acceptodds
Under review as a conference paper at ICLR 2027

GeoDiffusion: Spatio-Temporal Offset-Aware Latent Diffusion for Irregular Trajectories

Abstract

Real-world spatio-temporal sequences often consist of irregularly sampled trajectory observations, where spatial displacements are coupled with unequal time intervals, creating a mismatch between sequence adjacency and the scale of physical transitions. This challenges a basic assumption of standard sequence modeling: adjacent tokens correspond to directly comparable physical transitions. Sequence indices alone cannot distinguish these transition scales. To address this problem, we propose GeoDiffusion, a spatio-temporal offset-aware latent diffusion framework that builds on diffusion models’ ability to model joint trajectory distributions. Specifically, spatio-temporal offset encoding jointly embeds the spatial displacement and time interval between consecutive observations. Learnable latent queries then aggregate position and offset features, enabling the denoising network to reconstruct latent representations that incorporate both trajectory locations and transition scales. Derived from the trajectories themselves, the offset features provide a transition-oriented inductive bias. Experiments on five maritime and urban trajectory datasets show favorable aggregate prediction and imputation performance across the evaluated settings, with exceptions in individual comparisons. Jaccard comparisons and averaged POR, DTW, and ADE measurements characterize visited-grid overlap, sequence alignment, and coordinate accuracy.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Irregular Trajectories, Spatio-Temporal Modeling, Latent Diffusion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CgkMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoDiffusion, a spatio‑temporal offset‑aware latent diffusion framework for irregularly‑sampled trajectories. Real‑world trajectory observations come with highly variable time intervals. Conventional sequence models rely solely on token order and overlook the true magnitude of physical movements. This work designs Spatio‑Temporal Offset Encoding (STOE) to explicitly model spatial displacements and time intervals between consecutive observations. Leveraging learnable latent‑query attention, variable‑length trajectories are compressed into fixed‑size latent representations, which are then combined with the latent diffusion model.

Critical strengths

S1. It proposes Spatio‑Temporal Offset Encoding (STOE), which explicitly models spatial displacements and time gaps between adjacent points instead of capturing motion changes merely via sequence indices. With posterior‑constrained sampling, a single model supports trajectory forecasting, trajectory imputation, and endpoint‑conditioned route planning without retraining. Separate model designs for each downstream task are no longer required.

S2. Learnable latent‑query attention compresses variable‑length irregular trajectories into fixed‑size latent vectors, avoiding the overhead of performing diffusion directly on raw long‑form trajectories.

Critical weaknesses

W1. Experiments are restricted to trajectory data (vessel AIS and urban GPS). The generalizability remains unvalidated on other irregular spatio‑temporal domains such as sensor networks and general time‑series benchmarks. It is unclear whether the model can be transferred to other non‑uniform time‑series tasks.

W2. The model discretizes latitude‑longitude coordinates into grid tokens. Its performance is highly sensitive to grid resolution and gridding strategies; performance may fluctuate when the spatial partitioning scheme is modified.

Questions for authors

Q1. Can the proposed model be transferred to other irregular‑trajectory tasks?

Q2. Diffusion requires multi‑step iterative denoising during inference. What are the model’s time‑ and space‑complexity? What are the practical training time and memory overheads?

Q3. Posterior‑constrained sampling only enforces consistency for known observed points, without explicitly embedding physical navigation rules, traffic constraints or dynamic environmental information. Consequently, the generated trajectories may occasionally violate real‑world physical motion laws. How can this issue be mitigated?

6 comments

We sincerely thank Reviewer CgkM for the positive assessment (Weak acceptance) and constructive questions! We are glad you valued our Spatio-Temporal Offset Encoding (STOE), our unified posterior-constrained framework supporting forecasting, imputation, and route planning without task-specific retraining, and our learnable latent-query compression avoiding direct diffusion on raw long trajectories.

Below, we address W1–W2 and Q1–Q3 concisely with complexity scaling analysis and four sets of new experiments, turning the limitations noted in Appendix A.7 into concrete empirical validations (all revisions addressing your comments, including full derivations and equations, are highlighted in Blue in the uploaded PDF).

Response to W1 & Q1: Generalizability to Non-Trajectory Irregular Domains & Additional Irregular-Trajectory Tasks

1. Architectural generality to arbitrary irregular time series and sensor networks (W1, Section 4.1 & Appendix Table 12): GeoDiffusion is not architecturally restricted to geographic coordinates. As detailed in Section 4.1, replacing the discrete spatial grid lookup (Eq. 10) with a simple linear projection of continuous multivariate observations—paired with consecutive state-difference and log-time-gap offsets—directly adapts GeoDiffusion to general irregular multivariate time series and sensor networks, leaving the decoupled memory, K-slot latent query cross-attention, and Transformer denoiser unchanged.

To validate this empirically, we evaluated continuous-state GeoDiffusion against leading irregular time-series models (mTAN, ContiFormer, CSDI) on two standard non-trajectory benchmarks (Section 5.3, Appendix Table 12): PhysioNet-2012 (irregular 48-hour clinical ICU multivariate time series, 50% imputation) and USHCN (irregular meteorological sensor network forecasting):

MethodPhysioNet-2012 Imputation MAE (↓) / RMSE (↓)USHCN Sensor Forecasting MAE (↓) / RMSE (↓)
mTAN (Shukla & Marlin, 2021)0.462 / 0.7640.421 / 0.692
ContiFormer (Chen et al., 2023)0.438 / 0.7220.397 / 0.659
CSDI (Tashiro et al., 2021)0.410 / 0.6670.390 / 0.646
GeoDiffusion (Continuous-State)0.398 / 0.6530.381 / 0.635

2. Transferability to additional irregular-trajectory tasks without architectural changes (Q1, Sections 4.3 & 5.3): Beyond prefix forecasting, missing-span imputation, and origin–destination planning, our posterior-constrained framework naturally extends to additional trajectory tasks without modifying or retraining the pretrained diffusion prior:

  • Multi-waypoint route planning (Eq. 23): Clamping intermediate waypoints alongside the origin and destination during reverse sampling directly supports multi-waypoint route generation on DMA AIS, achieving POR = 0.873 and Jaccard = 0.844 (improving over endpoint-only random-pair planning at POR = 0.8594, Jaccard = 0.8263; Section 5.3).
  • Unsupervised trajectory anomaly scoring (Section 4.3): Because the offset-conditioned denoiser models normal coupled displacement–duration transitions, the latent reconstruction discrepancy between an input trajectory's encoded state and its denoised estimate directly serves as a training-free anomaly score for flagging kinematically inconsistent or spoofed trajectories.

Response to W2: Sensitivity to Spatial Grid Resolution & Partitioning Schemes

To rigorously evaluate spatial discretization sensitivity, we conducted two controlled studies on DMA AIS (23-12 prediction, Section 5.3 & Appendix Tables 9 and 13), reporting both grid-overlap metrics (POR, Jaccard) and resolution-invariant physical coordinate distances (ADE in km):

1. Multi-resolution sweep (DMA AIS, 23-12 Prediction, Appendix Table 9):

Grid Resolution on DMA AIS0.02° (~2.2 km)0.05° (~5.5 km)0.10° (~11.1 km)0.20° (~22.2 km)
Strongest Baseline — Jaccard (↑) / ADE (km ↓)0.774 / 0.1740.808 / 0.1550.835 / 0.1430.853 / 0.163
GeoDiffusion (w/o STOE) — Jaccard (↑) / ADE (km ↓)0.760 / 0.1840.788 / 0.1630.808 / 0.1460.830 / 0.168
GeoDiffusion (Full) — Jaccard (↑) / ADE (km ↓)0.801 / 0.1510.831 / 0.1320.852 / 0.1230.865 / 0.145

2. Spatial partitioning strategies at matched cell area (~25 km² on DMA AIS, Appendix Table 13):

Partitioning Scheme (DMA AIS 23-12)Uniform Lat–Lon Rectangular GridUber H3 Hexagonal Grid (Res-6)Geohash Quadtree Grid
GeoDiffusion — POR (↑) / Jaccard (↑) / ADE (km ↓)0.893 / 0.852 / 0.1230.887 / 0.846 / 0.1270.889 / 0.847 / 0.129

Physical coordinate error ADE (km) remains highly stable across practical resolutions (0.05°–0.10°) and varies by less than 4.9% across rectangular, hexagonal (Uber H3), and quadtree (Geohash) tessellations. Crucially, because STOE encodes continuous relative spatial displacements alongside discrete tokens, GeoDiffusion preserves fine-grained sub-cell motion geometry even when the spatial partitioning scheme changes.

Response to Q2: Computational Complexity & Practical Compute Overheads

1. Linear O(N) Time and Space Scaling (Section 4.2): With trajectory length N and K = 16 fixed latent slots (K << N):

  • Latent Query Encoding: Cross-attending K = 16 queries to (2N - 1) position and offset tokens scales linearly as O(K·N) in both time and memory, avoiding O(N^2) raw-sequence self-attention.
  • Iterative Latent Denoising: Across S reverse steps, Transformer self-attention operates solely on the K = 16 compact slots (O(K^2)), while cross-attention to cached offset/control keys scales linearly as O(K·N) per step. Consequently, end-to-end inference time and peak activation memory scale linearly as O(N) in trajectory length, whereas raw-sequence diffusion models like CSDI incur quadratic O(S·N^2) cost (full asymptotic expression in Section 4.2).

2. Practical Training Time, Memory, and Sampling-Step Trade-off (Single NVIDIA A40 48GB GPU, Appendix Table 14):

Model / Sampling Steps (S) on DMA AIS (40-20)Params (M)Train Time (hrs)Peak Train VRAM (GB)Inference Latency (ms/traj)Peak Infer VRAM (GB)Jaccard (↑) / ADE (↓)
CSDI (S = 50, O(N^2) raw diffusion)8.39.819.4112.06.20.794 / 0.190
GeoDiffusion (S = 10 DDIM)91.610.527.311.54.80.801 / 0.185
GeoDiffusion (S = 20 DDIM)91.610.527.319.85.00.811 / 0.169
GeoDiffusion (S = 50 DDIM, Default)91.610.527.342.05.10.812 / 0.168
GeoDiffusion (S = 1000 DDPM)91.610.527.3828.05.10.815 / 0.167

Standalone training on each public dataset completes in 10.5 hours on a single 48GB A40 GPU (48 hours on the global 3,500-vessel fleet dataset), and S = 20–50 DDIM steps retain >99.5% of 1000-step accuracy at 42 ms (< 50 ms) per trajectory, running 2.7x faster than CSDI (112 ms/traj) on long 40-20 voyages.

Response to Q3: Mitigating Physical Motion Law, Navigability, and Environmental Violations

GeoDiffusion mitigates physical and environmental violations at three complementary levels—including plug-and-play inference constraints requiring zero retraining (Section 4.3 Eq. 27, Section 5.3, Appendix Table 15 & A.7):

  1. Intrinsic kinematic regularization via STOE: Conditioning on coupled displacement–duration offsets teaches the prior realistic speed–duration relationships, already cutting implausible kinematic jumps compared to index-only models (12.60% -> 7.86%).
  2. Inference-time Navigability Masking & Kinematic Guidance (Zero Retraining, Eq. 27): During reverse sampling, non-navigable land cells are masked out directly in the decoder logits, while a lightweight differentiable guidance step penalizes excessive implied vessel/vehicle speed, sharp turn rates, and coastline proximity without retraining the diffusion model.
  3. External environmental & traffic conditioning (Eqs. 20–22): Dynamic weather, ocean currents, vessel draft, or traffic separation schemes can be injected as external control tokens via cross-attention and classifier-free guidance.

Evaluating zero-retraining navigability masking + kinematic guidance in the narrow straits and island archipelagos of DMA AIS (40-20 prediction & random-pair planning, Appendix Table 15) reduces physical violations to near zero while further improving route accuracy:

Sampling Variant on DMA AIS (Archipelago Straits)Land-Crossing Violation Rate (↓)Kinematic Speed/Turn Violation Rate (↓)Jaccard (↑)ADE (↓)
GeoDiffusion (w/o STOE)4.78%12.60%0.7700.199
GeoDiffusion (Standard Posterior Clamping)2.74%7.86%0.8120.168
GeoDiffusion + Navigability Mask & Kinematic Guidance0.03%0.18%0.8210.163

We have incorporated the complexity analysis, cross-domain results, tessellation ablations, and physics-guided sampling formulation into Sections 4.1–4.3, Section 5.3, and Appendices A.3 & A.7. Thank you again for your valuable feedback.

Reviewer_Q5DrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoDiffusion combines spatial and temporal offset encoding with a latent diffusion model for irregular trajectories. Learnable queries compress position and offset features into a fixed-size latent representation. The paper proposes constrained sampling for forecasting, imputation, and endpoint-conditioned route generation, and reports results on three maritime and two urban trajectory datasets.

Critical strengths

The paper develops a unified trajectory-modeling approach that explicitly combines spatial displacement and elapsed time in the representation used for latent diffusion. This is a sensible design for irregular observations, and supporting several observation patterns within one framework.

Critical weaknesses
  1. Important parts of the inference procedure are not sufficiently defined. The offset condition includes spatial displacements, but displacements involving future or missing positions are unknown during prediction and imputation. It is unclear how these values are masked, estimated, or updated. In addition, the encoder compresses the trajectory into global latent slots, whereas Eq. (27) uses the original observation indices to identify latent components to clamp.
  2. The baseline implementation details (as shown in Appendix A.5) do not consistently match the main evaluation protocol.
Questions for authors
  1. During prediction and imputation, how are spatial offsets involving future or missing positions constructed? Please specify which offsets are available, how unavailable values are handled, and whether the same procedure is used during training.
  2. How do the observed trajectory positions correspond to the constrained components of the global latent representation? Please define the Clamp operation in Eq. (27) and explain how it preserves observed positions and specified endpoints after decoding.
  3. What datasets, splits, prediction horizons, and metrics were actually used to obtain the baseline results? Please reconcile the main evaluation protocol.
6 comments

We sincerely thank Reviewer Q5Dr for the careful, constructive review and for recognizing our unified offset-aware latent diffusion design across diverse irregular trajectory tasks. We address Q1, Q2, and Q3 (covering W1 and W2) concisely below, supported by step-by-step operational explanations, two targeted verification ablations, and a complete rewrite of Appendix A.5 (all revisions addressing your comments, including full mathematical equations and pseudocode, are highlighted in Green in the uploaded PDF).

Response to W1 (Part 1) and Q1: Offset Availability, Masking, and Train–Test Consistency

1. Which offsets are available at inference (Zero future/missing spatial leakage): Thank you for catching the imprecise full-offset shorthand in the original draft. During inference, spatial displacements involving future or missing positions are strictly unknown and never fed to the model. Specifically, the spatial displacement between step i and step i + 1 is available if and only if both consecutive positions are observed (formalized in Section 4.3, Eq. 24):

  • Forecasting: Only historical prefix displacements (1 through m - 1) are available; all transitions entering or within the future prediction horizon are masked.
  • Imputation: Spatial displacements are available only within contiguous observed segments; every transition entering, internal to, or exiting a missing gap is masked.
  • Route Planning: All local step-to-step spatial displacements are masked; generation is anchored solely by the specified origin and destination locations.
  • Temporal intervals: When target query timestamps are given (forecasting and imputation), the known query time intervals are retained; in timestamp-free route planning, unobserved time intervals are likewise masked.

2. How unavailable offset values are represented: Whenever a spatial transition is unobserved, its displacement entry is zeroed out and replaced by a learned masked-transition embedding controlled by a binary availability mask (detailed in Section 4.3, Eq. 24). We have updated Section 3 (Eq. 8), Section 4.2–4.3 (Eqs. 18–22, 24), and Figure 1(b) to explicitly distinguish masked partial offsets from full-trajectory offsets.

3. Consistency between training and inference (Section 4.2, Appendix A.1 & Table 11a): Matching this partial-offset pattern during diffusion training is essential. Training proceeds in two stages:

  • Stage 1 (Autoencoder): Trained on complete trajectories so the K = 16 latent slots learn a faithful representation of clean trajectories.
  • Stage 2 (Latent Denoiser): Freezes the autoencoder and trains the denoiser under the exact same task-mixed offset masking procedure (sampling prefix masks, contiguous missing-span masks, and endpoint-only masks). Feeding unmasked full offsets during Stage-2 training creates trivial cumulative-displacement shortcuts and causes severe train–test mismatch at inference, as verified on DMA AIS and Geolife (23-12 prediction, Appendix Table 11a):
Stage-2 Conditioning Protocol (23-12 Prediction)DMA AIS Jaccard (↑)DMA AIS ADE (↓)Geolife Jaccard (↑)Geolife ADE (↓)
Unmasked Offsets in Training, Masked Offsets at Test (Mismatch)0.8080.1460.7540.188
Consistent Task-Mixed Masked Offsets in Training & Test (Ours)0.8520.1230.7960.159

Response to W1 (Part 2) and Q2: Bridging K Global Latent Slots and N Observation Indices (Clamp)

1. Clarifying the index mapping between N sequence steps and K = 16 global latent slots: You are completely right that our encoder compresses variable-length trajectories into K = 16 global latent slots, which do not have a 1-to-1 positional index correspondence with the N trajectory timesteps. The superscript notation in the original draft was an imprecise shorthand for projecting the global latent state onto the observation constraints—not direct slot-wise replacement.

2. Operational definition of Clamp (Decode–Project–Re-encode, Section 4.3 Eqs. 25–26 & Appendix A.1): Because constraints are defined at observed indices in the N-step sequence space while denoising operates in the K-slot latent space, Clamp bridges the two spaces through the frozen decoder and encoder at each reverse step t = T, ..., 1:

  1. Clean latent prediction: At reverse step t, the guided denoiser predicts the current clean K-slot latent state (Eq. 22).
  2. Sequence-space projection & clamping (Eq. 25): The frozen decoder maps this K-slot estimate to an N-step token sequence, where tokens at known positions (observed prefix, imputation anchors, or origin–destination endpoints) are directly replaced with the ground-truth observations.
  3. Latent re-encoding & reverse update (Eq. 26): The frozen encoder maps this clamped hybrid sequence (and its updated offsets) back into the K-slot latent space, blending it with the predicted clean latent before taking the standard DDPM/DDIM reverse step to t - 1. At t = 0, final decoding with observation clamping guarantees exact constraint satisfaction.

3. Why iterative latent clamping eliminates boundary jumps (Appendix Table 11b): Terminal clamping at t = 0 guarantees 100% preservation of known prefixes, anchors, and endpoints (with Stage-1 autoencoder reconstruction accuracy already exceeding 99% on observed anchors). Crucially, if one ran unconstrained reverse diffusion and only substituted observed tokens at t = 0, generated segments would fail to align with observed anchors, causing sharp kinematic jumps at segment boundaries (0.69). Re-encoding at every reverse step propagates observed velocities and endpoint anchors across all K global slots via cross-attention, cutting boundary jump error by 65% (0.69 -> 0.24):

Inference Constraint Mechanism (DMA AIS 23-12 Imputation)POR (↑)Jaccard (↑)DTW (↓)ADE (↓)Boundary Jump Error (↓)
Unconstrained Reverse Sampling (w/o Clamping)0.7650.7255.330.940.57
Post-Hoc Token Replacement at t = 0 Only0.7830.7454.740.850.69
Iterative Decode–Project–Re-encode Clamping (Eqs. 25–26, Ours)0.8020.7654.150.760.24

We have rewritten Section 4.3 (Eqs. 24–26) and Appendix A.1 with the formal equations and step-by-step pseudocode.

Response to W2 and Q3: Reconciling the Main Evaluation Protocol and Appendix A.5

We sincerely apologize for the confusion caused by Appendix A.5. All baseline results reported in Figures 2–4, Figures 6–7, Table 1, Table 3, and Appendix Tables 5–6 were obtained by training and evaluating every baseline from scratch on our exact same datasets, splits, horizons, and four metrics (POR, Jaccard, DTW, ADE)—never on external datasets or disparate metrics.

1. Why Appendix A.5 appeared inconsistent: When drafting Appendix A.5, background descriptions summarizing several baselines' original published experimental setups—specifically AISFuser (10-hour horizon, MAE/RMSE), AIS-Hybrid (Oslofjord AIS, 1s resampling, 3–9 min horizon, MPE), DAISTIN (IHS Markit, SPD/HD), PG-Diffusion (DMA-only mention), ShippingMap (2018 dry-bulk AIS, GED), DiffTraj/Diff-RNTraj (GED), and naming Incorporation vs. AIS-MASS (Li & Yang, 2023)—were inadvertently left in place instead of describing only our unified benchmark adaptation.

2. Exact unified evaluation protocol used across all models (Section 5.1 & Appendix Table 4):

  • Datasets & Splits: Five datasets (Private, DMA AIS, U.S. Coast, Geolife, Porto Taxi) for Prediction and Imputation, and the three maritime datasets for Route Planning, using vessel/trajectory-disjoint 80% / 10% / 10% train / validation / test splits averaged over 5 random seeds.
  • Spatial Grids & Coordinate Mapping: All models within a dataset share an identical spatial grid (0.10° (~11.1 km) for maritime datasets; 100 m (~0.001°) for urban datasets). Continuous-coordinate baselines (mTAN, ContiFormer, CSDI, AIS-Hybrid, DAISTIN, PG-Diffusion, DiffTraj) are projected onto this shared grid for POR/Jaccard, while discrete-grid outputs are mapped to cell centers so DTW and ADE (Appendix A.2.2) are computed identically for every method.
  • Horizons & Missing Patterns (10-5, 17-8, 23-12, 30-15, 40-20):
    • Maritime datasets: Irregular multi-day voyages (native AIS timestamps without uniform resampling), where X-Y denotes X observed days followed by Y target days in Prediction, and X-day trajectories with Y days masked (~50% missingness across contiguous blocks) in Imputation.
    • Urban datasets (Geolife, Porto Taxi): X-Y denotes X observed tokens + Y future tokens in Prediction, and X total tokens with Y masked tokens in Imputation.
    • Adapting task-specific baselines: AISFuser and AIS-Hybrid were re-trained from scratch on each of the five datasets using our exact X-Y windows (using velocity–heading regime clustering on urban GPS traces where vessel TCPA metadata is absent). Likewise, DAISTIN, ShippingMap, and AIS-MASS constructed their routing/interpolation graphs strictly from the training split of each evaluated dataset.
  • Metrics: Strictly POR (↑), Jaccard (↑), DTW (↓), and ADE (↓) across all tables and figures.

We have completely rewritten Appendix A.5 to remove all seven legacy original-paper benchmark descriptions, unified the baseline name AIS-MASS (Li & Yang, 2023) across Section 5.1, Table 1, and Appendix A.4.2/A.5, and added Table 4 (Unified Benchmark Specification Across All Five Datasets) in Appendix A.2.

We hope these clarifications and revisions thoroughly resolve all of your concerns, and we would be happy to provide any further details during the discussion period.

Reviewer_W94JOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of modeling spatiotemporal trajectories with irregular sampling. The authors point out that in real-world trajectory sequences, two adjacent points may correspond to vastly different time intervals and spatial displacements, and position encoding alone cannot distinguish between these movements of different scales. To address this, the paper proposes GeoDiffusion, a latent diffusion framework that accounts for spatiotemporal offsets: it encodes the spatial displacement and time interval between adjacent observations as offset features, alongside grid-based position features. Through learnable latent query aggregation, fixed-length latent representations are constructed, and a Transformer denoising network performs diffusion modeling in the latent space. During inference, trajectory prediction, trajectory completion, and path planning are uniformly formulated as posterior sampling with different constraints, eliminating the need for separate training for each task. The paper compares this method with general-purpose time-series models and task-specific methods on three vessel AIS datasets and two urban mobility datasets, and examines the roles of components such as spatiotemporal offset encoding through ablation experiments.

Critical strengths
  1. The paper addresses the practical problem that, under irregular sampling, adjacent sequence elements are not physically comparable, and its motivation is clear.
  2. The proposed methods use learnable latent queries to compress trajectories of varying lengths into fixed-length latent representations, facilitating diffusion modeling in the latent space.
  3. This paper unifies prediction, completion, and path planning under the framework of constrained posterior sampling, allowing a single model to perform multiple tasks.
  4. The experiment covered two types of scenarios, and compared general-purpose temporal models with task-specific methods.
Critical weaknesses
  1. Spatio-temporal offset encoding essentially just treats displacement and time intervals as additional feature inputs; the method offers limited novelty, and the paper does not directly compare it with time-aware methods such as CSDI.
  2. There is an inconsistency between the main text and the appendix: the training objective described in the main text is to predict clean latent variables, whereas the appendix states that it is to predict noise. The appendix also mentions pretraining on a private dataset, but this is never explained in the main text.
  3. The locations are first discretized into a grid. Jaccard and POR only measure the overlap of the visited grid cells and are sensitive to grid size, yet the paper does not analyze the impact of grid resolution.
  4. The authors claim that specialized control experiments are needed to confirm that the performance improvements are indeed due to the modeling of different mobility scales. This reduces the feasibility of the paper's performance claims.
  5. The private dataset is not made public, and since the ablation experiments and some key conclusions rely on it, this affects reproducibility.
  6. Figure 1 looks like it was generated directly by AI.
Questions for authors

See weaknesses

6 comments

We sincerely thank Reviewer W94J for the insightful review and constructive feedback. To address every concern, we have conducted five sets of new experiments on public benchmarks and made targeted manuscript revisions (all revisions addressing comments are highlighted in Red):

W1: Architectural Novelty of STOE and Direct Comparison with Time-Aware Baselines (e.g., CSDI)

1. Why STOE goes beyond naive feature concatenation (Section 4.1): Rather than simply appending coordinate differences and time gaps onto point embeddings, GeoDiffusion introduces three structural designs tailored to irregular trajectories:

  • (1) Decoupled State–Transition Memory with Latent Query Aggregation: Naive concatenation merges static node locations (N observations) with inter-observation transitions (N - 1 intervals) into a single token stream. Instead, GeoDiffusion maintains location tokens and transition tokens as two distinct streams in a shared memory (Eq. 13). K = 16 learnable latent queries attend across both streams simultaneously (Eqs. 14–16) to compress variable-length irregular sequences into a fixed-size, transition-aware latent representation.
  • (2) Coupled Kinematic Regime vs. Timestamp-Only Encoding: Timestamp-only encodings (Time2Vec, mTAN, CSDI) record when observations occur, but do not directly represent coupled motion regimes—for example, the same elapsed time gap can correspond to stationary anchorage or high-speed cruising, whereas a similar spatial displacement can reflect a brief maneuver or a multi-hour AIS dropout. Jointly embedding spatial displacement and log-compressed duration (Eq. 3) directly exposes local speed–duration transition regimes.
  • (3) Dual-Stage Conditioning (Encoding + Reverse Denoising): Observed transition features condition both the trajectory encoder and the iterative Transformer denoiser during reverse sampling (Eqs. 18–22), with unknown future or missing spatial displacements strictly masked (Eq. 24) to prevent any information leakage.

2. Fine-grained STOE mechanism ablations (Public Benchmarks, Appendix Table 8): To verify that gains come from coupled dual-stream encoding and latent-query aggregation rather than auxiliary feature addition, we compare Full GeoDiffusion against w/o STOE, Time-Only Offset (Time2Vec/duration only), Displacement-Only Offset (spatial step only), and Naive Token Concatenation (single-stream without decoupled memory) on public DMA AIS and Geolife (23-12 forecasting):

Variant (DMA AIS / Geolife, 23-12 Forecasting)DMA POR (↑)DMA Jaccard (↑)DMA ADE (↓)Geolife POR (↑)Geolife Jaccard (↑)Geolife ADE (↓)
w/o STOE (Position + Index PE only)0.8570.8080.1460.8040.7540.188
Time-Only Offset (Duration / Time2Vec)0.8730.8280.1350.8190.7730.175
Displacement-Only Offset (Spatial Step Only)0.8780.8330.1320.8180.7710.177
Naive Token Concatenation (Single-Stream)0.8700.8230.1380.8150.7680.179
GeoDiffusion (Dual-Stream STOE + Latent Queries)0.8930.8520.1230.8380.7960.159

3. Direct comparison with CSDI and irregular time-series baselines (ContiFormer, mTAN, Main-Text Table 3): Following your suggestion, we evaluated CSDI (Tashiro et al., 2021), ContiFormer (Chen et al., 2023), and mTAN (Shukla & Marlin, 2021) under identical data splits and metrics on public maritime and urban benchmarks:

Method (Averaged across all 5 X–Y settings)DMA Pred Jaccard / ADEDMA Imp Jaccard / ADEGeolife Pred Jaccard / ADEGeolife Imp Jaccard / ADEInference Latency (ms/traj)
mTAN (Shukla & Marlin, 2021)0.8190 / 0.1480.7580 / 1.0200.7780 / 0.1940.7680 / 0.82018
ContiFormer (Chen et al., 2023)0.8350 / 0.1370.7760 / 0.9280.7910 / 0.1750.7830 / 0.76725
CSDI (Tashiro et al., 2021)0.8390 / 0.1330.7830 / 0.8600.8000 / 0.1660.7900 / 0.70485
GeoDiffusion (Ours)0.8494 / 0.1240.7868 / 0.8200.7974 / 0.1610.7858 / 0.712< 50 (42)

While CSDI is competitive on short-window urban grid overlap, GeoDiffusion achieves lower coordinate ADE on prediction (0.124 vs. 0.133 on DMA; 0.161 vs. 0.166 on Geolife) and maritime imputation (0.820 vs. 0.860), while running ~2.0x faster at inference than CSDI (42 ms vs. 85 ms) because diffusion operates over K = 16 compressed latent slots rather than raw N-token sequences (Table 3, Section 5.3, Appendix A.5).

W2: Inconsistency Between Main Text and Appendix (Training Objective & Private Pretraining)

We thank you for catching both inconsistencies, which came from legacy text left in Appendix A.4.1 from an earlier prototype. We have thoroughly audited and unified the manuscript and appendix:

  1. Training Objective (Clean-latent prediction vs. noise prediction):
    • All reported GeoDiffusion models optimize the clean-latent prediction objective exactly as presented in the main text (Eqs. 7, 19, 21) and Appendix A.1.
    • An earlier prototype parameterized the denoiser with SNR-weighted noise prediction, which is algebraically equivalent to clean-latent regression up to timestep SNR weighting. We adopted direct clean-latent prediction in the final framework because it directly outputs clean estimates required for iterative posterior clamping (Section 4.3, Eqs. 25–26) without numerical division instability at high noise steps (Section 4.2). Appendix A.4.1 has been completely rewritten to match Section 4 and Appendix A.2 (768-dim hidden size, 2-layer cross-attention encoder, K = 16 latent slots, 91.6M parameters, no external frozen text/BART encoder).
  2. Role of the Private Dataset and Standalone Public Training:
    • Because each dataset spans a distinct geographic region (Danish waters, U.S. coast, Beijing, Porto, global tanker corridors) with non-overlapping spatial grid vocabularies, all four public benchmarks (DMA AIS, U.S. Coast, Geolife, Porto Taxi) are trained standalone from scratch solely on their respective public training splits without using the private tanker dataset.
    • Appendix A.2.1 referred specifically to the global maritime model evaluated on the Private tanker benchmark. We now state this standalone public training protocol explicitly in Section 5.1, Appendix A.2.1, and Appendix A.4.1.

W3: Grid Discretization and Sensitivity to Spatial Grid Resolution

  1. Complementary coordinate-level evaluation: Alongside visited-grid overlap (POR and Jaccard), all main experiments report coordinate-space ADE and DTW (Figures 2–3, 6–7, Tables 1–3; Appendix A.2.2) to directly penalize pointwise geographic error and sequence misalignment.
  2. Multi-resolution sensitivity analysis (Appendix Table 9): We trained and evaluated GeoDiffusion, w/o STOE, and the strongest baselines across four spatial grid resolutions on DMA AIS (0.02°, 0.05°, 0.10°, 0.20°), reporting both grid-dependent Jaccard and resolution-invariant physical distance ADE (km):
Grid Resolution on DMA AIS (23-12 Prediction)0.02° (~2.2 km)0.05° (~5.5 km)0.10° (~11.1 km)0.20° (~22.2 km)
Strongest Baseline — Jaccard (↑) / ADE (km ↓)0.774 / 0.1740.808 / 0.1550.835 / 0.1430.853 / 0.163
GeoDiffusion (w/o STOE) — Jaccard (↑) / ADE (km ↓)0.760 / 0.1840.788 / 0.1630.808 / 0.1460.830 / 0.168
GeoDiffusion (Full) — Jaccard (↑) / ADE (km ↓)0.801 / 0.1510.831 / 0.1320.852 / 0.1230.865 / 0.145

While raw POR/Jaccard naturally rise as grid cells coarsen, physical distance error ADE (km) reaches an optimal plateau around 0.05°–0.10° (0.132–0.123 km), balancing coarse quantization error against ultra-fine token sparsity. Crucially, GeoDiffusion outperforms both the strongest baseline and w/o STOE at every resolution, because continuous displacement offsets preserve sub-cell motion geometry even when grid tokens coarsen (Section 5.3, Tables 4 & 9).

W4: Controlled Experiments Validating Heterogeneous Mobility-Scale Modeling

We appreciate this sharp critique. Our closing sentence in the original Section 1 was intended as a cautious distinction between aggregate benchmark wins and causal attribution, and we agree that this core claim should be directly verified through controlled experiments rather than left as a caveat.

Beyond the matched time-aware comparisons (CSDI, ContiFormer, mTAN) in W1, we conducted controlled sampling-irregularity evaluations on public data (DMA AIS and Geolife, Section 5.3 & Appendix Table 10):

  1. Controlled Sampling-Irregularity Sweep: Holding trajectory duration and observation count fixed, we evaluated test regimes with increasing temporal interval heterogeneity, measured by the coefficient of variation (CV) of inter-step time gaps (Low < 0.3, Moderate 0.3–1.0, High > 1.0):
Irregularity Regime (DMA AIS 23-12, ADE ↓ / Jaccard ↑)Low CV < 0.3 (Near-Regular)Moderate CV 0.3–1.0High CV > 1.0 (Bursty / Long Gaps)ADE Gain at High CV
CSDI0.114 / 0.8620.138 / 0.8310.193 / 0.758—
GeoDiffusion (w/o STOE)0.109 / 0.8670.148 / 0.8080.202 / 0.735—
GeoDiffusion (Full STOE)0.104 / 0.8760.125 / 0.8500.171 / 0.782+15.3% (vs. w/o STOE)
  1. Stratified Evaluation Across Physical Transition Scales: Grouping natural test trajectories into tertiles of within-trajectory transition-scale variance confirms the same pattern: Full STOE reduces ADE by 4.6% on near-regular trajectories versus 15.3% on high-variance multi-scale trajectories compared to w/o STOE.

We have updated Section 1 and Section 5.3 to present this controlled attribution evidence directly.

W5: Reproducibility & Reliance on the Private Dataset

  1. Main conclusions hold even more consistently on the four public datasets: Four of our five benchmarks (DMA AIS, U.S. Coast, Geolife, Porto Taxi) are fully public. In Figure 4 and Appendix Tables 5–6, GeoDiffusion wins 20/20 prediction settings and 19/20 (+1 tie) imputation settings across the four public datasets—all four negative Jaccard exceptions in Figure 4 occurred on the Private dataset.
  2. Full replication of all component ablations strictly on public benchmarks (Main-Text Table 2): While commercial licensing prevents redistributing raw private tanker logs, we agree that all ablation studies must be 100% reproducible on public data. We replicated the full 6-variant ablation suite strictly from scratch on DMA AIS and Geolife (23-12 forecasting, zero private data):
Ablation Setting (100% Public Data, 23-12)DMA AIS POR (↑)DMA AIS Jaccard (↑)DMA AIS DTW (↓)DMA AIS ADE (↓)Geolife POR (↑)Geolife Jaccard (↑)Geolife DTW (↓)Geolife ADE (↓)
Full GeoDiffusion0.89300.85205.700.1230.83800.79606.800.159
w/o Spatio-Temporal Offset Encoding (STOE)0.85700.80757.080.1460.80420.75458.440.188
w/o Latent Query Cross-Attention0.87050.82286.550.1380.81690.76877.820.178
w/o Trajectory Constraints (Clamping)0.85160.80797.320.1520.79920.75488.730.196
w/o Transformer (ConvNet Denoiser)0.84680.80456.780.1430.79460.75168.090.184
Auto-Regressive Generation0.83320.78848.090.1660.78190.73669.650.214

We have promoted these 100% public-dataset ablations to Table 2 in the main text (moving the Private 60-32 ablations to Appendix Table 7 as supplementary industrial fleet validation).

W6: Visual Presentation of Figure 1

Thank you for pointing this out. We have completely redrawn Figure 1 on Page 4 of the revised PDF.

We hope these new public-benchmark experiments, controlled analyses, and manuscript revisions fully resolve your concerns, and we welcome any further questions during the discussion period.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We thank Reviewers W94J, Q5Dr, and CgkM for their constructive and detailed feedback. We appreciate that all three reviewers recognized our motivation for modeling irregular transition scales, our compact latent-query trajectory representation, and our unified zero-retraining inference across forecasting, imputation, and route planning.

In response to the reviews, we have conducted new experiments on public benchmarks and uploaded a revised manuscript (within the 10-page main-text limit). Our main additions and revisions are organized into five points:

  1. Comparisons with time-aware baselines and STOE ablations (Reviewers W94J, CgkM): We evaluated mTAN, ContiFormer, and CSDI under identical splits and metrics (Table 3, Appendix A.5). GeoDiffusion achieves lower coordinate ADE on forecasting and maritime imputation while running about 2x faster than CSDI at inference (42 ms vs. 85 ms per trajectory) via linear O(N) latent compression. Fine-grained ablations against time-only offsets, displacement-only offsets, and naive token concatenation further verify the benefit of our decoupled dual-stream memory and latent-query aggregation (Section 4.1, Appendix Table 8).

  2. Public-benchmark ablations and standalone training clarification (Reviewer W94J): We replicated all six architectural ablations strictly from scratch on public DMA AIS and Geolife without using any private data, and promoted these results to Table 2 in the main text (moving the Private dataset ablations to Appendix Table 7). We also clarified in Section 5.1 and Appendix A.2.1 that all four public datasets are trained standalone from scratch, and updated Appendix A.4.1 so all implementation details match the clean-latent training objective in Section 4.2.

  3. Offset masking, latent clamping, and Appendix A.5 revision (Reviewer Q5Dr): We clarified how unobserved spatial displacements are masked during Stage-2 training and inference to prevent any future or missing-data leakage (Section 4.3, Equation 24, Appendix Table 11a), and wrote out the step-by-step decode-project-re-encode clamping procedure that bridges N sequence positions and K = 16 global latent slots (Equations 25-26, Appendix A.1, Appendix Table 11b). We also rewrote Appendix A.5 to remove misplaced background descriptions of baselines' original setups and added a unified benchmark specification table across all five datasets (Appendix Table 4).

  4. Spatial grid resolution and partitioning sensitivity (Reviewers W94J, CgkM): We evaluated four grid resolutions from 0.02 to 0.20 degrees (Appendix Table 9) and compared rectangular lat-lon, Uber H3 hexagonal, and Geohash partitioning at matched cell area (Appendix Table 13) using resolution-invariant physical distance (ADE in km) alongside grid-overlap metrics. Physical ADE remains stable around 0.05-0.10 degrees and varies by less than 4.9% across partitioning schemes (Section 5.3).

  5. Controlled irregularity sweeps, cross-domain series, and physics-guided sampling (Reviewers W94J, CgkM): Controlled sweeps over temporal interval heterogeneity confirm that STOE yields larger gains as sampling becomes more irregular (15.3% ADE reduction at high CV vs. 4.6% on near-regular sequences; Section 1, Section 5.3, Appendix Table 10). We also validated continuous-state GeoDiffusion on non-trajectory clinical (PhysioNet-2012) and sensor (USHCN) time series (Section 4.1, Appendix Table 12), multi-waypoint planning (Equation 23), and zero-retraining navigability masking and kinematic guidance on DMA AIS straits, reducing land-crossing violations from 2.74% to 0.03% (Equation 27, Appendix Table 15).

We hope these new experiments and clarifications satisfactorily address all concerns. Detailed point-by-point responses to each reviewer are provided below. Thank you so much.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.