acceptodds
Under review as a conference paper at ICLR 2027

Learning Higher-Order Interactions from Incomplete Spatiotemporal Data with Mixtures of Hypergraph Laplacians

Abstract

Traffic sensor data are strongly structured in time and across sensors, but reconstructing missing values becomes increasingly difficult as gaps lengthen because missingness removes both local temporal context and the joint observations needed to infer cross-sensor relationships. We introduce MoHL, a mixture of hypergraph Laplacians that learns group-wise relational structure directly from incomplete spatiotemporal data. MoHL discovers overlapping sensor groups from observed co-variation, assigns each interaction order a Laplacian component, and learns their mixture from mask-matched holdouts. This group-wise parameterization pools evidence across related sensors rather than estimating every coupling independently. We show that the learned mixture induces structured weight sharing and characterize when such sharing is beneficial. Across three traffic networks, higher-order groups provide the largest gains under extended gaps, and shared weights consistently outperform freely learned weights on the same support when relational evidence is scarce. MoHL achieves the lowest error among evaluated methods under day-long gaps while using orders of magnitude fewer learned parameters than deep imputers. These results show that higher-order group structure is most useful precisely when missingness makes relational evidence scarce.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hypergraph Learning, Higher-Order Interactions, Spatiotemporal Imputation, Hypergraph Laplacians

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_m9EcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers the spatiotemporal imputation problem in traffic data. It introduces MoHL, a mixture of hypergraph Laplacians that learns group-wise relational structure directly from incomplete spatiotemporal data.

Critical strengths

MoHL discovers sensor groups from co-observed residual correlations, learns weights through mask-matched holdouts, and adapts regularization to local evidence. Its central contribution is a compact parameter-sharing scheme, with reported advantages under day-long outages.

Critical weaknesses
  1. The paper explicitly shows that its hypergraph penalty is equivalent to a graph penalty with shared edge weights. This raises a question about the group construction: would a simpler correlation-based graph perform similarly? A baseline with matched connectivity and a comparable number of learned parameters would help establish what the groups contribute.
  2. The experimental setting limits the conclusions. The main results use 100 high-degree sensors from a selected seven-day window, and the deep baselines are trained from scratch on that window. I would like to see whether the gains hold across full networks and different time windows, and how the comparison changes when deep models can use longer training histories.
  3. The shared-versus-free comparison is useful, but the same optimization budget may not suit both models. Learning a separate weight for every edge may require more training or stronger regularization. Without checking these factors, it is difficult to attribute the performance gap specifically to useful group structure.
  4. I am not fully convinced by the claim that sharing helps most when observations are scarce. In Table 9, its advantage does not shrink as the observation window grows, while Table 2 includes settings where free weights perform better. Please explain how these findings support the claim and report uncertainty estimates for the main comparisons.
Questions for authors

See Weakness

12 comments
AuthorsOpenReview ↗

Reply to Reviewer m9Ec (1/7)

Thank you for your thoughtful review, constructive suggestions, and patience. We apologize for the delay in responding. We took additional time to complete a substantial set of experiments so that we could address your concerns carefully and thoroughly.

We address your comments in the replies that follow. For reference, Table M1 reports the full-dataset evaluation on all sensors, averaged over the test weeks.

Table M1. Full-dataset evaluation: MAE on all sensors, averaged over the test weeks.

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.890 ± 0.1534.736 ± 0.2214.091 ± 0.16272.276 ± 2.33882.536 ± 1.98073.180 ± 2.51875.894 ± 1.55792.907 ± 1.70276.847 ± 0.951
HaLRTC1.263 ± 0.0351.993 ± 0.0863.507 ± 0.17110.978† ± 0.17813.397 ± 0.34472.171 ± 3.29218.038 ± 0.50723.205 ± 0.868106.606 ± 1.809
LETC2.409 ± 0.0642.712 ± 0.1014.949 ± 0.14731.816 ± 0.98547.821 ± 1.57582.026 ± 2.44248.419 ± 0.67564.663 ± 0.896116.571 ± 1.279
WDGTC2.286 ± 0.1112.679 ± 0.14713.990 ± 0.21016.889 ± 0.80017.778 ± 0.64268.535 ± 1.79923.442 ± 0.23926.569 ± 0.519104.503 ± 3.344
GRIN0.874† ± 0.0141.424† ± 0.0382.885† ± 0.17011.273 ± 0.26215.158 ± 0.48327.541 ± 0.27017.629† ± 0.41323.273 ± 0.70156.956 ± 1.899
SPIN0.937 ± 0.0171.447 ± 0.0413.299 ± 0.20512.085 ± 0.25214.401 ± 0.43224.390 ± 1.31418.003 ± 0.34821.535 ± 0.74035.038 ± 0.754
ImputeFormer1.071 ± 0.0281.474 ± 0.0353.061 ± 0.14012.356 ± 0.25313.949 ± 0.40620.385† ± 2.16717.931 ± 0.48520.946† ± 0.29929.662† ± 1.111
MoHL0.872 ± 0.0141.422 ± 0.0362.697 ± 0.09410.974 ± 0.24013.400† ± 0.28516.719 ± 1.27416.991 ± 0.35319.747 ± 0.60725.928 ± 0.674
Improvement (%) ↑+0.23+0.14+6.52+0.04−0.02+17.98+3.62+5.72+12.59

Mean ± s.d. across test weeks. Bold: lowest MAE; †: second lowest.

AuthorsOpenReview ↗

Reply to Reviewer m9Ec (2/7)

W1. Connectivity and parameter-count controls

Response.

Thank you for proposing this control. We agree that connectivity and parameter count need separate checks. Most controls trail MoHL on flow data, but simpler graphs win in some settings. The differing gate configurations limit attribution to grouping alone.

What we did.

We used MoHL’s residual correlations, profile, road-graph prior, solver and learning procedure, and tested:

  1. Matched connectivity with one weight.
  2. A top-5 graph with five per-rank weights.
  3. Top-5 graphs with one shared weight or free edge weights.

Tables M2 and M2b cover 100 sensors and all sensors. The controls omit the gate while MoHL retains it, so they do not isolate coupling structure.

Evidence and discussion.

Table M2. Correlation-graph controls: control − MoHL, grouped by what they match (100 sensors).

PairsLearned weightsPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
MoHL (MAE)600–96050.6991.2412.6509.3911.1713.8519.5122.9631.83
Matched connectivity
threshold graph with MoHL's pair count, one weight600–9601+0.005−0.001+0.074+0.89+0.78+0.62+0.08+0.19+0.60
all pairs of MoHL's groups, one weight600–9601+0.004+0.004+0.086+0.96+0.84+0.64+0.08+0.13+0.34
Matched number of weights
top-5 graph, one weight per rank340–4105+0.001+0.015+0.041+0.25+0.24+0.22+0.06+0.13+0.28
Fewer or more weights
top-5 graph, one shared weight340–4101+0.003−0.001+0.068+0.70+0.48+0.17+0.07+0.16+0.08
top-5 graph, free weight per edge340–410340–410+0.020+0.119−0.024−0.08+0.03+0.57+0.08+0.64+0.26

Positive differences favour MoHL; negative differences favour the control.

Matched controls. One-weight matched-support graphs trail on flow data; the per-rank control trails in every 100-sensor column. Connectivity or parameter count alone therefore does not reproduce the reported configuration’s accuracy, although the gate difference limits attribution.

What grouping changes. Couplings accumulate across shared groups and share weights by order, pooling evidence while allowing different effective pair couplings. This remains a pairwise quadratic penalty.

Full networks. The per-rank control still trails on PEMS03 day-long gaps; the threshold graph leads on PEMS-BAY Block-288. Most comparisons favour MoHL, with the same attribution limit.

Table M2b. Correlation-graph controls on the full networks: control − MoHL, grouped by what they match (all sensors, test weeks).

PairsLearned weightsPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
MoHL (MAE)2,030–3,37050.8801.4972.77311.4614.0518.9717.0020.5827.15
Matched connectivity
threshold graph with MoHL's pair count, one weight2,030–3,3701+0.012+0.026−0.040+0.71+0.20+0.26+0.12+0.23+0.46
all pairs of MoHL's groups, one weight2,030–3,3701+0.011+0.021+0.051+0.67+0.16+0.85+0.08+0.12+0.86
Matched number of weights
top-5 graph, one weight per rank1,190–1,3805+0.002+0.009+0.041+0.10+0.04+0.73+0.05+0.05+0.33
Fewer or more weights
top-5 graph, one shared weight1,190–1,3801+0.010+0.012+0.045+0.51+0.06+0.64+0.06+0.11+0.44
top-5 graph, free weight per edge1,190–1,3801,190–1,380+0.058+0.110+0.049+0.82+0.04+0.62+0.13+0.51+0.02

Positive differences favour MoHL; negative differences favour the control.

Conclusion.

MoHL generally outperforms simpler correlation graphs, including controls matched in connectivity or parameter count. The five-weight per-rank control has higher MAE in every reported setting at both sensor scales, supporting MoHL’s compact, structured design.

AuthorsOpenReview ↗

Reply to Reviewer m9Ec (3/7)

W2. Full networks, observation windows and training history

Response.

Thank you for these suggestions. MoHL is designed to reconstruct an incomplete recording when a long, representative training archive is unavailable. It learns sensor groups and shared weights from the current observations, pooling limited evidence while allowing the relational structure to be re-estimated for each window. Shared weights keep the estimation problem compact, while re-estimating the groups makes the model specific to the current window. We agree that robustness to network size and window choice, and the benefit of adding historical data, should be assessed separately.

Design rationale.

  1. Learn from the available window. The daily profile and sensor groups are estimated from observed entries, and the shared weights are learned through self-supervision on observed holdouts.
  2. Pool scarce evidence. Sharing weights by group order reduces the estimation burden compared with learning each pairwise coupling independently.
  3. Adapt the structure to the window. Re-estimating the groups allows the relational structure to reflect current co-variation. This supports flexible use when historical records are unavailable, limited or no longer representative.

What we did.

  1. Network size: we evaluated every method on all sensors and the test weeks, using only the week being completed.
  2. Window robustness: we tested additional weeks and longer windows at both 100 sensors and on the full networks (Tables M4 and M4b).
  3. Access to history: we trained the deep models on months of observations and gave MoHL the same history for its daily profile and groups.

Evidence and discussion.

Full-dataset accuracy. Table M3 repeats the M1 evaluation on all sensors and the test weeks.

Table M3. Full-dataset evaluation: MAE on all sensors, averaged over the test weeks.

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.8904.7364.09172.27682.53673.18075.89492.90776.847
HaLRTC1.2631.9933.50710.978†13.39772.17118.03823.205106.606
LETC2.4092.7124.94931.81647.82182.02648.41964.663116.571
WDGTC2.2862.67913.99016.88917.77868.53523.44226.569104.503
GRIN0.874†1.424†2.885†11.27315.15827.54117.629†23.27356.956
SPIN0.9371.4473.29912.08514.40124.39018.00321.53535.038
ImputeFormer1.0711.4743.06112.35613.94920.385†17.93120.946†29.662†
MoHL0.8721.4222.69710.97413.400†16.71916.99119.74725.928
Improvement (%) ↑+0.23+0.14+6.52+0.04−0.02+17.98+3.62+5.72+12.59

Bold: lowest MAE; †: second lowest. Positive improvement means lower MAE than the best baseline.

Performance in the intended setting. With only the current week available, MoHL has the lowest reported mean under day-long gaps on all three full networks. This extends the long-gap result beyond the original 100-sensor subset. Short-gap margins are smaller and rankings vary by regime, including HaLRTC’s slight lead on PEMS03 Cluster.

Different deployment scenarios. An established network may have months of representative records from which a history-trained model can learn recurring patterns; that model is more accurate in our history experiment. A new deployment or a short recording released for analysis may not provide such an archive, and older records may no longer reflect current sensor relationships. MoHL can be fitted from the observed entries of the current window and rebuild its groups for subsequent windows, so deployment does not depend on collecting a separate pretraining archive. The history experiment therefore evaluates the benefit of an additional data resource, while MoHL addresses flexible reconstruction from the data currently available.

Reviewer_SzvzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MoHL for imputing traffic sensor data with long, structured gaps, where cross-sensor structure matters most but the data to learn it are scarce. MoHL builds nested sensor groups from pairwise correlations of de-seasonalized residuals. It assigns one hypergraph Laplacian per group size, which is equivalent to a graph with shared pair weights. It learns a mixture of these Laplacians and the road-graph prior inside a convex spatiotemporal solve, trained on holdouts that mimic the real missingness. On three PEMS datasets, MoHL performs best mainly under day-long gaps. A control with free weights on the same sensor pairs is used to argue that the gains come from weight sharing.

Critical strengths
  1. The paper names a real tension in imputing data with long gaps. The regimes that most need cross-sensor structure are usually the ones that leave the least joint evidence for learning that structure.
  2. The paper frames the hypergraph Laplacian openly as weight sharing over a graph. That makes the inductive bias explicit.
Critical weaknesses
  1. The method has many interacting components, and it is unclear what each component contributes. The main text does not make clear which of them drive the main results.
  2. The "higher-order" framing overstates what is learned. The operator is a pairwise Laplacian. The missing control is a pairwise graph on the top-5 correlation neighbors with a single shared weight, or with learned per-rank weights.
  3. Only 100 sensors and one 7-day window per network are evaluated, and no runtime is reported. Scalability of group construction and of the CG solve to full networks is therefore not shown.
  4. Evidence for the shared-vs-free and SOTA claims is weaker than stated. Also, deep imputers are trained only on the same 7-day, 100-sensor window, far below their intended data regime.
  5. The evidence-scarcity mechanism is supported only on data simulated from the model itself. On real data, it is not confirmed. The claim that sharing helps because evidence is scarce is therefore not established.
Questions for authors
  1. How do the deep imputers perform when given the daily profile as input, or when trained on a longer history?
  2. What are the runtime and memory usage on a full network (all dataset sensors)?
17 comments
AuthorsOpenReview ↗

Reply to Reviewer Szvz (1/13)

Thank you for your thoughtful review, constructive suggestions, and patience. We apologize for the delay in responding. We took additional time to complete a substantial set of experiments so that we could address your concerns carefully and thoroughly.

We address your comments in the replies that follow. For reference, Table S1 reports the full-dataset evaluation on all sensors, averaged over the test weeks.

Table S1. Full-dataset evaluation: MAE on all sensors, averaged over the test weeks.

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.890 ± 0.1534.736 ± 0.2214.091 ± 0.16272.276 ± 2.33882.536 ± 1.98073.180 ± 2.51875.894 ± 1.55792.907 ± 1.70276.847 ± 0.951
HaLRTC1.263 ± 0.0351.993 ± 0.0863.507 ± 0.17110.978† ± 0.17813.397 ± 0.34472.171 ± 3.29218.038 ± 0.50723.205 ± 0.868106.606 ± 1.809
LETC2.409 ± 0.0642.712 ± 0.1014.949 ± 0.14731.816 ± 0.98547.821 ± 1.57582.026 ± 2.44248.419 ± 0.67564.663 ± 0.896116.571 ± 1.279
WDGTC2.286 ± 0.1112.679 ± 0.14713.990 ± 0.21016.889 ± 0.80017.778 ± 0.64268.535 ± 1.79923.442 ± 0.23926.569 ± 0.519104.503 ± 3.344
GRIN0.874† ± 0.0141.424† ± 0.0382.885† ± 0.17011.273 ± 0.26215.158 ± 0.48327.541 ± 0.27017.629† ± 0.41323.273 ± 0.70156.956 ± 1.899
SPIN0.937 ± 0.0171.447 ± 0.0413.299 ± 0.20512.085 ± 0.25214.401 ± 0.43224.390 ± 1.31418.003 ± 0.34821.535 ± 0.74035.038 ± 0.754
ImputeFormer1.071 ± 0.0281.474 ± 0.0353.061 ± 0.14012.356 ± 0.25313.949 ± 0.40620.385† ± 2.16717.931 ± 0.48520.946† ± 0.29929.662† ± 1.111
MoHL0.872 ± 0.0141.422 ± 0.0362.697 ± 0.09410.974 ± 0.24013.400† ± 0.28516.719 ± 1.27416.991 ± 0.35319.747 ± 0.60725.928 ± 0.674
Improvement (%) ↑+0.23+0.14+6.52+0.04−0.02+17.98+3.62+5.72+12.59

Mean ± s.d. across test weeks. Bold: lowest MAE; †: second lowest.

AuthorsOpenReview ↗

Reply to Reviewer Szvz (2/13)

W1. Contribution of each component

Response.

Thank you for asking us to clarify the contribution of each component. We agree that the submission did not explain this clearly enough. The sequential ablations identify the daily profile and group couplings as the main contributors to long-gap accuracy, with a smaller improvement from the gate. Changing one component at a time confirms this; it also shows that the temporal term carries short gaps and that weights learned on gap-shaped holdouts are needed under day-long gaps.

What we did.

  1. We restated Table 2 as sequential removals from full MoHL (Table S2), sharing masks, learning splits, solver and training budget.
  2. We repeated the ablation on all sensors (Table S2b).
  3. We changed one component at a time, removing it or replacing it with a simpler choice, and kept all the others (Table S2c, five mask seeds), then repeated this on all sensors (Table S2d).

Evidence and discussion.

How to read the ablation. Each row removes one additional component, so adjacent rows measure its contribution conditional on the components that remain. These are sequential removals, rather than independent leave-one-component-out effects; Table S2c gives the latter.

  • Gate: adapts regularization to local evidence; removing it gives MoHL-Static.
  • Orders 3–6: couple groups of three to six sensors with correlated residuals, sharing one weight per order.
  • Order 2: couples correlated sensor pairs with one shared weight.
  • Daily profile: supplies each sensor's typical time-of-day value; removing all these components leaves the prior-graph backbone, MoHL-Base.

Table S2. Ablation of MoHL (Table 2 of the paper): MAE as MoHL's components are removed in turn (100 sensors).

PEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
MoHL (full model)0.6991.2622.7519.62611.4314.0619.5323.0831.31
MoHL w/o gate (MoHL-Static)0.700†1.2682.808†9.652†11.45†14.12†19.61†23.14†31.38†
MoHL w/o gate, groups of orders 3–60.7031.2942.8149.98412.1915.3219.6223.3032.28
MoHL w/o gate, all groups0.7321.267†2.88210.9912.9716.0719.6323.6134.57
MoHL w/o gate, all groups, daily profile (MoHL-Base)0.7321.2733.12811.1113.3122.9819.8225.1752.71

Each row removes one additional component; “all groups” means orders 2–6. Bold: lowest MAE; †: second lowest.

Daily profile: a reference for missing days. Its removal causes the largest long-gap deterioration on the flow networks, while the effect under Cell is small. This fits its role: nearby observations can fill scattered gaps, but a missing day needs a reference shape.

Groups: information beyond the daily pattern. Removing the groups raises day-long-gap error on every network. Orders 3–6 account for a substantial part of the improvement on PEMS03. The PEMS-BAY Cluster result is also informative: pairs alone can hurt, while larger groups recover that loss. Adding cross-sensor links is therefore insufficient by itself; their organization matters.

Gate: a smaller adjustment. The gate provides smaller gains than the profile and groups in these sequential ablations. We describe it as an adaptation to local evidence, rather than the main source of the reported improvement.

The ablation on the full networks.

Full-network check. Each removal leaves the printed MAE unchanged or increases it in Table S2b. The profile and groups again account for the main long-gap improvements, supporting the same interpretation beyond the 100-sensor selection.

Table S2b. Ablation of MoHL on the full networks: MAE as MoHL's components are removed in turn (all sensors, test weeks).

PEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
MoHL (full model)0.8791.4992.75411.4614.0819.2817.0120.3926.89
MoHL w/o gate (MoHL-Static)0.881†1.508†2.781†11.54†14.0819.51†17.06†20.50†27.18†
MoHL w/o gate, groups of orders 3–60.8831.5182.80711.9414.3919.6617.0720.7528.50
MoHL w/o gate, all groups0.9071.5512.89712.2814.6421.4417.1421.1832.42
MoHL w/o gate, all groups, daily profile (MoHL-Base)0.9121.5573.30712.3515.1128.0217.2923.2054.29

Each row removes one additional component. Bold: lowest MAE; †: second lowest.

AuthorsOpenReview ↗

Reply to Reviewer Szvz (3/13)

W1 (continued)

Evidence and discussion (continued).

Table S2c. One component at a time: MAE of MoHL with one component removed or replaced (100 sensors, mean of five mask seeds).

PEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
Groups: which sensors share a weight
MoHL (full model)0.6991.2432.6709.58611.3413.8519.5122.9632.03
MoHL w/o groups of orders 3–60.6991.2542.715†9.841†11.90†14.69†19.52†23.07†32.77†
MoHL w/o groups0.703†1.2382.83311.0012.5516.2619.5423.2635.46
MoHL with random groups0.703†1.239†2.83811.0012.5615.5819.5423.2933.87
Daily profile and other terms
MoHL (full model)0.6991.2432.670†9.58611.3413.8519.5122.9632.03
MoHL w/o daily profile0.6991.1942.86810.4112.2415.8019.6623.5239.05
MoHL w/o temporal term1.7392.2622.65010.05†11.68†13.90†25.5028.7832.29†
MoHL w/o cross-sensor terms (road graph and groups)0.704†1.238†3.54511.1812.7322.7219.56†23.41†43.19
Learning: how the weights are set
MoHL (full model)0.6991.2432.6709.58611.34†13.8519.5122.9632.03
MoHL with untrained weights1.603†2.1562.771†11.5813.1416.2021.7926.7933.26†
MoHL with uniform-cell holdouts0.6991.267†2.9149.589†11.2314.03†19.52†23.11†44.56
Gate: how the weights adapt to local evidence
MoHL (full model)0.6991.2432.6709.58611.3413.8519.5122.9632.03
MoHL w/o gate (MoHL-Static)0.6991.255†2.690†9.597†11.38†13.89†19.57†23.04†32.06†

Each row changes one component of the full MoHL and keeps all the others, with the same masks, learning splits, solver and training budget; each category repeats the full model. Random groups keep the sizes of MoHL's groups with random members; untrained weights are the starting values of MoHL's learning. Bold: lowest MAE within each category; †: second lowest.

Long gaps: groups, daily profile and holdouts. Under Block-288, random groups of the same sizes lose most of the gain from MoHL's groups (PEMS03: 15.58 vs. 13.85, and 16.26 without groups), and road-neighbour groups do about as well as random ones: the gain comes from joining correlated sensors. Removing the daily profile or the cross-sensor terms raises PEMS04 Block-288 from 32.03 to 39.05 or 43.19, and uniform-cell holdouts raise it to 44.56.

Short gaps, learned weights and the gate. The temporal term carries short gaps (PEMS-BAY Cell: 1.739 without it vs. 0.699) but changes Block-288 by at most 0.26; removing the profile even helps on PEMS-BAY Cluster (1.194 vs. 1.243). Learned weights matter in every column, and the gate adds a small gain (at most 0.08).

Reviewer_tv87OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors of this paper consider the problem of missing data from traffic sensors, which not only affects the history of a single sensor but also the joint observations that are required to learn relationships to neighboring sensors. Addressing this issue, the authors introduce "Mixture of Hypergraph Laplacians" (MoHL), a method that seeks to identify groups of correlated sensors by calculating pairwise correlations in the observed deviations from typical daily patterns. Groups of different size, i.e. pairs, triplets, etc. are then represented as components in a mixture of hypergraph Laplacians, with a single weight capturing all sensors in a group. Weights of this mixture are learned in a self-supervised way by masking parts of the data. The method is tested on a data set with three real-world traffic networks, where it achieves the lowest error among the evaluated methods, while using orders of magnitude fewer learnable parameters.

Critical strengths
  • The authors address a clearly motivated problem with practical relevance in a meaningful way, using a meaningful approach.

  • To the best of my knowledge, the approach to use a learnable mixture of hypergraph Laplacians is novel and a genuinely innovative aspect of this work.

  • The proposed approach is relatively simple and has an order of magnitude fewer parameters than competing methods building on deep learning.

  • While the paper addresses a relatively narrow domain, the method itself is domain-agnostic and could be adapted to other settings where we have access to incomplete spatiotemporal sensor data.

  • Key design decisions, like the shared weight approach, are justified by a theoretical analysis.

Critical weaknesses
  • The biggest issue that I have with the paper is its relatively narrow scope of the empirical evaluation, see my question Q1 and suggestions below.

  • The method has been tested in a fixed, static time window. However, in a practical application e.g. for real-time monitoring it seems important that the method can also work in a streaming setting with continuously updated data, see Q2.

  • From the empirical evaluation, it seems that the performance of the method, and specifically its advantage to deep imputation methods, is sensitive to the size of the data gaps. See my question Q3.

  • Clearly, the way the method works, it strongly relies on a "cross-sensor extrapolation" that utilizes typical patterns, which means it may underperform on holidays or other atypical settings, where a forecasting might be most challenging but also useful. This limitation is discussed by the authors, but I still think it might limit the practical appeal of the approach.

  • Finally, the ablation reveals that a variant of the model where each sensor pair has its own weight outperforms the proposed approach, which is based on the idea that the sharing of weights within groups of a given order should reduce the estimation error when data is scarse. See my question Q4.

Questions for authors
  • Q1: As mentioned above, the scope of the empirical evaluation is very narrow, since the authors only test it in a single one-week data window from three US highways covered in a single data set only. How sensitive is the method to the size or choice of the observation window or the specific characteristics of that data? Also, given that the authors argue that their method is transferable to other types of sensor data, could you evaluate it in data from other settings as well? Concrete examples for readily available data that seem to be suitable for your method are:

  • PurpleAirSF air monitoring data: https://dl.acm.org/doi/pdf/10.1145/3589132.3625575

  • Smart Meter Data: https://arxiv.org/abs/2501.07276v2

  • LargeST benchmark: https://arxiv.org/pdf/2306.08259

  • Q2: Given the intended application domain, I would argue that a real-time monitoring setting, where the method is exposed to continuously updated data within a moving time window is important. I could not find a discussion on whether the method would be suitable in such a setting. Could the authors discuss this, possibly in the limitations section?

  • Q3: Carefully checking the results in the appendix, especially tables 4, 5 and 6, it seems that the proposed method underperforms in scenarios where the data gaps are short. In particular, table 4 shows that for 6 hour gaps deep imputation methods may perform better than the proposed method. I do not consider it a major issue that the proposed method does not outperform more complex methods in all circumstances, but I do think this might point to an issue that is not really discussed well in the work. Namely, that the basic group--based mechanism of MoHL mainly pays off when the temporal signal from a single sensor is so limited that the cross-sensor patterns become dominant. For shorter gaps. a simple interpolation mechanism might still work better. Since this raises a question in which condition a practitioner would want to use the proposed method, the authors should discuss this in the paper.

  • Q4: In the theoretical analysis, the authors argue why shared weights are expected to beat free weights in terms of estimation variance in a regime where evidence is scarce. However, this does not seem to be the case in some of the ablation results shown in Table 2, where the free weight variant outperforms the shared variant in three cases. This is also acknowledged in section 5.2 and further analyzed in table 7, which however misses exactly the data set PEMS03, which seems to break the theoretical prediction. Could the authors elaborate on whether their theoretical findings are fully backed by the experimental results? And why is the data set PEMS03 missing from table 7?

13 comments
AuthorsOpenReview ↗

Response to Reviewer tv87 (1/10)

Thank you for your thoughtful review, constructive suggestions, and patience. We apologize for the delay in responding. We took additional time to complete a substantial set of experiments so that we could address your concerns carefully and thoroughly.

We address your comments in the replies that follow. For reference, Table T1 reports the full-dataset evaluation on all sensors, averaged over the test weeks.

Table T1. Full-dataset evaluation: MAE on all sensors, averaged over the test weeks.

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.890 ± 0.1534.736 ± 0.2214.091 ± 0.16272.276 ± 2.33882.536 ± 1.98073.180 ± 2.51875.894 ± 1.55792.907 ± 1.70276.847 ± 0.951
HaLRTC1.263 ± 0.0351.993 ± 0.0863.507 ± 0.17110.978† ± 0.17813.397 ± 0.34472.171 ± 3.29218.038 ± 0.50723.205 ± 0.868106.606 ± 1.809
LETC2.409 ± 0.0642.712 ± 0.1014.949 ± 0.14731.816 ± 0.98547.821 ± 1.57582.026 ± 2.44248.419 ± 0.67564.663 ± 0.896116.571 ± 1.279
WDGTC2.286 ± 0.1112.679 ± 0.14713.990 ± 0.21016.889 ± 0.80017.778 ± 0.64268.535 ± 1.79923.442 ± 0.23926.569 ± 0.519104.503 ± 3.344
GRIN0.874† ± 0.0141.424† ± 0.0382.885† ± 0.17011.273 ± 0.26215.158 ± 0.48327.541 ± 0.27017.629† ± 0.41323.273 ± 0.70156.956 ± 1.899
SPIN0.937 ± 0.0171.447 ± 0.0413.299 ± 0.20512.085 ± 0.25214.401 ± 0.43224.390 ± 1.31418.003 ± 0.34821.535 ± 0.74035.038 ± 0.754
ImputeFormer1.071 ± 0.0281.474 ± 0.0353.061 ± 0.14012.356 ± 0.25313.949 ± 0.40620.385† ± 2.16717.931 ± 0.48520.946† ± 0.29929.662† ± 1.111
MoHL0.872 ± 0.0141.422 ± 0.0362.697 ± 0.09410.974 ± 0.24013.400† ± 0.28516.719 ± 1.27416.991 ± 0.35319.747 ± 0.60725.928 ± 0.674
Improvement (%) ↑+0.23+0.14+6.52+0.04−0.02+17.98+3.62+5.72+12.59

Mean ± s.d. across test weeks. Bold: lowest MAE; †: second lowest.

AuthorsOpenReview ↗

Response to Reviewer tv87 (2/10)

W1/Q1. Evaluation scope, observation windows and additional datasets

Response.

Thank you for the careful review and concrete suggestions. We agree that evaluating additional windows, full networks and other domains provides a stronger assessment of generality. We therefore expanded the experiments along each of these dimensions.

What we did: We expanded the experiments to all sensors, four additional weeks per network, windows of 2–28 days, and five additional datasets, including the LargeST, PurpleAirSF and smart-meter data you suggested. We also examined the data characteristics.

Evidence and discussion.

Full networks: the clearest gains are under day-long gaps. The original Table 1 uses the 100 best-connected sensors and the week with the fewest missing readings, which could favour cross-sensor methods. Table T1 therefore evaluates all sensors on the test weeks. Each method sees only the week it fills in. Baselines use their published settings; MoHL's profile-bin and additional settings were selected on the test weeks. MoHL has the lowest reported mean under day-long gaps on all three networks. These are descriptive comparisons; we do not attach paired significance claims to the T1 means.

Short gaps: small margins require a narrower interpretation. MoHL leads across the PEMS04 regimes, but several Cell and Cluster differences on PEMS-BAY and PEMS03 are very small. HaLRTC remains slightly lower on PEMS03 Cluster. The full-network results therefore support the long-gap finding more clearly than a general advantage for short gaps (see W3/Q3).

Choice of week. We repeated the 100-sensor experiment on four additional weeks per network, including two holidays, to test whether the Table 1 week was unusually favourable. Table T2 reports the mean over these weeks.

Table T2. Other weeks: mean MAE over four further weeks per network (100 sensors).

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.9494.9554.04581.5580.9983.7691.28104.1990.87
HaLRTC1.3812.2203.7169.65†11.1825.9422.3328.3980.17
LETC2.0232.4993.67435.1951.9486.5150.1167.83117.60
WDGTC2.2793.44815.68814.0415.0145.2626.7133.5297.26
GRIN0.848†1.457†3.085†9.9711.9822.2019.96†25.1850.10
SPIN0.9631.4623.51511.0012.3720.5720.6124.1739.78
ImputeFormer1.0541.5303.22110.4611.19†15.35†20.3123.42†32.72†
MoHL0.8461.4532.9319.6411.2615.2119.3523.3131.01
MoHL − best other ↓−0.002−0.004−0.154−0.01+0.08−0.14−0.61−0.11−1.71
The same in the Table 1 week−0.007−0.029−0.278−0.174−0.12−0.66−0.73−0.08−0.35

Bold: lowest MAE in each column; †: second lowest.

AuthorsOpenReview ↗

Reply to Reviewer tv87 (3/10)

W1/Q1 (continued)

Evidence and discussion (continued).

What changing the week establishes. MoHL retains the lowest day-long-gap mean on all three networks. The short-gap rankings remain mixed, including HaLRTC's lead on PEMS03 Cluster. Thus, the additional weeks strengthen the long-gap robustness claim without establishing uniform dominance across gap types.

All sensors on the additional weeks. Table T2b repeats the comparison at full scale. MoHL again leads under day-long gaps on every network. GRIN remains lower on PEMS-BAY Cluster, while several Cell and Cluster margins are small. This is a separate robustness check from Table T1.

Table T2b. Other weeks on the full networks: mean MAE over four further weeks per network (all sensors).

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.6984.4133.91074.4884.5574.4075.1491.3376.66
HaLRTC1.2251.8763.32111.54†14.20†73.8517.9922.80102.86
LETC2.3412.5714.75132.8149.2083.6247.2063.34113.91
WDGTC2.1652.50414.43117.5118.4471.6223.7626.76104.77
GRIN0.862†1.3582.816†11.6815.5629.4917.61†23.6456.36
SPIN0.9291.4173.07212.7314.9426.2918.0221.7137.32
ImputeFormer1.0281.4282.92112.7614.5321.01†17.8421.12†30.46†
MoHL0.8591.401†2.69411.5214.1819.7917.0420.6727.53
MoHL − best other ↓−0.003+0.043−0.122−0.02−0.02−1.22−0.57−0.45−2.93
The same on the full dataset−0.002−0.002−0.188−0.004+0.003−3.666−0.638−1.199−3.734

The full-dataset reference uses Table T1. Bold: lowest MAE; †: second lowest.

Window length at 100 sensors. Table T3 varies the densest window from 2 to 28 days under day-long gaps.

Table T3. Window length: MAE under Block-288 for windows of 2 to 28 days (100 sensors).

MethodPEMS-BAY 2 daysPEMS-BAY 4 daysPEMS-BAY 7 daysPEMS-BAY 14 daysPEMS-BAY 28 daysPEMS03 2 daysPEMS03 4 daysPEMS03 7 daysPEMS03 14 daysPEMS03 28 daysPEMS04 2 daysPEMS04 4 daysPEMS04 7 daysPEMS04 14 daysPEMS04 28 days
GRIN2.4302.875†2.826†2.505†2.04754.5224.0722.2320.6618.68112.6253.5050.9043.0439.99
ImputeFormer2.760†2.7772.9852.6541.91628.39†16.49†14.35†12.28†9.8143.97†41.08†31.88†27.08†25.42
MoHL3.7243.1072.6502.3252.040†25.9016.4413.6712.2111.50†33.7638.1331.0526.9725.76†
MoHL − best other ↓+1.294+0.330−0.176−0.180+0.124−2.49−0.05−0.68−0.07+1.69−10.21−2.95−0.83−0.11+0.34

Bold: lowest MAE; †: second lowest.

Comments on the paper

2 comments
Area_Chair_HJoEOpenReview ↗

Please join the author-reviewer discussion

Dear the authors and reviewers,

We noticed the authors did not provide responses. Our author and reviewer discussion will end soon in 18 November (AOE). If you still have some issues to discuss, this may still be a good time to do it. We highly appreciate your time on discussing and exchanging thoughts to keep our ICLR standards. Many thanks!

Bests,
Your AC

AuthorsOpenReview ↗

General Response

We sincerely thank the reviewers for their careful reading and constructive feedback. In response, we have conducted substantial additional experiments and analyses to address their concerns, strengthen the evaluation, and clarify the presentation.

The additions cover five areas:

  • Broader evaluation: full traffic networks, four additional weeks per network, windows of 2–28 days, and five additional datasets.
  • Baseline comparisons: deep models trained on months of history or given the daily profile, and MoHL given the same history.
  • Model analysis: ablations, correlation-graph controls, removal of real readings, and larger training budgets and stronger regularization for free weights.
  • Practical evaluation: gap-length analysis, interpolation baselines, holidays, parameter reuse, streaming, runtime and memory.
  • Uncertainty: repeated mask seeds and paired intervals for matched comparisons.

For detailed responses to each comment, please see our replies to Reviewers tv87, Szvz and m9Ec below.

Definitions used in the replies.

  • Gaps: Cell hides individual readings; Cluster hides a sensor and its four nearest neighbours for 30 minutes. Block-6/72/288 hides one sensor for 30 minutes/6 hours/24 hours. Traffic experiments hide 50% the observed readings.
  • Settings: Table 1 uses the 100 highest-degree sensors on the densest week. Full networks use all 325/358/307 sensors of PEMS-BAY/03/04 over all the test weeks.
  • History: current week only means each method sees only the week it imputes. With months of history, deep models train on the first 70% of weeks; MoHL’s history variants estimate their profile and groups from those weeks.
  • Weights: MoHL-Static omits the gate and shares weights by group order. MoHL-Free uses the same coupled pairs with a separate weight per pair.

Table G1 reports the full-dataset evaluation of PEMS-BAY, PEMS03 and PEMS04, using all sensors and averaging over the test weeks.

Table G1. Full-dataset evaluation: MAE on all sensors, averaged over the test weeks.

MethodPEMS-BAY CellPEMS-BAY ClusterPEMS-BAY Block-288PEMS03 CellPEMS03 ClusterPEMS03 Block-288PEMS04 CellPEMS04 ClusterPEMS04 Block-288
kNN-spatial3.890 ± 0.1534.736 ± 0.2214.091 ± 0.16272.276 ± 2.33882.536 ± 1.98073.180 ± 2.51875.894 ± 1.55792.907 ± 1.70276.847 ± 0.951
HaLRTC1.263 ± 0.0351.993 ± 0.0863.507 ± 0.17110.978† ± 0.17813.397 ± 0.34472.171 ± 3.29218.038 ± 0.50723.205 ± 0.868106.606 ± 1.809
LETC2.409 ± 0.0642.712 ± 0.1014.949 ± 0.14731.816 ± 0.98547.821 ± 1.57582.026 ± 2.44248.419 ± 0.67564.663 ± 0.896116.571 ± 1.279
WDGTC2.286 ± 0.1112.679 ± 0.14713.990 ± 0.21016.889 ± 0.80017.778 ± 0.64268.535 ± 1.79923.442 ± 0.23926.569 ± 0.519104.503 ± 3.344
GRIN0.874† ± 0.0141.424† ± 0.0382.885† ± 0.17011.273 ± 0.26215.158 ± 0.48327.541 ± 0.27017.629† ± 0.41323.273 ± 0.70156.956 ± 1.899
SPIN0.937 ± 0.0171.447 ± 0.0413.299 ± 0.20512.085 ± 0.25214.401 ± 0.43224.390 ± 1.31418.003 ± 0.34821.535 ± 0.74035.038 ± 0.754
ImputeFormer1.071 ± 0.0281.474 ± 0.0353.061 ± 0.14012.356 ± 0.25313.949 ± 0.40620.385† ± 2.16717.931 ± 0.48520.946† ± 0.29929.662† ± 1.111
MoHL0.872 ± 0.0141.422 ± 0.0362.697 ± 0.09410.974 ± 0.24013.400† ± 0.28516.719 ± 1.27416.991 ± 0.35319.747 ± 0.60725.928 ± 0.674
Improvement (%) ↑+0.23+0.14+6.52+0.04−0.02+17.98+3.62+5.72+12.59

Mean ± s.d. describes variation across test weeks. All entries average three mask seeds. Improvements are calculated from the printed means. Bold: lowest MAE; †: second lowest.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.