acceptodds
Under review as a conference paper at ICLR 2027

PAST: A Primary-Auxiliary Spatio-Temporal Network for Traffic Time Series Imputation

Abstract

Traffic time series imputation is vital for transportation reliability, yet diverse missing patterns–including random, fiber, and block missing–pose significant challenges. Existing models typically disentangle spatial and temporal dependencies based on internal relationships but struggle to adapt to stochastic missing positions or capture the long-term and large-scale dependencies required for extensive missingness. We propose the Primary-Auxiliary Spatio-Temporal network (PAST), which categorizes data patterns into primary patterns (derived from internal relationships) and auxiliary patterns (driven by external factors like timestamps and node attributes). PAST employs a Graph-Integrated Module (GIM) to capture primary patterns via dynamic graphs with an interval-aware dropout mechanism, and a Cross-Gated Module (CGM) to extract auxiliary patterns through bidirectional gating. These modules interact via shared hidden vectors under an ensemble self-supervised framework. Experiments on three datasets across 27 missing conditions demonstrate that PAST consistently outperforms seven state-of-the-art baselines, achieving accuracy improvements of up to 26.2% in RMSE.

Primary area:
learning on time series and dynamical systems
Keywords:
traffic, time-series prediction

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_jz2AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses traffic time-series imputation under random, single-sensor interval, and multi-sensor block missingness. PAST has two components. The Graph-Integrated Module (GIM) propagates observed measurements over a temporal graph defined by the observation mask and then over the road network. It uses interval-aware edge dropout to reduce reliance on nearby observations. The Cross-Gated Module (CGM) encodes sensor identity and calendar context, injects these features into the GIM layers, and predicts a residual that is added to the GIM estimate at missing entries.

Critical strengths
  1. The primary and auxiliary distinction is well matched to structured missingness: the GIM uses measurements where they are available, while per-entry calendar and sensor context is injected into its layers to support estimates inside a gap.

  2. Interval-aware edge dropout is a targeted regularization choice. It encourages the model to rely less on the nearest observations while retaining the possibility of using local evidence, which fits the paper’s motivation for handling long gaps.

  3. The model keeps the road graph sparse during spatial propagation and uses repeated one-hop products instead of forming dense graph powers. This gives the design a clear computational advantage for networks with many sensors and relatively few road edges.

Critical weaknesses
  1. Every missing receiver has the same incoming temporal neighbors: all observed steps and the CGM context node. The aggregation therefore does not explicitly weight observations by their temporal distance to each target. Target-specific distinctions must instead arise through the node features and CGM injection, and the paper does not establish how well this handles different temporal relationships to the observed data.

  2. The residual-learning procedure is not fully specified. The CGM target depends on the current GIM prediction, but both losses are stated to be minimized together. If gradients from the CGM loss pass through the GIM prediction, the training is not equivalent to fitting a residual after a fixed primary model, as suggested by the gradient-boosting interpretation. The resulting interaction between the two branches is unclear.

  3. The residual branch is trained from observed entries but is applied only at missing entries. The paper does not describe a calibration or consistency constraint that ensures residuals learned on observed locations transfer to entries with different mask-dependent support. This may be especially relevant for long gaps, where the GIM feature distribution differs substantially from that at observed entries.

  4. The auxiliary branch uses sensor identity and calendar embeddings but no dynamic exogenous signals. It can represent recurring profiles, but it has no direct input for incidents, weather effects, special events, or other nonrecurring regime changes inside a gap. The measurement branch can mitigate this when nearby evidence remains, but the limitation is most relevant to the long and wide gaps that motivate the method.

  5. CSDI, ImputeFormer, and OPCR are evaluated on one representative setting for each missingness type rather than across the full range of random rates and gap lengths or widths. Their relative performance over the broader configuration grid therefore remains unclear, which limits the strength of the comparison with recent methods.

Questions for authors
  1. For two missing time steps with different temporal distances from the observed entries, what part of the GIM representation makes their primary estimates different before CGM context is injected? Please clarify whether the implementation includes any target-dependent temporal encoding or weighting.

  2. Are GIM predictions detached when forming the CGM residual target, or are both branches optimized jointly through that target? Please state the exact gradient path and explain how it relates to the gradient-boosting interpretation.

  3. What criteria were used to select the representative settings for CSDI, ImputeFormer, and OPCR? Were these settings chosen to reflect the range of gap severities within each missingness type?

5 comments
AuthorsOpenReview ↗

Response to Reviewer jz2A (Part 1/3): Target-Specific Temporal Weighting

We sincerely thank the reviewer for the detailed and technically precise review. We are encouraged that the reviewer found the primary/auxiliary distinction well matched to structured missingness, interval-aware dropout a targeted regularization choice, and the sparse spatial propagation computationally advantageous. The questions helped us state several parts of the method much more precisely. We answer each point below, and all changes are included in the revised PDF (changes in blue).

Q1 / W1: "For two missing time steps with different temporal distances from the observed entries, what part of the GIM representation makes their primary estimates different before CGM context is injected? Please clarify whether the implementation includes any target-dependent temporal encoding or weighting."

Yes, the implementation includes target-dependent temporal weighting. Every missing receiver has the same set of sources (the observed steps and the context node), but the weights on these edges are not shared. The temporal adjacency is a learnable L×LL\times L matrix Θ\Theta, indexed by the target step uu and the source step vv, together with a learnable weight θu\theta_u for the context vertex:

[AT]uv=Θuv[A_T]_{uv} = \Theta_{uv} if step vv is observed, θu\theta_u if vv is the context vertex, and 00 otherwise.

Each row is normalized by the sum of absolute weights over the available sources. Each target therefore has its own weighting over the observed steps, and this weighting encodes temporal distance. The case definition in the original Section 4.2 gave the support of this matrix; the revised Eq. (3) writes out the learnable weights on it.

The learned weights show this directly.

Table A. Mean ∣AT[i,j]∣\lvert A_T[i,j]\rvert by temporal distance ∣i−j∣\lvert i-j\rvert in the first forward temporal layer, and the share of each row's weight within one hour (∣i−j∣≤4\lvert i-j\rvert\le 4; uniform weighting gives 0.088). New Appendix M.

Model1248163264Share within 1 h
METR-LA, random r=0.45.572.340.590.0920.0310.0250.0320.84
METR-LA, fiber l=963.571.490.520.2580.1980.1120.0550.54
PeMS-Bay, block (48,5)3.040.590.130.0650.0550.0470.0550.65

Three observations follow:

  1. The weights decay sharply with distance. For the METR-LA random model, 84% of each row's weight lies within one hour, versus 8.8% under uniform weighting.
  2. The rows are highly target-specific. The mean pairwise cosine similarity between rows is 0.06.
  3. The weighting adapts to the gap. Because the weights are renormalized over the available sources, a missing step next to an observation concentrates on it, while a step deep inside a gap shifts its weight to more distant observations and to the context node. Under fiber missingness (l=96), the learned profile is flatter: 54% of the weight lies within one hour, and the weight at distance 32 is 4.5 times that of the random model. This is the intended effect of interval-aware dropout.

Two missing steps at different distances from the observations therefore aggregate different observations with different weights, and obtain different primary estimates before any CGM context is injected.

AuthorsOpenReview ↗

Response to Reviewer jz2A (Part 2/3): Gradient Path and Transfer of the Residual

Q2 / W2: "Are GIM predictions detached when forming the CGM residual target, or are both branches optimized jointly through that target? Please state the exact gradient path and explain how it relates to the gradient-boosting interpretation."

The GIM prediction is detached. The revised Eq. (2) reads LGIM=MSE(M⊙YGIM, M⊙X)\mathcal L_{\rm GIM} = \mathrm{MSE}(M\odot Y_{\rm GIM},\, M\odot X) and LCGM=MSE(M⊙(sg(YGIM)+YCGM), M⊙X)\mathcal L_{\rm CGM} = \mathrm{MSE}(M\odot(\mathrm{sg}(Y_{\rm GIM}) + Y_{\rm CGM}),\, M\odot X), where sg\mathrm{sg} denotes stop-gradient. The exact gradient path is:

LossUpdates GIMUpdates CGM
LGIM\mathcal L_{\rm GIM}YesYes, through the context injected into every GIM layer
LCGM\mathcal L_{\rm CGM}No (stop-gradient)Yes

The interaction through the injected context is the intended coupling between the two modules. As the reviewer points out, both objectives are minimized jointly, and the residual target moves as GIM is updated, so the training is not stage-wise gradient boosting. Section 4.1 now names the scheme joint residual learning with a stop-gradient and states this gradient path, and the introduction uses the same name.

W3: "The residual branch is trained from observed entries but is applied only at missing entries. The paper does not describe a calibration or consistency constraint that ensures residuals learned on observed locations transfer to entries with different mask-dependent support. This may be especially relevant for long gaps [...]"

The residual branch is supervised on gap-like entries. At every training step, a random subset of the observed entries is additionally hidden from the model input while remaining in both losses:

  • under fiber and block configurations, a contiguous interval is hidden in half of the sensor series of each batch;
  • under random configurations, 10% of the entries are hidden at random.

At these hidden entries, GIM must infer the value from support that mimics real gaps, including long gaps, and the residual branch learns to correct exactly these estimates. The residual branch is therefore trained on the same kind of entries that it corrects at inference. The revised Section 4.4 describes this procedure. The real-world evaluation supports the transfer: on real sensor failures, adding the residual branch to the injection-only model lowers RMSE from 0.497 to 0.458 (see our response to F7jZ).

AuthorsOpenReview ↗

Response to Reviewer jz2A (Part 3/3): Exogenous Signals and Representative Settings

W4: "The auxiliary branch uses sensor identity and calendar embeddings but no dynamic exogenous signals. It can represent recurring profiles, but it has no direct input for incidents, weather effects, special events, or other nonrecurring regime changes inside a gap. [...]"

We agree with the reviewer that calendar and sensor-identity context captures recurring profiles, and that non-recurring events such as incidents, weather or special events are outside its scope. The revised Section 6 now states this explicitly. The CGM interface is generic: any per-entry exogenous signal can be embedded and gated in the same way as the calendar features, without changing the rest of the model. The standard benchmarks used here and in prior work (METR-LA, PeMS-Bay and LargeST) do not provide such signals, and we consider them a natural extension of PAST.

Q3 / W5: "What criteria were used to select the representative settings for CSDI, ImputeFormer, and OPCR? Were these settings chosen to reflect the range of gap severities within each missingness type?"

The criterion was one configuration per missingness type at the common missing rate of 0.4, fixed before any results were obtained: random r=0.4, fiber l=24 and block (48,10). The reason for evaluating a single configuration per type is computational cost, in particular CSDI's diffusion sampling over 27 configurations and two evaluation settings.

These settings cover both ends of the severity range. Fiber l=24 is the shortest fiber gap, and block (48,10) has the widest spatial span. Because the margin of PAST grows with gap length, the shortest fiber gap is the fiber setting least favorable to PAST; even there, PAST leads all three recent methods out of sample on METR-LA.

Table B. Out-of-sample RMSE on METR-LA in the representative configurations (original Table 1).

MethodRandom r=0.4Fiber l=24Block (48,10)
CSDI0.27110.47740.5499
ImputeFormer0.35240.43990.5417
OPCR0.28740.53030.5510
PAST0.21080.30830.3442

How the methods behave as gaps become longer and wider is covered by the seven established baselines, including the traffic-oriented GRIN and STCPA, on all 27 configurations. There, PAST is best in 53 of 54 settings, and its margin grows with gap length and spatial span. The revised Section 5.1 states this selection criterion.

We hope these responses address the reviewer's concerns. We would be sincerely grateful if the reviewer would take them into account in the final assessment, and we would be glad to provide any further clarification during the discussion period.

Reviewer_r4jEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PAST for traffic time series imputation, aiming to mitigate accuracy degradation caused by limited observational evidence under structured missingness. PAST organizes imputation by information availability, combining observed measurements with temporal and sensor context that remains available at missing entries. PAST's primary module, the Graph-Integrated Module (GIM), learns primary patterns from observations through temporal graph aggregation followed by spatial propagation over the road graph. For GIM's temporal graph, the authors introduce interval-aware dropout during training. It preferentially removes edges from nearby observed time points to missing targets, discouraging overreliance on the closest observations. PAST's auxiliary module, the Cross-Gated Module (CGM), learns auxiliary patterns by combining temporal and spatial embeddings through cross-gating. It injects the resulting context into each GIM layer and learns a residual correction to the GIM prediction. The authors compare PAST with baselines across multiple datasets and missingness patterns to evaluate its overall performance. They further investigate the main design choices through analyses of the primary–auxiliary organization, component ablations, and evaluation of the dropout strategy.

Critical strengths
  1. The paper addresses traffic time series imputation under random, fiber, and block missingness, a problem relevant to the reliable operation of intelligent transportation systems. Experimental comparisons with diverse baselines across multiple datasets and missingness patterns demonstrate the overall effectiveness of the combined approach, particularly in handling structured missingness.

  2. Targeted experiments provide empirical support for the effectiveness and rationale of the main design choices. The comparison with the temporal/spatial split variant supports the value of PAST's primary–auxiliary organization. Ablation studies on components such as CGM and cross-gating assess their contributions to model performance. Comparisons among dropout strategies further support the effectiveness of interval-aware dropout, which improves imputation accuracy under structured missingness while maintaining good performance under random missingness.

Critical weaknesses
  1. The ablations support the auxiliary module (CGM) as a whole but do not isolate the contributions of context injection and residual correction. Removing CGM eliminates both mechanisms, making their individual contributions to the performance gains unclear. Additional ablation experiments retaining only context injection or only residual correction would help reveal the sources of the method's performance gains.

  2. In Tables 8 and 9, four out-of-sample results for the variant without GIM show MAE values exceeding the corresponding RMSE values. For example, on METR-LA, MAE exceeds RMSE under both random missingness (1.1333 vs. 1.0633) and fiber missingness (1.2147 vs. 1.1021). For the same set of prediction errors, MAE cannot exceed RMSE. The authors should review the relevant experimental procedures and clarify or correct these entries to ensure accurate reporting and strengthen the rigor of the paper.

Questions for authors

Q1. Could you evaluate variants retaining only context injection or only residual correction to isolate their individual contributions?

Q2. Could you review the experimental procedures associated with the four entries where MAE exceeds RMSE in Tables 8 and 9, and clarify or correct these results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer r4jE (Part 1/2): Context Injection vs. Residual Correction

We sincerely thank the reviewer for the careful review, and for recognizing the relevance of the problem and the targeted experiments behind our main design choices. Both questions are precise and very helpful, and we answer them below. All changes are included in the revised PDF (changes in blue).

Q1: "Could you evaluate variants retaining only context injection or only residual correction to isolate their individual contributions?"

We added the two requested variants (new Appendix J):

  • Injection-only: CGM injects context into each GIM layer but produces no residual (Y=YGIMY = Y_{\rm GIM}, and LCGM\mathcal L_{\rm CGM} is removed).
  • Residual-only: CGM produces the residual, but no context is injected into GIM (the context node is removed from the temporal graph).

All variants are trained on METR-LA under an identical budget.

Table A. Out-of-sample RMSE on METR-LA (change relative to PAST).

VariantRandom r=0.4Fiber l=48Block (48,5)
PAST (injection + residual)0.28960.43220.4150
Injection-only0.2915 (+0.7%)0.4383 (+1.4%)0.4298 (+3.6%)
Residual-only0.2908 (+0.4%)0.4810 (+11.3%)0.4640 (+11.8%)
w/o CGM (neither)0.2906 (+0.3%)0.4997 (+15.6%)0.4736 (+14.1%)

We also evaluated the same variants on real sensor-failure patterns extracted from METR-LA (see our response to F7jZ for the protocol).

Table B. Imputation under real-world missingness (METR-LA test period).

VariantRMSEMAE
PAST (injection + residual)0.4580.309
Injection-only0.4970.368
Residual-only0.5150.354
w/o CGM (neither)0.5230.346

The two tables give a consistent picture:

  1. Context injection provides most of CGM's gain under structured missingness. Removing it raises out-of-sample RMSE by 11.3% (fiber) and 11.8% (block).
  2. The residual correction adds a further gain, 1.4% under fiber and 3.6% under block missingness, the largest where gaps span several sensors.
  3. Under random missingness, all variants are within 1%, as expected when local observations are abundant.
  4. On real failures, both mechanisms are needed for the best result, and the ordering of the variants is the same as under synthetic structured missingness.
AuthorsOpenReview ↗

Response to Reviewer r4jE (Part 2/2): MAE and RMSE Entries in Tables 8-9

Q2: "Could you review the experimental procedures associated with the four entries where MAE exceeds RMSE in Tables 8 and 9, and clarify or correct these results?"

We thank the reviewer for catching this. The reviewer is correct: MAE cannot exceed RMSE on the same set of errors. The four affected entries are MAE values of the w/o-GIM variant that were entered incorrectly when the ablation table was assembled. We re-ran this variant with the main evaluation code, which asserts MAE ≤ RMSE for every reported number.

Table C. Out-of-sample RMSE / MAE of the w/o-GIM variant.

SettingRecomputedPreviously in Tables 8-9
METR-LA, random r=0.41.0710 / 0.65841.0633 / 1.1333
METR-LA, fiber l=481.0845 / 0.63871.1021 / 1.2147
METR-LA, block (48,10)1.0647 / 0.62461.1031 / 1.2169
PeMS-Bay, fiber l=480.5291 / 0.26310.5385 / 0.5422

The recomputed RMSE values closely reproduce Table 8, which confirms that only the MAE entries were affected. Table 9 now reports the recomputed MAE values, and its caption notes the correction. We checked every entry of Tables 6-10, and all of them satisfy MAE ≤ RMSE. The conclusion of the ablation is unchanged: removing GIM causes by far the largest degradation of all components.

We are grateful to the reviewer for this careful check, which has improved the accuracy of the paper. We hope these responses address the reviewer's questions, and we would be sincerely grateful if the reviewer would take them into account in the final assessment. We would be glad to provide any further clarification during the discussion period.

AuthorsOpenReview ↗

Kind Reminder on Our Responses

Dear Reviewer r4jE,

Thank you again for your thoughtful feedback! We hope our responses have addressed your concerns. If any concerns remain, we'd be happy to discuss them further. If you feel your concerns have been resolved, we'd appreciate it if you would reconsider your overall score in light of our responses.

Best regards,

The Authors

Reviewer_KV1hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PAST, a traffic time-series imputation network organized by information availability rather than by the conventional temporal/spatial axes. A Graph-Integrated Module (GIM) extracts "primary" (mask-dependent) patterns via mask-adaptive temporal graphs in which observed steps message missing ones, distance-based multi-order spatial propagation over the road network, and an interval-aware edge dropout that removes short observed-to-missing edges with higher probability during training. A Cross-Gated Module (CGM) extracts "auxiliary" (mask-invariant) patterns from timestamp and sensor-identity embeddings via bidirectional gating, injects them into every GIM layer, and is trained to predict the residual of GIM. PAST is evaluated on METR-LA, PeMS-Bay, and LargeST-SD under 27 missing configurations (random, fiber, block) against ten baselines, of which the three most recent (CSDI, ImputeFormer, OPCR) are compared on one representative configuration per missingness type. PAST attains the lowest RMSE in 53 of 54 settings against the seven established baselines and in 17 of 18 against the three recent ones, with the largest gains under fiber and block missingness. The paper’s headline reduction is 36.9% relative to BRITS in one cell; relative to the strongest baseline in that cell (ImputeFormer, 0.5417), the reduction is 36.5%, as Table 1’s relative-change row reports.

Critical strengths

The availability-based decomposition is formalized cleanly, and its effect is substantially isolated by the temporal/spatial-split control. The primary/auxiliary distinction is not merely an architectural choice but a testable hypothesis, and the temporal/spatial-split variant degrades far more under structured gaps (+33.4% fiber, +20.2% block) than under random gaps (+12.4%) while keeping the same components. This control is what lifts the paper from a module combination to a substantiated claim about where robustness to structured missingness comes from, and it is the main reason I lean toward acceptance despite limited component-level novelty. The control is not perfectly clean—it also relocates the spatial operator between branches—but it is substantially more convincing than an ablation that merely removes a module. The empirical study is broad and checkably reported. The evaluation covers three datasets, 27 configurations, both in-sample and out-of-sample evaluation, full per-cell tables, sensitivity analyses, run-to-run variability, and training cost, and explicitly reports the one setting where OPCR beats PAST. Because every cell is tabulated, the aggregate claims can be recomputed, and they check out: the reported average reductions are 8.2%/8.5% (random), 12.8%/18.2% (fiber), and 14.4%/19.9% (block) for in-sample and out-of-sample evaluation, respectively, with average degradations from in-sample to out-of-sample evaluation of 7.7% for PAST and 8.7% for GRIN, and a median run SD of 0.0016. The margin over strong baselines also tends to grow with gap length and spatial span, consistent with the paper’s explanation.

Critical weaknesses

The architectural claim is not tested against the simpler alternative the paper itself names. The related work notes that imputation models typically feed positional/temporal encodings and identity embeddings as additional input features of a single network, yet no such variant is run. The existing ablations do not cover it: w/o CGM removes context entirely, w/o GIM is context-only, CG→LL and G→RNN swap components inside the two-module design, and the T/S split compares two ways of building a two-branch network. These experiments support the availability-based decomposition over the temporal/spatial split, but they do not establish that a dedicated two-module architecture is necessary or superior to simply exposing the same context to a single network. Since the latter is the simpler alternative identified in the related work, it remains an important missing control for isolating the architectural contribution.

(i) The mask-generation procedure is underspecified in the paper—for fiber and block masks, the text gives interval lengths up to l and a fixed rate of 0.4, but not the number of intervals per sensor, how the rate is realized, how sensors are selected, or the seeds—so the masks cannot be regenerated from the paper alone. (ii) The paper does not report multiple independent mask draws per configuration: Table 11’s repeated runs hold the configuration fixed and are not identified as mask resampling, so mask-sampling uncertainty is unquantified. This matters because some win counts rest on near-ties: for block missingness on PeMS-Bay in the in-sample setting, PAST scores 0.3220 against 0.3222 for OPCR, a 0.06% margin inside PAST’s own run SD of 0.0013 for that cell. Moreover, OPCR’s run-to-run variability is not reported, so the uncertainty of the pairwise difference cannot be assessed from Table 11 alone. (iii) The recent strong baselines appear in only 3 of 27 configurations, with fiber represented by l=24 , the shortest gap.

Questions for authors

(1) How does PAST compare with a single-module variant that receives the same context (time-of-day, day-of-week, sensor-ID embeddings) as additional input features—for example, GIM with context concatenated to its inputs and no CGM—at comparable parameter count and compute? This would separate the effect of the two-module organization from the effect of simply exposing calendar and sensor-identity features.

(2) In Equation (2), is YGIM​ detached when constructing the residual target M⊙(X−YGIM​)? Without a stop-gradient, LCGM​ also backpropagates into GIM through this target. Even with a stop-gradient, the target changes as GIM is updated jointly. Could you clarify the optimization procedure and whether “joint residual learning” or “residual parameterization” would therefore be more accurate than “gradient boosting”?

(3) How exactly is the missing rate of 0.4 realized for fiber and block masks—number of intervals per sensor, sensor selection, interval placement, and seeds? Can you also report mean ± SD over at least five independent mask draws for the three representative configurations, ideally including both PAST and the recent baselines?

3 comments
AuthorsOpenReview ↗

Response to Reviewer KV1h (Part 1/2): Single-Network Control

We sincerely thank the reviewer for the exceptionally careful reading, including recomputing our aggregate statistics from the per-cell tables. We greatly appreciate that the reviewer treated the primary/auxiliary distinction as a testable hypothesis, and the suggested single-network control has made this hypothesis considerably sharper. We answer each question below, and all new results are included in the revised PDF (changes in blue).

Q1: "How does PAST compare with a single-module variant that receives the same context (time-of-day, day-of-week, sensor-ID embeddings) as additional input features—for example, GIM with context concatenated to its inputs and no CGM—at comparable parameter count and compute?"

We implemented exactly this control, GIM-ctx. It is GIM without CGM, with the same context embeddings (sensor ID, day-of-week, hour, minute) added to the input projection of every entry. Its width is matched to PAST's parameter count. All variants share an identical training budget on METR-LA.

Table A. RMSE on METR-LA, in-sample / out-of-sample (new Appendix J).

VariantParamsRandom r=0.4Fiber l=48Block (48,5)
PAST287.2K0.2819 / 0.28960.4016 / 0.43220.4017 / 0.4150
GIM-ctx284.7K0.2661 / 0.27700.3833 / 0.45100.3848 / 0.4401
w/o CGM184.7K0.2831 / 0.29060.4753 / 0.49970.4664 / 0.4736

Table B. Increase of RMSE from in-sample to out-of-sample evaluation.

VariantFiber l=48Block (48,5)
PAST+7.6%+3.3%
GIM-ctx+17.7%+14.4%

This control supports the availability-based view in a precise way.

  • Random missingness. Local observations are abundant, and context as an input feature is sufficient: GIM-ctx reaches a 4.3% lower out-of-sample RMSE.
  • Structured missingness. GIM-ctx fits the training period best, but PAST generalizes better to the held-out period: the out-of-sample RMSE of GIM-ctx is 4.3% higher under fiber and 6.0% higher under block missingness, and its RMSE increases from in-sample to out-of-sample evaluation are more than twice (fiber) and four times (block) those of PAST (Table B).
  • Real failures. On the real sensor-failure patterns of METR-LA, the RMSE of GIM-ctx is 34% higher than that of PAST (0.616 vs. 0.458; Table B in our response to F7jZ).

When context enters only as an input feature, it is entangled with the mask-dependent message passing. A separate mask-invariant branch keeps it fully available inside long gaps, and the results show that this is what generalizes beyond the training period. Appendix J now states this distinction explicitly.

AuthorsOpenReview ↗

Response to Reviewer KV1h (Part 2/2): Optimization, Masks and Recent Baselines

Q2: "In Equation (2), is YGIMY_{\rm GIM} detached when constructing the residual target M⊙(X−YGIM)M\odot(X-Y_{\rm GIM})? [...] Could you clarify the optimization procedure and whether "joint residual learning" or "residual parameterization" would therefore be more accurate than "gradient boosting"?"

Yes, YGIMY_{\rm GIM} is detached. The revised Eq. (2) reads LGIM=MSE(M⊙YGIM, M⊙X)\mathcal L_{\rm GIM} = \mathrm{MSE}(M\odot Y_{\rm GIM},\, M\odot X) and LCGM=MSE(M⊙(sg(YGIM)+YCGM), M⊙X)\mathcal L_{\rm CGM} = \mathrm{MSE}(M\odot(\mathrm{sg}(Y_{\rm GIM}) + Y_{\rm CGM}),\, M\odot X), where sg\mathrm{sg} denotes stop-gradient. LCGM\mathcal L_{\rm CGM} therefore updates only the CGM parameters and never back-propagates into GIM. CGM additionally receives gradients from LGIM\mathcal L_{\rm GIM} through the context injected into every GIM layer, which is the intended interaction between the two modules. We agree with the reviewer: because both objectives are minimized jointly and the residual target moves as GIM is updated, "joint residual learning" is the accurate name. Section 4.1 and the introduction now use this name, and Section 4.1 states the exact gradient path.

Q3: "How exactly is the missing rate of 0.4 realized for fiber and block masks—number of intervals per sensor, sensor selection, interval placement, and seeds? Can you also report mean ± SD over at least five independent mask draws [...]?"

The full procedure is in the new Appendix I:

  • Random: each entry is missing independently with probability r.
  • Fiber: for each sensor, segments with lengths drawn uniformly from {1,…,l}\lbrace 1,\dots,l\rbrace are placed at uniformly random start positions until the target rate is reached. Overlapping segments merge, so the realized rate is slightly below the nominal one.
  • Block: each block covers s sensors that are connected on the road graph, with a length drawn uniformly from {1,…,l}\lbrace 1,\dots,l\rbrace and a uniformly random start. Blocks are added until the global rate reaches r.

Table C. Realized mask statistics on METR-LA.

ConfigRealized ratePer-sensor rateMedian gapGaps per sensor
Random r=0.2 / 0.4 / 0.60.199 / 0.400 / 0.600within ±0.021 / 1 / 21825 / 2743 / 2743
Fiber l=240.3700.33-0.4312339
Fiber l=480.3820.32-0.4524180
Fiber l=960.3880.29-0.504892
Block (48,5)0.4000.30-0.4921206
Block (48,10)0.4000.33-0.4716247
Block (96,5)0.4000.27-0.5040106

The exact mask files of all configurations are released with the code, so every result can be reproduced on identical masks.

For mask-sampling variability (new Appendix L), we drew two additional independent masks for each structured configuration and retrained PAST on each under the same budget. The out-of-sample RMSE is 0.4446 ± 0.0031 (fiber l=48) and 0.4396 ± 0.0047 (block (48,5)), so the SD across draws is about 1% of the error. The redrawn fiber masks reach the nominal rate exactly (0.401 vs. 0.382 for the released mask), which accounts for their slightly higher absolute error. Following the reviewer's observation, we mark the PeMS-Bay in-sample block cell (0.3220 vs. 0.3222 for OPCR) as a tie, and the text now reports that PAST is best in 16 of 18 settings against the three recent methods, with one tie.

W(iii): "The recent strong baselines appear in only 3 of 27 configurations, with fiber represented by l=24, the shortest gap."

CSDI, ImputeFormer and OPCR are evaluated on one representative configuration per missingness type because of their computational cost, in particular CSDI's diffusion sampling over 27 configurations and two evaluation settings. The configurations were fixed before any results were obtained: one per type at the common missing rate of 0.4 (random r=0.4, fiber l=24, block (48,10)). As the reviewer observes, the margin of PAST tends to grow with gap length; the shortest fiber gap is therefore the fiber setting least favorable to PAST. Even there, PAST leads all three recent methods out of sample on METR-LA (RMSE 0.3083 vs. 0.4399-0.5303). The full grid of 27 configurations is covered by the seven established baselines, where PAST is best in 53 of 54 settings.

We hope these responses address the reviewer's questions. We would be sincerely grateful if the reviewer would take them into account in the final assessment, and we would be glad to provide any further clarification during the discussion period.

AuthorsOpenReview ↗

Kind Reminder on Our Responses

Dear Reviewer KV1h,

Thank you again for your thoughtful feedback! We hope our responses have addressed your concerns. If any concerns remain, we'd be happy to discuss them further. If you feel your concerns have been resolved, we'd appreciate it if you would reconsider your overall score in light of our responses.

Best regards,

The Authors

Reviewer_F7jZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a data imputation model and algorithm to address the problem of missing data in traffic time series. Specifically, the proposed method consists of two components: a primary module and an auxiliary module. Leveraging the temporal continuity of traffic data, the primary module estimates missing values based on neighboring observations. Meanwhile, the auxiliary module captures typical temporal patterns by considering the periodicity of traffic time series, as well as the node-specific scales and patterns, thereby facilitating data imputation under extensive missing-data scenarios. Experimental results on three publicly available traffic datasets demonstrate that the proposed method consistently outperforms baseline methods.

Critical strengths
  1. This paper is well-written, clearly structured, and easy to understand.

  2. The proposed method is well-motivated, with each carefully designed component addressing specific data characteristics and challenges inherent to the problem.

  3. The experimental evaluation is sufficiently comprehensive, including comparisons with baseline methods across multiple datasets and well-designed ablation studies. The experimental results provide convincing evidence of the rationale and effectiveness of the proposed method.

Critical weaknesses
  1. The paper lacks a sufficiently thorough review of the literature on traffic data imputation. Although the related work section discusses time series imputation in general, traffic data imputation has a well-established and extensive body of research that deserves more comprehensive discussion in the main text.

  2. Due to the insufficient coverage of existing research on traffic time series imputation, the experimental evaluation omits comparisons with several relevant baseline methods specifically developed for this task. This makes it difficult to convincingly demonstrate the superiority of the proposed method over existing approaches.

  3. Regarding the context nodes, the paper incorporates context features as nodes in the graph. However, it does not provide sufficient empirical evidence to demonstrate the effectiveness or necessity of this design choice.

  4. The experimental evaluation can, to some extent, be regarded as simulation-based, as the missing data patterns are artificially generated using simple predefined rules. Such simulated missingness may deviate from the actual missing data patterns observed in real-world traffic systems. The authors are encouraged to consider learning the distribution of missingness from real-world data and then generating more realistic missing data patterns based on the learned distribution.

Questions for authors

Are there any datasets that provide real-world missingness patterns in traffic data? If so, could these patterns be used to construct the test set, rather than the training set, to better evaluate the proposed methodology?

3 comments
AuthorsOpenReview ↗

Response to Reviewer F7jZ (Part 1/2): Real-World Missingness

We sincerely thank the reviewer for the positive assessment of the writing, the motivation and the experiments, and for the concrete and constructive suggestions. We are especially grateful for the suggestion to evaluate on real missingness patterns, which led to one of the most informative experiments in the revision. We answer each comment below, and all new results are included in the revised PDF (changes in blue).

Q1 / W4: "Are there any datasets that provide real-world missingness patterns in traffic data? If so, could these patterns be used to construct the test set, rather than the training set, to better evaluate the proposed methodology?"

Yes. METR-LA itself records real sensor failures as zero readings, and these failures have exactly the structure that our fiber and block settings are designed to model.

Table A. Real failure patterns in METR-LA.

StatisticValue
Share of readings that are failures7.5%
Median failure length5 steps (75 min)
90th percentile of failure length42 steps (10.5 h)
Longest failure560 steps
Time steps at which more than half of the sensors fail together668

Following the reviewer's suggestion, we use these patterns to build the test set. We take the failure pattern of the training period, shift it forward by exactly four weeks onto the test period, and use it as the test mask, so that every real failure keeps its time of day and day of week. Entries whose own ground truth is a failure reading are excluded, which leaves 10,182 evaluated entries. All models are the ones trained under block (48,5) missingness, and nothing is tuned on the real masks.

Table B. Imputation under real-world missingness on the METR-LA test period (new Appendix K).

MethodRMSEMAE
PAST0.4580.309
Injection-only0.4970.368
Residual-only0.5150.354
w/o CGM0.5230.346
GIM-ctx (single network with context inputs)0.6160.403
Linear interpolation0.6570.363
Historical average (time-of-week)0.8370.683

PAST achieves the lowest RMSE and MAE. Its RMSE is 12.4% lower than that of the same model without CGM, 30% lower than linear interpolation, and 45% lower than the time-of-week historical average. The variants rank in the same order as under synthetic fiber and block missingness (see our responses to KV1h and r4jE). The synthetic protocol, which follows the standard settings of GRIN, SPIN and ImputeFormer, is therefore a faithful proxy for real failures, and the conclusions of the paper carry over to them.

On the suggestion to learn the distribution of missingness from real data: using the real failure patterns directly as test masks is, in our view, the most faithful form of this idea, since it requires no modeling assumption about the missingness process. We thank the reviewer for pointing us in this direction.

AuthorsOpenReview ↗

Response to Reviewer F7jZ (Part 2/2): Context Nodes and Traffic-Imputation Literature

W3: "Regarding the context nodes, the paper incorporates context features as nodes in the graph. However, it does not provide sufficient empirical evidence to demonstrate the effectiveness or necessity of this design choice."

We provide two kinds of evidence.

(a) How much the model relies on the context node. In each temporal layer, the context node competes with the observed steps for a row-normalized share of the aggregation weight, so its share measures how much the model draws on context. The share is 0.29 (METR-LA) and 0.34 (PeMS-Bay) under random missingness (r=0.4), and rises to 0.46-0.56 under fiber and block missingness. On LargeST-SD, it is 0.47-0.61 in all settings. The model thus learns to draw on context precisely when local observations are scarce, which is the behavior the design targets.

(b) Removing the context node. We remove only the context node from the temporal graph and keep the CGM residual branch.

Table C. Out-of-sample RMSE on METR-LA with and without the context node.

SettingPASTw/o context nodeChange
Random r=0.40.28960.2908+0.4%
Fiber l=480.43220.4810+11.3%
Block (48,5)0.41500.4640+11.8%
Real-world missingness (Table B)0.4580.515+12.4%

The context node matters little when observations are dense, and becomes essential inside long and wide gaps, including real failures. Both analyses are in the revised PDF (Appendices J and M).

W1 / W2: "The paper lacks a sufficiently thorough review of the literature on traffic data imputation. [...] the experimental evaluation omits comparisons with several relevant baseline methods specifically developed for this task."

We thank the reviewer for this suggestion. Section 2 now includes a dedicated paragraph on traffic data imputation that covers two families:

  • Low-rank matrix and tensor completion: BATF (Chen et al., TR-C 2019), BTMF (Chen & Sun, TPAMI 2022), LRTC-TNN (Chen et al., TR-C 2020) and LATC (Chen et al., IEEE T-ITS 2022). These methods organize traffic data as a sensor × time-of-day × day tensor and exploit its low rank and temporal smoothness. They capture recurring daily and weekly structure globally, but they are transductive: imputing a new period requires solving the completion problem again.
  • Deep traffic imputers: GRIN, STCPA, SPIN, PriSTI and ImputeFormer, which learn spatio-temporal dependencies from data and apply directly to unseen periods.

PAST belongs to the second family, but makes the recurring structure that low-rank methods exploit explicit through calendar and sensor context, which remains available inside long gaps. Our comparison includes the deep traffic imputers GRIN, STCPA and ImputeFormer, and on real failures PAST also clearly outperforms the classical traffic references, linear interpolation and the historical average (Table B).

We hope these responses address the reviewer's concerns. We would be sincerely grateful if the reviewer would take the new results into account in the final assessment, and we would be glad to provide any further clarification during the discussion period.

AuthorsOpenReview ↗

Kind Reminder on Our Responses

Dear Reviewer F7jZ,

Thank you again for your thoughtful feedback! We hope our responses have addressed your concerns. If any concerns remain, we'd be happy to discuss them further. If you feel your concerns have been resolved, we'd appreciate it if you would reconsider your overall score in light of our responses.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of New Results and Revisions

Dear Reviewers and Area Chair,

We sincerely thank all four reviewers for their time and for their careful and constructive reviews. We are encouraged that the reviewers found the availability-based primary/auxiliary decomposition well motivated and supported by the temporal/spatial-split control (KV1h, r4jE, jz2A), the empirical study broad and checkably reported (KV1h, F7jZ, r4jE), and the paper clear and easy to follow (F7jZ, KV1h, r4jE).

We have uploaded a revised PDF, with all changes marked in blue. All new experiments are on METR-LA, and all variants within a comparison are trained under an identical budget. The table below lists how each concern is addressed.

ConcernReviewer(s)What we didWhere in the revision
Real-world missingnessF7jZTest on real sensor-failure patterns from METR-LAAppendix K
Single network with context as inputKV1hParameter-matched GIM-ctx controlAppendix J
Context injection vs. residual correctionr4jEInjection-only and residual-only variantsAppendix J
Evidence for the context nodeF7jZLearned weight share and removal ablationAppendices J and M
Gradient path of the residual objectiveKV1h, jz2AStop-gradient stated; scheme named joint residual learningSection 4.1, Eq. (2)
Target-specific temporal weightingjz2ALearnable weights written out; learned profiles reportedSection 4.2, Eq. (3); Appendix M
Transfer of the residual to missing entriesjz2ATraining with artificially hidden entries describedSection 4.4
Mask generation and sampling variabilityKV1hFull procedure, realized rates, independent redrawsAppendices I and L
Traffic-imputation literatureF7jZNew related-work paragraphSection 2
MAE > RMSE entriesr4jEFour entries recomputed; all tables checkedTable 9
Representative settings of recent baselinesKV1h, jz2ASelection criterion statedSection 5.1
Dynamic exogenous signalsjz2ALimitations paragraphSection 6

The main findings of the new experiments are:

  1. Real failures. On real sensor-failure patterns, PAST achieves the lowest error: RMSE 0.458, versus 0.523 without CGM, 0.616 for a single network with context inputs, and 0.657 for linear interpolation. The variants rank in the same order as under synthetic fiber and block missingness.
  2. Two-module organization. Context as an input feature of a single network is sufficient under random missingness. Under fiber and block missingness, PAST generalizes better to the held-out period (out-of-sample RMSE 0.4322 vs. 0.4510 and 0.4150 vs. 0.4401), and on real failures the RMSE of the single network is 34% higher.
  3. Inside CGM. Context injection provides most of the gain under structured missingness: removing it raises out-of-sample RMSE by 11.3% (fiber) and 11.8% (block). The residual correction adds a further 1.4% and 3.6%.
  4. Temporal weights. The weights are learned per target and decay with temporal distance: 84% of each row's weight lies within one hour, versus 8.8% under uniform weighting.
  5. Mask sampling. Across independent mask draws, the SD of out-of-sample RMSE is about 1% of the error.

We are grateful for the reviewers' suggestions, which have clearly strengthened the paper, and we warmly welcome any further questions during the discussion period.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.