The pseudocode computes L_r as the noise-matching term only, while Eq. (15) defines L_r = L_segment + lambda_p * L_phy; please align the algorithm with the loss definition.
[Sec. 4.2.3] Fig. 6(b) states that the model generates 1,024 trajectories for the sampling-interval study, whereas Sec. 4.1.3 specifies that 2,000 synthetic trajectories are sampled for the main metrics; please clarify which sample size is used in the sensitivity analysis.
[Sec. 4.2.7] The enumerated training settings jump from '(ii) applying DP training only during the first stage' to '(iv) applying DP training only during the second stage'; the list skips (iii).
[Sec. 4.2.2] The text attributes the Chengdu JSD-LD degradation to 'physical constraints introduced in the second-stage fine-grained generation,' but L_phy is defined and applied in the first-stage segment-level loss (Eq. 14-15); please correct the attribution or explain the intended mechanism.
[Sec. 3.1.2, Eq. (13)] There are minor notation issues: 'x_o' should be 'x_0', and Eq. (13) refers to z_t^r while the text defines z_t as the noised latent; please unify the notation.
[Sec. 4.2.1] The text says the model generates 'the same number of synthetic trajectories as the synthetic test dataset,' which is inconsistent with the fixed 2,000-sample protocol in Sec. 4.1.3; please specify the exact sample count used.

Circularity Check

1 steps flagged · score 6.0 of 10

Main evaluation metrics overlap with the conditioning input: JSD-LD and JSD-trip are partly forced by c, making part of the reported advantage constructional rather than independent.

fitted input called prediction [Sec. 3.3 (Eq. 11 context), Sec. 3.4.1, Sec. 4.1.3, Algorithm 2]
"c is the external conditional information, which consists of the departure time, origin, destination, and some statistics (e.g., lengths, duration) of trajectories. ... JSD-LD: which means the Jensen-Shannon Divergence of lengths between real trajectory and synthetic trajectory data. ... JSD-trip: which means the Jensen-Shannon Divergence of the origin and destination grid distribution of the trajectory between real trajectory data and synthetic trajectory data."

The metric JSD-LD is literally the length distribution of generated vs. real trajectories, and JSD-trip is the origin/destination grid distribution. Both quantities are members of the conditioning vector c that Cardiff consumes at every denoising step: Sec. 3.3 lists lengths and duration in c, Sec. 3.4.1 lists start/end locations in c, and Algorithm 2 takes c as an input. Thus, if the 2000 evaluation trajectories in Table 3 are generated with c drawn from the real/test set, the model is handed the exact marginals it is then credited with reproducing; the 'prediction' of length and O/D distributions reduces to reading off the input conditions. The paper never states how c was sampled for Table 3, and the conditional generation experiment (Fig.

full rationale

The paper's evaluation metrics JSD-LD and JSD-trip are exactly the length and origin/destination distributions that the model receives as conditioning inputs c. Since generation is conditional on c (Algorithm 2), if the test conditions are sampled from the real test set, the generated length and O/D marginals will match the real ones by construction to the extent the model follows the condition. The paper provides no evidence that evaluation conditions were sampled independently of the real data distribution, and its own conditional-generation experiment (Fig. 13) demonstrates that length and OD are controllable inputs. Thus two of the three headline metrics reduce to conditional-fidelity checks rather than independent predictions of the data distribution. The JSD-SD metric, which measures the emergent spatial distribution of generated points, is not directly conditioned and provides independent evidence for the method; therefore the circularity is partial, not total. A score of 6 reflects this partial reduction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical or mathematical entities; its novel components are architectural. The main free parameters are hyperparameters selected by sensitivity experiments on the evaluation datasets. The key domain assumptions are the hierarchy decomposition, the quality of the learned latent, and the accuracy of the road network graph.

free parameters (3)

lambda_p (physical validity loss weight) = 0.05 (Singapore, Chengdu), 0.01 (Porto)
Chosen via sensitivity scan in Fig 6(a); directly controls how strongly the segment-level diffusion is penalized for non-adjacent road transitions and affects the final JSD-LD.
Sampling interval (DDIM skip) = 50
Selected from a range of interval values as a speed-quality tradeoff in Fig 6(b); the paper's reported results use this default.
Latent length and dimension = L=32, d=128
Fixed model hyperparameters that determine the capacity of the segment-level latent; no ablation is shown for these values.

assumptions (5)

standard math Standard DDPM forward and reverse process equations (Eqs. 1-6) hold as described.
Background diffusion formulation adopted from Ho et al. [20] without modification.
domain assumption Trajectories can be decomposed into road-segment sequences and fine-grained GPS sequences without losing task-relevant information.
Central modeling premise in Definition 2.4 and Section 3.1, motivated by [3,64]; if the hierarchy is not a sufficient representation, the coarse stage cannot guide the fine stage.
domain assumption The learned latent autoencoder maps variable-length discrete segment sequences into fixed-size continuous latents suitable for diffusion modeling.
Relies on the BART/perceiver-style reconstruction objective in Eq. 10; without near-lossless compression, the first-stage diffusion inherits reconstruction errors.
domain assumption The road network adjacency matrix A correctly encodes physical connectivity of map-matched segments.
The spatial validity loss in Eq. 14 assumes the OpenStreetMap-derived graph and map-matching are accurate.
ad hoc to paper Applying noise augmentation only during training while using clean z0 during sampling remains effective.
This is a design choice without theoretical justification; the paper only empirically compares with and without augmentation, and the train-inference mismatch is not analyzed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion." pith.science (2026). https://pith.science/paper/HMTXFAAN

@misc{pith2026250713366,
  author       = {Pith},
  title        = {Pith review of: Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HMTXFAAN}},
  note         = {Machine review of arXiv:2507.13366}
}

read the original abstract

Urban mobility data has significant connections with economic growth and plays an essential role in various smart-city applications. However, due to privacy concerns and substantial data collection costs, fine-grained human mobility trajectories are difficult to become publicly available on a large scale. A promising solution to address this issue is trajectory synthesizing. However, existing works often ignore the inherent structural complexity of trajectories, unable to handle complicated high-dimensional distributions and generate realistic fine-grained trajectories. In this paper, we propose Cardiff, a coarse-to-fine Cascaded hybrid diffusion-based trajectory synthesizing framework for fine-grained and privacy-preserving mobility generation. By leveraging the hierarchical nature of urban mobility, Cardiff decomposes the generation process into two distinct levels, i.e., discrete road segment-level and continuous fine-grained GPS-level: (i) In the segment-level, to reduce computational costs and redundancy in raw trajectories, we first encode the discrete road segments into low-dimensional latent embeddings and design a diffusion transformer-based latent denoising network for segment-level trajectory synthesis. (ii) Taking the first stage of generation as conditions, we then design a fine-grained GPS-level conditional denoising network with a noise augmentation mechanism to achieve robust and high-fidelity generation. Additionally, the Cardiff framework not only progressively generates high-fidelity trajectories through cascaded denoising but also flexibly enables a tunable balance between privacy preservation and utility. Experimental results on three large real-world trajectory datasets demonstrate that our method outperforms state-of-the-art baselines in various metrics.

Figures

Figures reproduced from arXiv: 2507.13366 by the authors.

**Figure 1.** The spatial hierarchy of trajectories. Definition 2.4 (Trajectory Hierarchical Decomposition). Inspired by the analysis of the hierarchical organization of urban mobility [3, 64], as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** The framework of Cardiff. where z0 is the clean latent representation of the input segmentlevel trajectory. 𝛼¯𝑡 is the constant hyperparameter to control the noise injection degree for sampling. 𝜖𝑡𝑟 is a noise sampled from the Gaussian distribution. At the fine-grained GPS-level forward diffusion, we define the clean input 𝑥𝑜 as the fine-grained trajectory sample 𝜏. We can then obtain the 𝑥𝑡 at any given diffusion … view at source ↗

**Figure 3.** Illustration of trajectory autoencoder. Segment-level latent encoding and compression: As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Structure of segment-level denoising network. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: (a) Cascaded denoising structure (bottom left), (b) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Parameter sensitivity analysis on Singapore dataset: [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Results visualization of Singapore dataset: (a) real [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: Results visualization of Chengdu dataset: (a) real [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Results visualization of Porto dataset: (a) real trajec [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

**Figure 13.** Figure 13: Visualization of conditional generation. [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]

**Figure 15.** Figure 15: CDF of the minimum Hausdorff distance to real data. Model JSD-SD JSD-LD JSD-Trip 𝑤/𝑜 DP 0.0043 0.0089 0.0164 DP1 0.0052 0.0126 0.0184 DP2 0.0246 0.0447 0.0348 𝑤/ DP 0.0223 0.0399 0.0416 [PITH_FULL_IMAGE:figures/full_fig_p011_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 54 canonical work pages

[62]

Veraset. 2024. Global Mobility and Location Data Provider. https://www.veraset. com/. Accessed: 2024-06-17

work page 2024
[47]

Xuan Rao, Shuo Shang, Renhe Jiang, Peng Han, and Lisi Chen. [n.d.]. Seed: Bridging Sequence and Diffusion Models for Road Trajectory Generation. InTHE WEB CONFERENCE 2025

work page 2025
[1]

Open Street Map

2019. Open Street Map. https://www.openstreetmap.org. accessed 27 June 2024

work page 2019
[2]

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. 308–318

2016
[3]

Aleix Bassolas, Hugo Barbosa-Filho, Brian Dickinson, Xerxes Dotiwalla, Paul Eastham, Riccardo Gallotti, Gourab Ghoshal, Bryant Gipson, Surendra A Hazarie, Henry Kautz, et al. 2019. Hierarchical organization of urban mobility and its connection with city livability. Nature communications 10, 1 (2019), 4817

work page 2019
[4]

Kunlin Cai, Jinghuai Zhang, Zhiqing Hong, William Shand, Guang Wang, Desh- eng Zhang, Jianfeng Chi, and Yuan Tian. 2024. Where Have You Been? A Study of Privacy Risk for Point-of-Interest Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 175–186

work page 2024
[5]

Chu Cao and Mo Li. 2021. Generating mobility trajectories with retained data utility. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2610–2620

work page 2021
[6]

Xinyu Chen, Jiajie Xu, Rui Zhou, Wei Chen, Junhua Fang, and Chengfei Liu. 2021. Trajvae: A variational autoencoder model for trajectory generation. Neurocom- puting 428 (2021), 332–339

work page 2021

Show all 79 references

[7]

Yile Chen, Gao Cong, and Cuauhtemoc Anda. 2023. Teri: An effective framework for trajectory recovery with irregular time intervals. Proceedings of the VLDB Endowment 17, 3 (2023), 414–426

2023
[8]

Yuhang Chen, Chaoyun Zhang, Minghua Ma, Yudong Liu, Ruomeng Ding, Bowen Li, Shilin He, Saravan Rajmohan, Qingwei Lin, and Dongmei Zhang. 2023. ImD- iffusion: Imputed Diffusion Models for Multivariate Time Series Anomaly Detec- tion. Proceedings of the VLDB Endowment 17, 3 (202...

2023
[9]

Yves-Alexandre De Montjoye, César A Hidalgo, Michel Verleysen, and Vincent D Blondel. 2013. Unique in the crowd: The privacy bounds of human mobility. Scientific reports 3, 1 (2013), 1376

2013
[10]

Bangchao Deng, Xin Jing, Tianyue Yang, Bingqing Qu, Philippe Cudre-Mauroux, and Dingqi Yang. 2024. Revisiting Synthetic Human Trajectories: Imitative Generation and Benchmarks Beyond Datasaurus. arXiv preprint arXiv:2409.13790 (2024)

2024 arXiv
[11]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

2021
[12]

Didi Chuxing. 2017. GAIA Initiative. https://outreach.didichuxing.com. Accessed: 2019-10-01

2017
[13]

Yuntao Du, Yujia Hu, Zhikun Zhang, Ziquan Fang, Lu Chen, Baihua Zheng, and Yunjun Gao. 2023. Ldptrace: Locally differentially private trajectory synthesis. arXiv preprint arXiv:2302.06180 (2023)

2023 arXiv
[14]

Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. 2022. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933 (2022)

2022 arXiv
[15]

Zengyang Gong, Yuxiang Zeng, and Lei Chen. 2024. Real-time insertion operator for shared mobility on time-dependent road networks. Proceedings of the VLDB Endowment 17, 7 (2024), 1669–1682

2024
[16]

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems , Vol. 27

2014
[17]

Klaus Greff, Rupesh K Srivastava, Jan Koutník, Bas R Steunebrink, and Jürgen Schmidhuber. 2016. LSTM: A search space odyssey. IEEE transactions on neural networks and learning systems 28, 10 (2016), 2222–2232

2016
[18]

Mehmet Emre Gursoy, Ling Liu, Stacey Truex, Lei Yu, and Wenqi Wei. 2018. Utility-aware synthesis of differentially private and attack-resilient location traces. In Proceedings of the 2018 ACM SIGSAC conference on computer and com- munications security. 196–211

2018
[19]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

2016
[20]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

2020
[21]

Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. 2022. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research 23, 47 (2022), 1–33

2022
[22]

Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

2022 arXiv
[23]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models. Advances in Neu- ral Information Processing Systems 35 (2022), 8633–8646

2022
[24]

Yujia Hu, Yuntao Du, Zhikun Zhang, Ziquan Fang, Lu Chen, Kai Zheng, and Yunjun Gao. 2024. Real-time trajectory synthesis with local differential privacy. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 1685–1698

2024
[25]

Dou Huang, Xuan Song, Zipei Fan, Renhe Jiang, Ryosuke Shibasaki, Yu Zhang, Haizhong Wang, and Yugo Kato. 2019. A variational autoencoder based gen- erative model of urban human mobility. In 2019 IEEE conference on multimedia information processing and retrieval (MIPR) . IEEE, 425–430

2019
[26]

Daniel P Huttenlocher, Gregory A Klanderman, and William J Rucklidge. 1993. Comparing images using the Hausdorff distance. IEEE Transactions on Pattern Analysis and Machine Intelligence 15, 9 (1993), 850–863. https://doi.org/10.1109/ 34.232073

1993
[27]

Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shel- hamer, et al. 2021. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795 (2021)

2021 arXiv
[28]

Wenjun Jiang, Wayne Xin Zhao, Jingyuan Wang, and Jiawei Jiang. 2023. Contin- uous trajectory generation based on two-stage GAN. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 4374–4382

2023
[29]

Alexandra Kapp, Julia Hansmeyer, and Helena Mihaljević. 2023. Generative models for synthetic urban mobility data: A systematic literature review.Comput. Surveys 56, 4 (2023), 1–37

2023
[30]

Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. 2022. Elucidating the Design Space of Diffusion-Based Generative Models. In Proc. NeurIPS

2022
[31]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[32]

Diederik P Kingma, Max Welling, et al. 2013. Auto-encoding variational bayes

2013
[33]

Solomon Kullback and Richard A Leibler. 1951. On information and sufficiency. The annals of mathematical statistics 22, 1 (1951), 79–86

1951
[34]

Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mo- hamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.1...

2019 arXiv
[35]

Maocheng Li, Yuxiang Zeng, Libin Zheng, Lei Chen, and Qing Li. 2023. Accurate and efficient trajectory-based contact tracing with secure computation and geo- indistinguishability. InInternational Conference on Database Systems for Advanced Applications. Springer, 300–316

2023
[36]

J. Lin. 1991. Divergence measures based on the Shannon entropy. IEEE Transac- tions on Information Theory 37, 1 (1991), 145–151

1991
[37]

Xi Liu, Hanzhou Chen, and Clio Andris. 2018. trajGANs: Using generative adversarial networks for geo-privacy protection of trajectory data (Vision paper). In Location privacy and security workshop . 1–7

2018
[38]

Qingyue Long, Huandong Wang, Tong Li, Lisi Huang, Kun Wang, Qiong Wu, Guangyu Li, Yanping Liang, Li Yu, and Yong Li. 2023. Practical synthetic human trajectories generation based on variational point processes. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Disc...

2023
[39]

Justin Lovelace, Varsha Kishore, Chao Wan, Eliot Shekhtman, and Kilian Q Weinberger. 2023. Latent diffusion for language generation. Advances in Neural Information Processing Systems 36 (2023), 56998–57025

2023
[40]

Jesse Merhi, Erik Buchholz, and Salil S Kanhere. 2024. Synthetic Trajectory Gener- ation Through Convolutional Neural Networks. In2024 21st Annual International Conference on Privacy, Security and Trust (PST) . IEEE, 1–12

2024
[41]

moreiraMatias and Wendy Kan. 2015. ECML/PKDD 15: Taxi Trip Time Predic- tion (II). https://kaggle.com/competitions/pkdd-15-taxi-trip-time-prediction-ii. Kaggle

2015
[42]

Mehmet Ercan Nergiz, Maurizio Atzori, and Yucel Saygin. 2008. Towards tra- jectory anonymization: a generalization-based approach. In Proceedings of the SIGSPATIAL ACM GIS 2008 International Workshop on Security and Privacy in GIS and LBS. 52–61

2008
[43]

Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffusion probabilistic models. In International conference on machine learning . PMLR, 8162–8171

2021
[44]

William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. In Proceedings of the IEEE/CVF international conference on computer vision . 4195–4205

2023
[45]

Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen
[46]

Jinmeng Rao, Song Gao, Yuhao Kang, and Qunying Huang. 2020. LSTM-TrajGAN: A deep learning approach to trajectory privacy protection. arXiv preprint arXiv:2006.10521 (2020)

2020 arXiv
[48]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

2022
[49]

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Baoshen Guo, Zhiqing Hong, Junyi Li, Shenhao W...

2015
[50]

Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural infor...

2022
[51]

Robin M Schmidt. 2019. Recurrent neural networks (rnns): A gentle introduction and overview. arXiv preprint arXiv:1912.05911 (2019)

2019 arXiv
[52]

Chenyang Shao, Fengli Xu, Bingbing Fan, Jingtao Ding, Yuan Yuan, Meng Wang, and Yong Li. 2024. Beyond imitation: Generating human mobility from context- aware reasoning with large language models. arXiv e-prints (2024), arXiv–2402

2024
[53]

Dingyuan Shi, Yongxin Tong, Zimu Zhou, Ke Xu, Zheng Wang, and Jieping Ye
[54]

Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

2020 arXiv
[55]

Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradi- ents of the data distribution. Advances in neural information processing systems 32 (2019)

2019
[56]

Shun Takagi, Li Xiong, Fumiyuki Kato, Yang Cao, and Masatoshi Yoshikawa
[57]

Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. 2022. Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)

2022 arXiv
[58]

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

2017
[59]

Proceedings of the VLDB Endowment 17, 11 (2024), 3058–3071

HRNet: Differentially Private Hierarchical and Multi-Resolution Network for Human Mobility Data Synthesization. Proceedings of the VLDB Endowment 17, 11 (2024), 3058–3071

2024
[60]

Huandong Wang, Changzheng Gao, Yuchen Wu, Depeng Jin, Lina Yao, and Yong Li. 2023. PateGail: A privacy-preserving mobility trajectory generator with imitation learning. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 14539–14547

2023
[61]

Yi Wang, Wai Y Szeto, Ke Han, and Terry L Friesz. 2018. Dynamic traffic assign- ment: A review of the methodological advances for environmentally sustainable road transportation applications. Transportation Research Part B: Methodological 111 (2018), 370–394

2018
[63]

Haomin Wen, Youfang Lin, Yutong Xia, Huaiyu Wan, Qingsong Wen, Roger Zimmermann, and Yuxuan Liang. 2023. Diffstg: Probabilistic spatio-temporal graph forecasting with denoising diffusion models. In Proceedings of the 31st ACM International Conference on Advances in Geographic ...

2023
[64]

Wikipedia. 2025. Transportation forecasting — Four-step models. https://en. wikipedia.org/wiki/Transportation_forecasting/Four-step_models. [Online; accessed 19-March-2025]

2025
[65]

Tonglong Wei, Youfang Lin, Shengnan Guo, Yan Lin, Yiheng Huang, Chenyang Xi- ang, Yuqing Bai, and Huaiyu Wan. 2024. Diff-rntraj: A structure-aware diffusion model for road network-constrained trajectory generation. IEEE Transactions on Knowledge and Data Engineering (2024)

2024
[66]

Haitao Yuan, Gao Cong, and Guoliang Li. 2024. Nuhuo: An effective estimation model for traffic speed histogram imputation on a road network. Proceedings of the VLDB Endowment 17, 7 (2024), 1605–1617

2024
[67]

Yuan Yuan, Jingtao Ding, Huandong Wang, and Depeng Jin. 2024. Generating daily activities with need dynamics. ACM Transactions on Intelligent Systems and Technology 15, 2 (2024), 1–28

2024
[68]

Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. 2023. InterDiff: Generating 3D Human-Object Interactions with Physics-Informed Diffusion. In ICCV

2023
[69]

Jing Zhang, Qihan Huang, Yirui Huang, Qian Ding, and Pei-Wei Tsai. 2023. DP- TrajGAN: A privacy-aware trajectory generation model with differential privacy. Future Generation Computer Systems 142 (2023), 25–40

2023
[70]

Liming Zhang, Liang Zhao, and Dieter Pfoser. 2022. Factorized deep genera- tive models for end-to-end trajectory generation with spatiotemporal validity constraints. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems. 1–12

2022
[71]

Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. 2023. PhysDiff: Physics-Guided Human Motion Diffusion Model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

2023
[72]

Yizhe Zhang, Jiatao Gu, Zhuofeng Wu, Shuangfei Zhai, Joshua Susskind, and Navdeep Jaitly. 2023. Planner: Generating diversified paragraph via latent lan- guage diffusion model. Advances in Neural Information Processing Systems 36 (2023), 80178–80190

2023
[73]

Yuemin Zhang, Qingqing Ye, Rui Chen, Haibo Hu, and Qilong Han. 2023. Trajec- tory Data Collection with Local Differential Privacy. Proceedings of the VLDB Endowment 16, 10 (2023), 2591–2604

2023
[74]

Ruixing Zhang, Yunqi Liu, Liangzhe Han, Leilei Sun, Chuanren Liu, Jibin Wang, and Weifeng Lv. 2025. Large-scale Human Mobility Data Regeneration for Open Urban Research. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 . 2827–2836

2025
[75]

Yuanshao Zhu, Yongchao Ye, Shiyao Zhang, Xiangyu Zhao, and James Yu. 2023. Difftraj: Generating gps trajectory with diffusion probabilistic model. Advances in Neural Information Processing Systems 36 (2023), 65168–65188

2023
[76]

Yuanshao Zhu, James Jianqiao Yu, Xiangyu Zhao, Qidong Liu, Yongchao Ye, Wei Chen, Zijian Zhang, Xuetao Wei, and Yuxuan Liang. 2024. Controltraj: Controllable trajectory generation with topology-constrained diffusion model. In Proceedings of the 30th ACM SIGKDD Conference on Kn...

2024
[77]

Jensen, and Panos Kalnis

Silin Zhou, Shuo Shang, Lisi Chen, Christian S. Jensen, and Panos Kalnis. 2024. RED: Effective Trajectory Representation Learning with Comprehensive Infor- mation. Proc. VLDB Endow. 18, 2 (Oct. 2024), 80–92. https://doi.org/10.14778/ 3705829.3705830

2024
[2022]

arXiv preprint arXiv:2204.06125 1, 2 (2022), 3

Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1, 2 (2022), 3

2022 arXiv
[2024]

In The Twelfth International Conference on Learning Representations

Graph-Constrained Diffusion for End-to-end Path Planning. In The Twelfth International Conference on Learning Representations

Pith tools

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

[1] [62]

Veraset. 2024. Global Mobility and Location Data Provider. https://www.veraset. com/. Accessed: 2024-06-17

work page 2024

[2] [47]

Xuan Rao, Shuo Shang, Renhe Jiang, Peng Han, and Lisi Chen. [n.d.]. Seed: Bridging Sequence and Diffusion Models for Road Trajectory Generation. InTHE WEB CONFERENCE 2025

work page 2025

[3] [1]

Open Street Map

2019. Open Street Map. https://www.openstreetmap.org. accessed 27 June 2024

work page 2019

[4] [2]

Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. 308–318

2016

[5] [3]

Aleix Bassolas, Hugo Barbosa-Filho, Brian Dickinson, Xerxes Dotiwalla, Paul Eastham, Riccardo Gallotti, Gourab Ghoshal, Bryant Gipson, Surendra A Hazarie, Henry Kautz, et al. 2019. Hierarchical organization of urban mobility and its connection with city livability. Nature communications 10, 1 (2019), 4817

work page 2019

[6] [4]

Kunlin Cai, Jinghuai Zhang, Zhiqing Hong, William Shand, Guang Wang, Desh- eng Zhang, Jianfeng Chi, and Yuan Tian. 2024. Where Have You Been? A Study of Privacy Risk for Point-of-Interest Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 175–186

work page 2024

[7] [5]

Chu Cao and Mo Li. 2021. Generating mobility trajectories with retained data utility. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2610–2620

work page 2021

[8] [6]

Xinyu Chen, Jiajie Xu, Rui Zhou, Wei Chen, Junhua Fang, and Chengfei Liu. 2021. Trajvae: A variational autoencoder model for trajectory generation. Neurocom- puting 428 (2021), 332–339

work page 2021