acceptodds
Under review as a conference paper at ICLR 2027

ARCI-T: ANCHOR-RESIDUAL CONTROLLED IMPUTATION FOR MISSING TRAFFIC SENSOR DATA

Abstract

Missing values in traffc sensor records may manifest as isolated points, consecutive blocks, or sensor-level outages, yielding heterogeneous temporal and spatial evidence across different missing positions. Deterministic recovery provides a stable reference, while learned models can capture richer spatiotemporal dynamics. Rather than treating these two solutions as competing alternatives, we formulate traffc imputation as controlled correction around a retained deterministic reference. We formalize Anchor-Residual Controlled Imputation for Missing Traffc Sensor Data (ARCI-T), which employs the deterministic recovery as an explicit Anchor, learns an anchor-relative correction, and determines the contribution magnitude of this correction at each missing position based on available observational evidence. This design preserves the deterministic reference as a fallback while enabling learned corrections to adapt under heterogeneous missingness conditions. Across four traffc benchmarks under random and structured missingness protocols, ARCI-T achieves state-of-the-art or highly competitive imputation accuracy, demonstrating that controlled correction is an effective approach for traffic sensor imputation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Imputation, Spatiotemporal Learning, Missing Data Traffic Data Imputation Residual Learning, Adaptive Gating

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yWykOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ARCI-T, a traffic imputation method that keeps a deterministic recovery (LinearFill) as an explicit anchor and learns a residual correction on top of it. Experiments on four traffic datasets under random and structured missingness show the lowest MAE in almost all settings.

Critical strengths
  • Simple, interpretable formulation with a guaranteed fallback to the deterministic reference. I think this approach is very practical.
  • Consistent improvements across four datasets and both random and structured missingness.
  • Extensive ablation studies are provided
Critical weaknesses
  • According to the ablation studies, the gains seem driven mostly by the spatial component rather than the proposed mechanism. Please justify the motivation of the paper based on the experimental results.
  • Standard graph-based imputation baselines (GRIN, PriSTI) are cited but not compared. I think these methods should be included as baselines.
  • Only MAE is reported in the main tables. It would be better to report other metrics such as RMSE and MRE.
Questions for authors

Please see weakness above.

4 comments
AuthorsOpenReview ↗

W1 | Role of the Spatial Component vs. Contribution of the Proposed Acceptance Mechanism

Thank you for recognizing the practical utility of our approach and the effectiveness of our deterministic restoration reference. Below, we address your comments regarding the roles of the spatial component versus our proposed acceptance mechanism, graph-based imputation baselines, and multi-metric evaluation:

W1 | Role of the Spatial Component vs. Contribution of the Proposed Acceptance Mechanism

Thank you for pointing out that in our ablation study, a large portion of the performance gain comes from the spatial component. We agree with this observation, and we realize that our original manuscript did not explain clearly enough how the spatial modeling component and our proposed controlled-correction mechanism work together. To make our explanation self-contained without requiring a re-read of the paper: ARCI-T operates in two stages around a deterministic anchor (e.g., linear interpolation). First, the backbone extracts temporal features and cross-sensor spatial features—where the spatial component is implemented by our correlation-graph module (CorrTopK), which aggregates information from the top-KK correlated neighboring sensors—to predict a candidate residual correction relative to the anchor. Second, our proposed position-conditioned acceptance mechanism (the gating branch) evaluates local observational evidence to decide what fraction (G∈[0,1]G \in [0,1]) of that candidate correction should actually be applied at each missing position. To directly test whether our proposed acceptance mechanism still provides meaningful gains when the spatial component is held constant, we trained an ungated model (G1, which keeps the exact same CorrTopK spatial component but fixes G≡1G \equiv 1) from scratch. The results confirm that the spatial modeling component and the proposed acceptance mechanism play complementary roles, and that adding our acceptance mechanism consistently outperforms G1 across all 20/20 paired seed runs.

Specifically, the spatial component (CorrTopK) is indeed an essential source of information in ARCI-T—especially during long temporal gaps or whole-sensor outages where a target sensor has no nearby temporal observations of its own and must rely on neighboring sensors to reconstruct traffic states. However, our proposed acceptance mechanism is not intended to replace spatial modeling; instead, the spatial modeling component and the proposed acceptance mechanism address two different questions in our pipeline: the spatial component answers "what candidate correction can be proposed from neighboring sensors," whereas the proposed acceptance mechanism answers "how much of that proposed correction should be accepted and applied at each specific position."

In our original ablation table, the w/o CorrTopK variant removed the spatial component entirely. Because both the residual proposal head and the acceptance gate take the upstream spatial features produced by CorrTopK as input, removing CorrTopK degraded both branches at once and prevented the model from proposing informed spatial corrections in the first place. As a result, that ablation showed the necessity of spatial information, but could not isolate the standalone contribution of the acceptance mechanism. To properly separate the effect of our proposed acceptance mechanism from the spatial component, we trained the ungated control model G1 (G≡1G \equiv 1) from scratch under our unified mechanism-analysis setup (fixed 100-epoch training selected by validation loss), keeping the exact same deterministic anchor, CorrTopK spatial component, and residual proposal network while removing position-conditioned gating.

Table R8 | Controlled Comparison of the Proposed Acceptance Mechanism under an Identical Spatial Component (Trained from Scratch)

DatasetProtocolFull MAEG1 MAEMAE Reduction
PeMS-BAYMixed251.53241.5503+1.15%
PeMS-BAYSensor251.69301.7432+2.88%
PEMS04Mixed2521.731122.2097+2.16%
PEMS04Sensor2523.278123.8053+2.21%

As shown in Table R8, across all four matched settings (5 random seeds per setting, 20 paired runs in total), the full ARCI-T model consistently outperforms G1 (1.5324 vs. 1.5503 on PeMS-BAY Mixed25; 1.6930 vs. 1.7432 on PeMS-BAY Sensor25; 21.7311 vs. 22.2097 on PEMS04 Mixed25; and 23.2781 vs. 23.8053 on PEMS04 Sensor25). In short, the spatial component supplies cross-sensor context to form strong correction candidates, while our proposed acceptance mechanism prevents applying full corrections blindly at positions where the local anchor is already accurate or where spatial signals are noisy. We will revise the motivation and ablation discussions in the manuscript to make this relationship clear.

AuthorsOpenReview ↗

W2 | Complete Five-Seed Comparison with GRIN

We agree that evaluating graph-based baselines on only a subset of missingness protocols gives an incomplete picture, and that a full multi-seed comparison against standard graph imputation models is necessary to properly evaluate our method against graph-based approaches. Accordingly, we completed all 80 evaluations (4 datasets×4 missingness protocols×5 seeds4\text{ datasets} \times 4\text{ missingness protocols} \times 5\text{ seeds}) for GRIN, showing that ARCI-T achieves lower mean MAE in 14/16 settings—including all 8/8 structured missingness settings (Mixed25 and Sensor25)—while also explaining the hardware memory constraints encountered with PriSTI.

The complete 5-seed MAE results across all 16 settings are presented in Table R10(a), where relative error reduction is computed as (GRIN−ARCI-T)/GRIN(\text{GRIN} - \text{ARCI-T}) / \text{GRIN} (positive values indicate lower MAE for ARCI-T).

Table R10(a) | GRIN Across All 16 Settings: MAE Mean ±\pm Sample Standard Deviation

DatasetProtocolARCI-TGRINARCI-T Error Reduction
METR-LAER252.1181±0.27782.1181 \pm 0.27782.1003±0.01092.1003 \pm 0.0109−0.85%-0.85\%
METR-LAER602.4351±0.32962.4351 \pm 0.32962.4119±0.02782.4119 \pm 0.0278−0.96%-0.96\%
METR-LAMixed253.2387±0.01453.2387 \pm 0.01453.6397±0.07163.6397 \pm 0.0716+11.02%
METR-LASensor253.4722±0.06523.4722 \pm 0.06523.9598±0.05763.9598 \pm 0.0576+12.31%
PeMS-BAYER250.6867±0.00110.6867 \pm 0.00110.7321±0.00150.7321 \pm 0.0015+6.20%
PeMS-BAYER600.8648±0.00200.8648 \pm 0.00200.9063±0.00230.9063 \pm 0.0023+4.58%
PeMS-BAYMixed251.4817±0.01481.4817 \pm 0.01481.5369±0.01371.5369 \pm 0.0137+3.59%
PeMS-BAYSensor251.6458±0.01851.6458 \pm 0.01851.7013±0.00701.7013 \pm 0.0070+3.27%
PEMS04ER2515.8302±0.025815.8302 \pm 0.025816.0229±0.029716.0229 \pm 0.0297+1.20%
PEMS04ER6016.7568±0.040216.7568 \pm 0.040216.9366±0.014516.9366 \pm 0.0145+1.06%
PEMS04Mixed2520.6271±0.186020.6271 \pm 0.186021.3815±0.237221.3815 \pm 0.2372+3.53%
PEMS04Sensor2521.9462±0.131621.9462 \pm 0.131622.5430±0.197222.5430 \pm 0.1972+2.65%
PEMS08ER2511.4385±0.033911.4385 \pm 0.033912.1455±0.048212.1455 \pm 0.0482+5.82%
PEMS08ER6012.4842±0.052312.4842 \pm 0.052313.0291±0.063213.0291 \pm 0.0632+4.18%
PEMS08Mixed2516.2881±0.177316.2881 \pm 0.177317.0377±0.097917.0377 \pm 0.0979+4.40%
PEMS08Sensor2517.2828±0.066017.2828 \pm 0.066018.0706±0.207018.0706 \pm 0.2070+4.36%

Across all 16 settings, the only two exceptions are METR-LA ER25 and ER60, where GRIN's mean MAE is slightly lower by 0.0178 and 0.0232, respectively; across all structured missingness settings (Mixed25 and Sensor25), ARCI-T achieves consistent gains of up to +11.02% and +12.31% on METR-LA.

Note on PriSTI: Regarding the conditional diffusion model PriSTI that you also mentioned, we tried to include it in a full benchmark comparison; however, because global spatiotemporal attention combined with multi-step iterative diffusion denoising has quadratic memory overhead across 300+ graph nodes, running PriSTI across all 80 evaluations (16×516 \times 5) frequently triggers out-of-memory (OOM) errors and requires compute time beyond our laboratory GPU capacity. To ensure broad baseline coverage, alongside our complete GRIN evaluation, we included diffusion/generative and recent spatiotemporal imputation models such as FENCE, GiFlow, and StamImputer in our experiments and discussions, and we will add the complete GRIN results as well as a dedicated discussion of PriSTI's architecture and computational cost to the revised manuscript.

AuthorsOpenReview ↗

W3 | Multi-Metric Evaluation: MAE, RMSE, MRE, and MSE

W3 | Multi-Metric Evaluation: MAE, RMSE, MRE, and MSE

We agree with your suggestion: relying on MAE alone can hide how a model handles large outlier errors and does not fully reflect reconstruction quality across different error metrics. To provide a more complete evaluation, we compiled 5-seed average results for MAE, RMSE, MRE, and MSE across all 16 primary settings, showing that ARCI-T achieves lower mean MAE and MRE in 15/16 settings and lower mean RMSE and MSE in 16/16 settings—reducing RMSE by 6.62% and MSE by 12.82% even on PEMS04 Sensor25 where MAEs are similar.

Table R11 | Multi-Metric Results Across All 16 Original Benchmark Settings (ARCI-T / Baseline per Cell)

DatasetProtocolOriginal Benchmark Strongest BaselineMAERMSEMREMSE
METR-LAER25OPCR2.1181 / 2.32874.3029 / 4.53380.0418 / 0.045918.9723 / 20.5571
METR-LAER60OPCR2.4351 / 2.71425.1519 / 5.59560.0480 / 0.053527.1056 / 31.3145
METR-LAMixed25OPCR3.2387 / 4.27847.0348 / 9.61350.0642 / 0.084849.4899 / 92.5633
METR-LASensor25BRITS3.4722 / 4.45747.4073 / 9.52450.0689 / 0.088554.8948 / 90.7180
PeMS-BAYER25OPCR0.6867 / 0.77811.3477 / 1.54400.0110 / 0.01251.8164 / 2.3849
PeMS-BAYER60Linear0.8648 / 0.97731.7575 / 1.97500.0138 / 0.01563.0889 / 3.9009
PeMS-BAYMixed25OPCR1.4817 / 1.80813.2686 / 4.05000.0237 / 0.029010.6852 / 16.4064
PeMS-BAYSensor25OPCR1.6458 / 1.98983.5584 / 4.31820.0264 / 0.031912.6665 / 18.6546
PEMS04ER25ATTST15.8302 / 16.690826.0394 / 26.87860.0742 / 0.0782678.0570 / 722.4748
PEMS04ER60Linear16.7568 / 18.766927.4088 / 30.68540.0786 / 0.0880751.2475 / 941.5973
PEMS04Mixed25BRITS20.6271 / 21.786132.6985 / 37.12100.0973 / 0.10281069.3093 / 1378.6993
PEMS04Sensor25BRITS21.9462 / 21.862134.6281 / 37.08420.1036 / 0.10321199.2053 / 1375.5825
PEMS08ER25ATTST11.4385 / 13.001918.8336 / 20.15100.0499 / 0.0567354.7077 / 406.0781
PEMS08ER60Linear12.4842 / 14.406620.1022 / 22.46870.0544 / 0.0628404.1057 / 504.8494
PEMS08Mixed25BRITS16.2881 / 18.528826.0279 / 31.95340.0715 / 0.0813677.7251 / 1021.3238
PEMS08Sensor25BRITS17.2828 / 18.919327.3970 / 32.75370.0757 / 0.0829750.6318 / 1073.1410

To compare multiple error metrics on the exact same method pairs, Table R11 keeps the strongest external baseline identified for each setting in the original benchmark (while comparisons with the newly added GRIN baseline are presented separately in Table R10(a)). PEMS04 Sensor25 clearly shows the value of multi-metric evaluation: while ARCI-T and BRITS have very similar 5-seed mean MAEs (21.9462 vs. 21.8621), ARCI-T reduces RMSE by 6.62% (34.6281 vs. 37.0842) and MSE by 12.82% (1199.2053 vs. 1375.5825), showing much stronger suppression of large outlier errors. Reporting MAE alone would miss these gains in squared error metrics. We will include these complete multi-metric results and analyses in the revised manuscript.

AuthorsOpenReview ↗

Manuscript Revision Guide for Reviewer yWyk

Manuscript Revision Guide for Reviewer yWyk

Thank you again for your valuable feedback. Following up on the experimental results reported in our response above, we have uploaded the revised manuscript (with all additions highlighted in blue). Below is the exact mapping of your three concerns to the sections, pages, and tables added or updated in the revised PDF:

  • Regarding W1 (Clarifying the role of the spatial component CorrTopK and isolating the incremental contribution of the proposed acceptance mechanism):

    • Section 4.3 (Page 9, Lines 432–452, New Table 7) & Appendix C.7 (Page 25, Lines 1345–1349): Clarified in both the main text and appendix that bypassing CorrTopK (w/o CorrTopK) alters the upstream spatial representation consumed by both the residual proposal head and the acceptance gate, and introduced the independently trained ungated variant G1 (G≡1G \equiv 1 fixed throughout training and evaluation), which retains the exact same anchor and CorrTopK spatial backbone.
    • Appendix C.1 (Page 22, Lines 1178–1179; Page 23, Table 15) & Appendix C.5 (Page 24, Lines 1286–1295; Page 25, Lines 1313–1330, New Table 20): Added the formal definition of G1 (distinguishing it from the inference-only All-R intervention) and Table 20, reporting five-seed paired differences (Δ\Delta), ordinary 95% confidence intervals, and unadjusted two-sided paired tt-test pp-values confirming that full ARCI-T outperforms from-scratch G1 across all 20/20 paired seed runs (raw p < 0.01 across all four settings, remaining significant at α=0.05\alpha = 0.05 under a Holm step-down check).
  • Regarding W2 (Adding full benchmark comparisons with graph-based imputation baselines GRIN and PriSTI):

    • Section 4.1 (Page 6, Lines 272–276) & Section 4.2 (Page 6, Lines 318–321): Integrated the 80-run (4 datasets×4 protocols×5 seeds4\text{ datasets} \times 4\text{ protocols} \times 5\text{ seeds}) comparison with GRIN (Cini et al., 2022) into the main text, highlighting that ARCI-T achieves lower mean MAE in 14 of 16 settings and leads in all 8/8 structured missingness settings (Mixed25 and Sensor25).
    • Appendix B.10 (Page 19, Line 1012; Page 21, Lines 1114–1116): Added full implementation and training details for GRIN.
    • Appendix D.1 (Page 32, Updated Table 33; Pages 32–33, Updated Tables 35–38; Page 34, New Table 40): Updated the best-external summary (Table 33) and the complete protocol-level MAE tables for all four datasets (Tables 35–38) to include GRIN, and added Table 40 reporting GRIN's five-seed MAE, RMSE, and MRE across all 16 settings.
    • Status of PriSTI: As noted in our response, PriSTI (cited in Section 1, Page 1, Line 42) relies on global spatiotemporal attention combined with multi-step iterative diffusion sampling, which incurs prohibitive quadratic memory and runtime overhead on 300+ node graphs and could not be scaled to the full 80-run benchmark campaign under our GPU budget; therefore, we completed the full 80-run graph-imputer evaluation with GRIN while covering diffusion/generative baselines via CSDI, GiFlow, and FENCE.
  • Regarding W3 (Reporting additional evaluation metrics beyond MAE—RMSE, MRE, and MSE):

    • Section 4.1 (Page 6, Lines 272–276), Section 4.2 (Page 6, Lines 321–323) & Appendix D / D.1 (Page 31, Lines 1622–1627, 1650–1655): Added formal definitions and main-text pointers for the secondary metrics (RMSE, MRE, and MSE), noting that against the fixed strongest comparator from the original benchmark, ARCI-T achieves lower mean MAE and MRE in 15/16 settings and lower mean RMSE and MSE in all 16/16 settings (distinct from the expanded GRIN pool).
    • Appendix D.1 (Page 32, Table 34; Page 34, New Table 40; Page 35, New Table 41 & Table 42): Added Table 41 (Page 35), providing the complete five-seed MAE, RMSE, MRE, and MSE comparison between ARCI-T and the fixed strongest original baseline in every setting (including PEMS04 Sensor25, where ARCI-T lowers RMSE by 6.62% and MSE by 12.82%), alongside multi-metric tables for GRIN (Table 40) and DCCN-SPF/FENCE (Table 42).
Reviewer_4XrBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper aim to solve imputation of missing traffic-sensor data under heterogeneous missingness (isolated points, contiguous blocks, sensor-level outages), where the evidence available near each missing position differs. ARCI-T formulates imputation as *controlled correction around a retained deterministic reference. Validation covers four benchmarks (METR-LA, PeMS-BAY, PEMS04, PEMS08) under four missingness protocols (ER25/ER60/Mixed25/Sensor25) with five seeds (80 runs) against 13+ baselines.

Critical strengths

1.A clean, explicit conceptual framework with an ablation design that directly tests the claimed mechanism. The separation of correction content (ARP) from bounded, position-conditioned acceptance (PBRA), with the deterministic reference retained in the output, is clearly differentiated from FENCE (guidance scale of a generative process), DCCN-SPF (fixed S+T), OPCR (pointwise confidence propagation), and SAITS (attention weighting). 2.Thorough, well-controlled, and unusually carefully reported validation. 80 dataset–protocol–seed runs with matched baseline configurations, per-component ablations, paired t-based 95% CIs in appendix analyses, stress tests under heavier missingness and temporal distance on frozen checkpoints, and a serious reproducibility apparatus (source commits, mask SHA-256 hashes, seed conventions, parameter counts).

Critical weaknesses

Each ingredient — residual parameterization around an initial estimate (DCCN-SPF), per-position adaptive weighting (SAITS), pointwise confidence propagation (OPCR), guidance-strength control (FENCE), and sparse sensor-correlation graphs — already exists in prior work. The paper's contribution is a clean factorization and a rigorous validation of it, rather than a new mechanism or a deeper understanding.

Questions for authors

1.Are the main-benchmark wins over the best external baselines significant under paired tests across the five seeds?

2.A per-position analysis of G (its relation to gap length, distance to support, sensor identity, time-of-day) would substantially strengthen the "position-conditioned control" claim beyond the coarse protocol-level geometry summary.

8 comments

post comments below your review discussing if your (human-generated) assessment of the paper differs from any AI tools you may have used. The differences between my reviews and those generated by AI include the article ratings, the actual strengths and weaknesses of the articles, and the authors' questions all of which show clear differences.

The following are the review comments from the Direct AI tool: Paper Summary The paper addresses imputation of missing traffic-sensor data under heterogeneous missingness (isolated points, contiguous blocks, sensor-level outages), where the evidence available near each missing position differs. Rather than treating deterministic recovery and learned models as competing alternatives, ARCI-T formulates imputation as controlled correction around a retained deterministic reference: (i) a deterministic anchor A (LinearFill by default) is computed and kept in the final output; (ii) ARP (Anchor-Relative Residual Parameterization) learns a residual proposal R relative to A, using a multi-scale trend/fluctuation temporal encoder, a lightweight contrast state, and a training-set-frozen sparse sensor-correlation graph (CorrTopK); (iii) PBRA (Position-Conditioned Bounded Residual Acceptance) produces a per-position acceptance coefficient G ∈ [0,1] from temporal, spatial, missingness-geometry (9-dim mask descriptor q), and robust–light evidence, with the acceptance path not reading R or A+R. The final estimate is X = A + G⊙R, with observed values clamped. Training uses a Huber loss on hidden positions plus an auxiliary adjacent temporal-difference loss. Validation covers four benchmarks (METR-LA, PeMS-BAY, PEMS04, PEMS08) under four missingness protocols (ER25/ER60/Mixed25/Sensor25) with five seeds (80 runs) against 13+ baselines. ARCI-T attains the lowest MAE in 15/16 dataset–protocol settings (the sole inversion is PEMS04 Sensor25, where BRITS is marginally better). Ablations examine anchor-relative vs. direct parameterization, three deterministic anchors, gate variants (Anchor, All-R, Const-G, Shuffle-G, Generic-G, Self-G, Global-G, Shared-E), and removal of q, the output anchor, and CorrTopK. Stress tests (Entry40/Mixed40/Sensor40) and T1–T4 temporal-distance bins on frozen checkpoints, plus a detailed reproducibility appendix (commits, hashes, masks, parameter counts), round out the study. Critical advatanges:Innovation in this submission is incremental. Rather than inventing new operational mechanisms or uncovering deeper theoretical understanding, the paper’s contribution consists of a tidy factorized formulation and its systematic validation. The mechanistic study on what the acceptance coefficient encodes is shallow, restricted only to geometry at the protocol level. The Gate+R ablation further reveals that architecturally separating proposal and acceptance does not produce measurable gains. Effects are weak, with inconsistent signs across experimental environments. The central acceptance question for ICLR is whether the idea of fixed deterministic reference plus bounded residual acceptance represents meaningful conceptual advancement. Critical Weaknesses:The practical significance of the empirical edge is moderate and somewhat diluted by the baseline set. Several baselines collapse under structured missingness (ATTST, GiFlow), so "wins in almost all settings" is weaker than it first appears; against the strongest matched baseline per setting, gains are solid but moderate (5–24%). Notably, in the hardest stress regime the paper itself highlights (Sensor40), the position-conditioned gate adds essentially nothing over the shuffled gate (ARCI-T ≈ Shuffle-G on both datasets), and the main tables report MAE only, with MSE relegated to the appendix.Questions For Authors

  1. Statistical testing: Some ARCI-T cells carry large variance (e.g., METR-LA ER25: 2.1181±0.2778 vs. 2.3287±0.0091), so paired-seed analysis would materially strengthen the headline claims.
  2. Sensor40 behavior: Under 40% whole-horizon sensor outages, ARCI-T and Shuffle-G are essentially tied. Is this expected — i.e., position-specific evidence is uninformative when the entire horizon of a sensor is missing? How should readers reconcile this with the paper's framing that controlled correction matters most precisely where learned correction is most needed?
  3. Anchor dependence: Multi-anchor results are close for LinearFill vs. NearestFill but clearly worse for SeasonalProfile (e.g., PEMS04 Mixed25: 20.63 vs. 21.46). Is the framework's benefit anchor-dependent in expectation, and is there guidance on anchor selection in practice?
  4. Efficiency: Beyond parameter counts (≈123k trainable), are there runtime or memory comparisons with the baselines? Training at 100 epochs/batch 8 and the inference cost would help assess practical adoption.
AuthorsOpenReview ↗

Q1 | Statistical Significance of the Five-Seed Advantage

Thank you for your generous assessment of our conceptual framework, ablation design, and reproducibility efforts. Your questions focus on the key empirical and mechanistic aspects of our study: whether our 5-seed gains are statistically significant under multiple-testing correction, how the gating coefficient GG varies at the position level across gap geometry, sensors, and time of day, and whether the full acceptance branch adds measurable value over an ungated baseline. We present our findings below:

Q1 | Statistical Significance of the Five-Seed Advantage

We appreciate your focus on statistical reliability. Reporting 5-seed mean error reductions across multiple benchmarks without formal hypothesis testing or multiple-testing control leaves open whether observed improvements could stem from random seed variance or false positives. To ensure statistical rigor, we conducted two-sided paired tt-tests across all 16 "Dataset ×\times Missingness Protocol" settings in our original benchmark and applied the Holm step-down multiple-comparison correction, confirming that ARCI-T achieves lower mean MAE in 15/16 settings and retains statistically significant superiority (p<0.05p < 0.05) in 13/16 settings after correction.

In our testing procedure, ARCI-T was paired seed-for-seed (n=5n = 5 paired seeds per setting) against the strongest external baseline from the original benchmark. All 16 raw pp-values were ordered from smallest to largest and evaluated against Holm-adjusted thresholds for the family size (m=16m = 16). Table R4 summarizes the paired comparisons at α=0.05\alpha = 0.05.

Table R4 | Paired MAE Comparisons and Holm-Adjusted Significance Across the 16 Original Benchmark Settings

DatasetProtocolOriginal Benchmark Strongest BaselineARCI-T MAEBaseline MAEHolm-Adjusted (α=0.05\alpha = 0.05)
METR-LAER25OPCR2.11812.3287n.s.
METR-LAER60OPCR2.43512.7142n.s.
METR-LAMixed25OPCR3.23874.2784Sig. (p<0.05p < 0.05)
METR-LASensor25BRITS3.47224.4574Sig. (p<0.05p < 0.05)
PeMS-BAYER25OPCR0.68670.7781Sig. (p<0.05p < 0.05)
PeMS-BAYER60Linear0.86480.9773Sig. (p<0.05p < 0.05)
PeMS-BAYMixed25OPCR1.48171.8081Sig. (p<0.05p < 0.05)
PeMS-BAYSensor25OPCR1.64581.9898Sig. (p<0.05p < 0.05)
PEMS04ER25ATTST15.830216.6908Sig. (p<0.05p < 0.05)
PEMS04ER60Linear16.756818.7669Sig. (p<0.05p < 0.05)
PEMS04Mixed25BRITS20.627121.7861Sig. (p<0.05p < 0.05)
PEMS04Sensor25BRITS21.946221.8621n.s.
PEMS08ER25ATTST11.438513.0019Sig. (p<0.05p < 0.05)
PEMS08ER60Linear12.484214.4066Sig. (p<0.05p < 0.05)
PEMS08Mixed25BRITS16.288118.5288Sig. (p<0.05p < 0.05)
PEMS08Sensor25BRITS17.282818.9193Sig. (p<0.05p < 0.05)

As shown in Table R4, across all 16 fixed-baseline paired tests, 13 settings remain statistically significant after Holm multiple-comparison correction, with only METR-LA ER25, METR-LA ER60, and PEMS04 Sensor25 not reaching significance. We will include these paired statistical test results in the revised manuscript.

AuthorsOpenReview ↗

Q2 | Does Vary Across Gaps, Sensors, and Time of Day?

Q2 | Does GG Vary Across Gaps, Sensors, and Time of Day?

We agree that overall average metrics alone do not show how the gating mechanism works internally, and that checking whether GG adapts across missingness geometries, different sensors, and times of day—rather than staying a fixed constant—is important for verifying our design hypothesis. Following your suggestion, we analyzed both GG and the actual applied correction magnitude across 10,906,313 missing entries from 20 frozen models, finding that mean GG rises consistently from 0.082–0.2200.082\text{--}0.220 at gap boundaries (1 step) to 0.434–0.4660.434\text{--}0.466 deep inside gaps (16–31 steps), while also adapting across individual sensors and hours of the day.

Here, "steps" measure the number of 5-minute sampling intervals separating a missing entry from the nearest valid observation on the same sensor: "1 step" (5 minutes) denotes gap boundaries right next to a known observation where local temporal interpolation is reliable, whereas "16–31 steps" (~1.3–2.6 hours) denotes positions deep inside a gap far from any valid temporal observation. In addition, we report both the dimensionless gate ratio GG and the actual applied correction magnitude ∣G⊙R∣|G \odot R| in original data units, because GG measures the accepted share of the proposed residual whereas the actual correction magnitude reflects the absolute adjustment made to the deterministic reference. In this analysis, entries were grouped by position within each seed and averaged with equal weight across the 5 seeds.

Table R5 | Acceptance Coefficient GG and Actual Correction Magnitude from 1 Step to 16–31 Steps

Dataset & ProtocolMean GG: 1 step →\rightarrow 16–31 stepsMean Actual Correction Magnitude: 1 step →\rightarrow 16–31 steps
PeMS-BAY Mixed250.082→0.4350.082 \rightarrow 0.4350.201→2.8960.201 \rightarrow 2.896
PeMS-BAY Sensor250.125→0.4340.125 \rightarrow 0.4340.397→2.9140.397 \rightarrow 2.914
PEMS04 Mixed250.128→0.4660.128 \rightarrow 0.4668.204→30.8408.204 \rightarrow 30.840
PEMS04 Sensor250.220→0.4600.220 \rightarrow 0.4609.748→32.0519.748 \rightarrow 32.051

Across all 20 models, both mean GG and mean actual correction magnitude increase substantially from 1 step (gap boundary) to 16–31 steps (deep gap interior). On PeMS-BAY Mixed25, for example, mean GG rises from 0.082 to 0.435, while the applied correction magnitude grows from 0.201 to 2.896. Grouping by total contiguous gap length shows the same upward trend: for contiguous gap lengths of 1, 16–31, 32–63, and 64–127 steps under the same setting, mean GG increases to 0.063, 0.278, 0.353, and 0.453, respectively. Thus, where local temporal references are weaker, the model systematically accepts a larger share of the learned spatiotemporal correction.

Table R6 | Distribution of Mean GG Across Sensors and Hours of the Day

Dataset & Protocol# SensorsPer-Sensor Mean GG (10th–90th Percentile)24-Hour Mean GG (Min–Max)
PeMS-BAY Mixed253250.267–0.3210.262–0.336
PeMS-BAY Sensor253250.304–0.3670.270–0.389
PEMS04 Mixed253070.312–0.3590.169–0.433
PEMS04 Sensor253070.363–0.4040.174–0.490

Table R6 summarizes the 10th–90th percentile range of per-sensor mean GG across all sensors, along with the minimum-to-maximum range of hourly mean GG across the 24 hours of the day. These results confirm that GG is not a static constant applied equally across all missing entries, but adjusts dynamically to local missingness geometry, sensor characteristics, and time-of-day traffic patterns. We view GG as an adaptive correction ratio rather than a calibrated probability or confidence score, and we will include these position-level analyses in the revised manuscript.

Reviewer_TS2uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ARCI-T for traffic sensor imputation. A deterministic recovery (LinearFill) is kept as an explicit anchor A, a network predicts an anchor-relative residual R, and a bounded per-position coefficient G ∈ [0,1], computed from temporal, spatial and mask-geometry evidence, decides how much of R is applied, giving A + G ⊙ R. The method is validated on METR-LA, PeMS-BAY, PEMS04 and PEMS08 under four synthetic missingness protocols against 13 baselines, with ablations and stress tests, and reports the lowest MAE in 15 of 16 dataset–protocol settings.

Critical strengths
  1. No comparison with methods that follow the same route. The experiments compare ARCI-T with purely learned imputers and with deterministic recovery. They do not compare it with methods that also refine a base estimate with a learned correction, although the paper discusses several of them in the introduction and in Table 8: DCCN-SPF (linear interpolation plus a learned deviation), RDPI (diffusion refinement of an initial imputation) and FENCE (adaptive guidance strength). This is consequential because the paper’s central claim concerns how a learned correction is combined with a deterministic reference; without such a comparison, the gains cannot be attributed to this specific design rather than to the general “base estimate + learned correction” strategy. A brief discussion of related classical ideas, such as Kalman filtering and regression-kriging, would also help position the contribution.
  2. All missingness is synthetic. The method is motivated by real sensor outages and long gaps (Section 1, Section 2.2). However, the datasets are used in versions without natural missing entries (Appendix B.10), and all evaluated missingness is generated by the authors’ masks (Appendix B.7). In the main protocols, the longest outages are blocks of 12–48 steps (Sensor25), and whole-sensor outages appear only in a stress test (Sensor40). This is consequential because the position-conditioned design targets exactly these real outage patterns, so it remains open whether the advantage carries over. I suggest adding an experiment on real sensor data with naturally occurring gaps, or with longer outages in the main protocols; if this is not feasible, a discussion of why the synthetic protocols are representative of real outages would also help.
Critical weaknesses
  1. No comparison with methods that follow the same route. The experiments compare ARCI-T with purely learned imputers and with deterministic recovery. They do not compare it with methods that also refine a base estimate with a learned correction, although the paper discusses several of them in the introduction and in Table 8: DCCN-SPF (linear interpolation plus a learned deviation), RDPI (diffusion refinement of an initial imputation) and FENCE (adaptive guidance strength). This is consequential because the paper’s central claim concerns how a learned correction is combined with a deterministic reference; without such a comparison, the gains cannot be attributed to this specific design rather than to the general “base estimate + learned correction” strategy. A brief discussion of related classical ideas, such as Kalman filtering and regression-kriging, would also help position the contribution.
  2. All missingness is synthetic. The method is motivated by real sensor outages and long gaps (Section 1, Section 2.2). However, the datasets are used in versions without natural missing entries (Appendix B.10), and all evaluated missingness is generated by the authors’ masks (Appendix B.7). In the main protocols, the longest outages are blocks of 12–48 steps (Sensor25), and whole-sensor outages appear only in a stress test (Sensor40). This is consequential because the position-conditioned design targets exactly these real outage patterns, so it remains open whether the advantage carries over. I suggest adding an experiment on real sensor data with naturally occurring gaps, or with longer outages in the main protocols; if this is not feasible, a discussion of why the synthetic protocols are representative of real outages would also help.
Questions for authors
  1. Could you add at least one “base estimate + learned correction” method (e.g. DCCN-SPF, or RDPI / FENCE if code is available) under the same protocol?
  2. Could you report results on real sensor data with naturally occurring missingness, or on protocols with longer or whole-sensor outages? If not, could you discuss how representative the synthetic protocols are of real outages?
  3. Section 2.2 argues that imputation difficulty depends on where a position lies relative to the observed values. Could you report MAE as a function of gap length or of the distance to the nearest observation, for ARCI-T and the strongest baselines? This would show directly whether the position-conditioned acceptance helps most where local support is weakest.
4 comments
AuthorsOpenReview ↗

Q1|Comparison with "Initial Estimate + Learned Correction" Methods

Thank you for your thorough and insightful review. Your three questions address core empirical and mechanistic aspects of our study—specifically the need for direct comparisons against same-paradigm "base estimate + learned correction" baselines, empirical verification inside the deep interiors of long gaps where local temporal support is weakest, and model behavior under sustained outages. These suggestions have been instrumental in strengthening our experimental validation. We detail our point-by-point responses below:

Q1 | Comparison with "Initial Estimate + Learned Correction" Methods

We fully agree with your point. Since our core method focuses on performing controlled corrections around a deterministic reference, comparing only with models that predict missing values directly is not enough to isolate the benefit of our gated acceptance design—a direct comparison with methods that also follow the "initial estimate + learned correction" paradigm is necessary. Following your suggestion, we evaluated DCCN-SPF and FENCE across 5 random seeds on PeMS-BAY under both Mixed25 and Sensor25, finding that ARCI-T substantially outperforms both same-paradigm baselines across all evaluation metrics.

Specifically, DCCN-SPF was re-implemented from its original codebase and adapted to traffic time series by applying linear interpolation followed by learned residual correction; FENCE uses its full cluster variant, which adjusts diffusion guidance via feedback. All methods were evaluated across 5 random seeds under matched data splits and missingness masks, with errors computed on masked entries in the original data scale.

Table R1 | Five-Seed Results of Same-Paradigm Methods on PeMS-BAY

ProtocolMethodMAE (Mean ±\pm SD)RMSEMRE
Mixed25DCCN-SPF2.2439±0.02552.2439 \pm 0.02555.06940.0360
Mixed25FENCE5.0208±3.00285.0208 \pm 3.002810.20720.0805
Mixed25ARCI-T1.4817±0.01481.4817 \pm 0.01483.26860.0237
Sensor25DCCN-SPF2.5508±0.02362.5508 \pm 0.02365.56570.0409
Sensor25FENCE7.7510±4.65667.7510 \pm 4.656613.79530.1242
Sensor25ARCI-T1.6458±0.01851.6458 \pm 0.01853.55840.0264

As shown in Table R1, compared to DCCN-SPF, ARCI-T lowers 5-seed mean MAE by 33.97% on Mixed25 and 35.48% on Sensor25, along with consistent improvements in per-seed averaged RMSE and MRE. This provides a direct comparison within the "initial estimate plus learned correction" paradigm rather than comparing only against direct prediction models.

These results also clarify the architectural difference: unlike unconstrained residual updates, ARCI-T keeps the deterministic restoration as an explicit output reference and controls how much of each proposed correction is accepted based on the local spatiotemporal evidence available at that position. We will include these comparative results and discussions in the revised manuscript.

(Note on FENCE on PEMS04: Evaluation of FENCE on PEMS04 has also been completed; however, due to severe numerical instability and large error divergence observed on this dataset, we are still checking its configuration and refrain from drawing conclusions on its PEMS04 performance at this stage.)

AuthorsOpenReview ↗

Q2 | Natural Missingness and Sustained Outages

You raise a very practical point regarding real-world deployment. We fully agree that evaluating imputation accuracy on real sensor failures would make the empirical validation even more convincing, and we recognize that synthetic missingness masks alone cannot capture every physical cause of real-world hardware outages. However, due to the objective constraints of publicly available benchmarks, the strictly preprocessed traffic datasets do not contain natural failure intervals paired with ground-truth labels; consequently, we must rely on synthetic masking to conduct reproducible quantitative evaluations, which is the established convention in traffic flow imputation research, followed by essentially all studies in the field. With this in mind, we clarified the ground-truth boundaries of existing preprocessed benchmarks alongside standard evaluation practices in this field, and further analyzed our Sensor40 persistent full-sensor outage stress test—showing that even when 40% of sensors remain completely offline across the entire test horizon, ARCI-T uses spatial correlations to reduce MAE on PEMS04 by 62.00% compared to the deterministic anchor.

Across all four traffic benchmarks used in our study, the preprocessed data assets contain zero natural-missing entries with recoverable ground truth, making it impossible to compute meaningful error metrics on real natural outages. In traffic data imputation research, synthetic masking is the standard and necessary methodology for reproducible benchmarking because original observations must be known in order to mask target entries and evaluate reconstructions quantitatively. For instance, CoFILL adopts controlled point-wise and contiguous block missingness on PEMS-BAY and METR-LA; recent traffic diffusion imputation studies construct random and temporally contiguous missing masks synthetically; and MagiNet similarly relies on controlled missingness protocols. While synthetic masking guarantees fairness and reproducibility, we recognize that it cannot replicate all physical mechanisms of field sensor failures, and we will state this data boundary clearly in the revised manuscript.

To evaluate robustness under prolonged disconnections, we highlight the Sensor40 stress test in our manuscript, where approximately 40% of sensors are randomly selected and held completely unobserved across the entire test horizon.

Table R2 | MAE under Heavy Missingness and Persistent Full-Sensor Outages

DatasetProtocolDeterministic AnchorARCI-TError Reduction
PeMS-BAYMixed402.61231.6451+37.03%
PeMS-BAYSensor405.32103.2628+38.68%
PEMS04Mixed4032.607223.1204+29.09%
PEMS04Sensor40105.289840.0085+62.00%

As shown in Table R2, under PEMS04 Sensor40 (persistent whole-sensor outage), holding a subset of sensors completely unobserved throughout the test horizon simulates real-world scenarios such as sustained power loss or equipment failure: whereas relying solely on the temporal deterministic anchor leads to an MAE of 105.2898, ARCI-T uses spatial evidence from operational neighboring sensors to reduce MAE to 40.0085 (a 62.00% error reduction).

AuthorsOpenReview ↗

Q3 | How Error Varies with Gap Length and Observation Distance

We agree with your point that reporting only protocol-level average MAE hides local performance differences and does not fully support our motivation regarding variations in local observational support. To address this, we stratified test-set missing entries along two dimensions: (1) distance to the nearest observed temporal step, and (2) total length of the contiguous missing gap; this breakdown shows that while ARCI-T and the strongest baseline perform similarly right at gap boundaries (1 step away), ARCI-T maintains a 14.4%–21.8% error reduction deep inside long gaps (16–127 steps) where local temporal support is weak or absent.

To make the physical meaning of "time steps" in these tables clear, we note that the traffic benchmarks used in our study are sampled at 5-minute intervals (i.e., 1 time step = 5 minutes). Thus, our two grouping dimensions correspond to clear physical and geometric conditions:

  1. Distance to the nearest valid observation (Table R3(a)) measures how many time steps separate a missing entry from the closest known observation (either before or after it) on the same sensor. Specifically, "1 step" (5 minutes) denotes positions at the immediate boundary of a gap directly adjacent to a valid observation, where local temporal continuity is strong; "2–3 steps" (10–15 minutes) and "4–7 steps" (20–35 minutes) represent transition zones moving into the gap; and "16–31 steps" (~1.3–2.6 hours) or "32–63 steps" (~2.7–5.2 hours) represent positions deep inside a long outage, far from any valid temporal reference on that sensor.
  2. Total contiguous gap length (Table R3(b)) measures the overall duration of the continuous missing block in which a missing entry is located. Specifically, "16–31 steps" corresponds to a contiguous outage lasting roughly 1.3–2.6 hours, "32–63 steps" corresponds to roughly 2.7–5.2 hours, and "64–127 steps" represents an ultra-long continuous disconnection lasting roughly 5.3–10.6 hours.

In our evaluation, group-level MAEs were computed independently within each seed before averaging across the five seeds. Table R3(a) reports the MAE difference relative to the strongest baseline in the original benchmark (OPCR), where positive values indicate lower MAE for ARCI-T; Table R3(b) reports absolute MAEs across long contiguous gap lengths.

Table R3(a) | Stratification by Distance to Nearest Observation (in Steps): OPCR MAE −- ARCI-T MAE

Dataset / BaselineProtocol1 step2–3 steps4–7 steps16–31 steps32–63 steps
PeMS-BAY / OPCRMixed25+0.062+0.062+0.506+0.506+0.485+0.485+0.304+0.304+0.283+0.283
PeMS-BAY / OPCRSensor25+0.085+0.085+0.437+0.437+0.476+0.476+0.255+0.255+0.248+0.248

Table R3(b) | MAE Stratified by Total Contiguous Gap Length (PeMS-BAY Mixed25)

MethodGap Length 16–31 StepsGap Length 32–63 StepsGap Length 64–127 Steps (Ultra-Long Outage)
Strongest Baseline (OPCR)1.9272.0192.209
ARCI-T (Ours)1.5061.6501.891
ARCI-T Error Reduction+21.8%+18.3%+14.4%

These results directly support our design motivation: immediately adjacent to observed entries (1 step), the deterministic anchor is already very accurate, so the difference between the two methods is small (+0.062+0.062 / +0.085+0.085). Once inside gap interiors where local temporal support disappears (from 2–3 steps out to 32–63 steps), or during ultra-long contiguous outages of 64–127 steps (over 5–10 hours), ARCI-T dynamically increases its spatiotemporal correction, maintaining a clear and consistent accuracy advantage. We will add this position-stratified analysis to the revised manuscript.

AuthorsOpenReview ↗

Manuscript Revision Guide for Reviewer TS2u

Manuscript Revision Guide for Reviewer TS2u

Thank you again for your constructive comments. Following up on our detailed response above, we have uploaded the revised manuscript (with updates highlighted in blue). Below is the exact mapping of your three questions to the sections, pages, figures, and tables added or updated in the revised PDF:

  • Regarding Q1 & W1 (Comparison with "initial estimate + learned refinement" methods and discussion of classical estimators):

    • Section 4.1 (Page 6, Lines 272–276) & New Section 4.5 (Page 10, Lines 489–500, 526–529): Added main-text experimental setup and discussion comparing ARCI-T against DCCN-SPF and FENCE (cluster) on PeMS-BAY under Mixed25 and Sensor25.
    • Appendix D.1 (Page 35, New Table 42) & Appendix B.10 (Page 21, Lines 1116–1120): Added the complete five-seed MAE, RMSE, and MRE comparison table (Table 42) and documented the implementation protocols for DCCN-SPF and FENCE.
    • Appendix A.3 (Page 13, Lines 694–697; Page 14, Table 9): Added a dedicated paragraph discussing how ARCI-T parallels and differs from classical two-stage estimators—specifically Kalman filtering (applying a covariance-derived gain to an observation innovation) and regression kriging (adding an interpolated spatial residual to a fitted trend).
  • Regarding Q2 & W2 (Natural missingness vs. synthetic masks and performance under sustained sensor outages):

    • Section 4.1 (Page 6, Lines 270–272) & Appendix B.10 (Page 21, Lines 1121–1126): Clarified that the preprocessed benchmark assets contain zero natural-missing entries with ground-truth labels (V=1V=1), and explained the standard benchmark convention (Cini et al., 2022; Deng et al., 2024) of applying point, block, and sensor-level masks to valid recorded observations so ground-truth targets remain available for evaluation.
    • Section 4.4 (Page 9, Table 8(a); Page 10, Lines 486–488) & Appendix D.2 (Pages 35–36, Lines 1882–1895, Table 43): Expanded the discussion of the Sensor40 synthetic sustained-outage stress test (where 40% of sensors are completely unobserved across the entire test horizon), demonstrating how frozen ARCI-T checkpoints leverage cross-sensor spatial corrections to reduce PEMS04 Sensor40 MAE by 62.00% (105.2898 →\rightarrow 40.0085), while transparently noting that Sensor40 remains a synthetic stress protocol rather than natural field-failure data.
  • Regarding Q3 (MAE as a function of contiguous gap length and distance to the nearest visible reading):

    • New Section 4.5 (Page 10, Lines 489–537, New Figure 5(a,b)): Added position-stratified error curves comparing ARCI-T and the representative baseline (OPCR) on PeMS-BAY Mixed25 and Sensor25 across seven step bins (1, 2–3, 4–7, 8–15, 16–31, 32–63, and 64–127 five-minute steps) by contiguous missing-run length (Figure 5(a)) and nearest-reading distance (Figure 5(b)).
    • Appendix C.9 (Pages 29–30, Lines 1546–1595, New Table 30 & Table 31) & Appendix B.8 (Page 20, Table 11): Added complete bin-level tables reporting seed coverage nn, total missing positions, contiguous run counts, and bin-level MAEs by contiguous run length (Table 30, noting that Sensor25 has zero samples in the short 1 to 4–7 step bins and populated bins from 8–15 to 64–127 steps) and by distance to the nearest visible reading (Table 31).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.