acceptodds
Under review as a conference paper at ICLR 2027

SPAQR: Sampling-Aware Patch Allocation and Query-conditioned Retrieval for IMTS Forecasting

Abstract

Irregular multivariate time series (IMTS) forecasting is challenging because observation intervals are uneven within variables and timestamps are unaligned across variables. Patch-based methods can preserve the original observations and their timestamps, yet they commonly determine partitions without allocating a shared patch budget according to cross-variable sampling heterogeneity. Moreover, multi-scale views are typically constructed before a target variable-time query identifies the historical evidence required for a forecast. To this end, we propose SPAQR, Sampling-aware Patch Allocation and Query-conditioned Retrieval for IMTS Forecasting. SPAQR first performs sampling-aware patch allocation, distributing a fixed per-sample patch budget across variables and feasible temporal splits according to observation density, temporal span, and internal gaps. It then applies multi-scale support encoding to retain coarse, boundary, and transition evidence within the resulting fixed supports. Finally, query-conditioned evidence retrieval uses each target variable and forecast time to read global context and softly weight the supports and their internal states without modifying the partition. Across four public IMTS benchmarks and sixteen representative baselines, SPAQR achieves the lowest mean MSE and MAE on every dataset under a common protocol. Ablations and sensitivity studies further characterize the behavior of its allocation and retrieval designs.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_v6YbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPAQR for irregular multivariate time-series forecasting, combining sampling-aware patch allocation with query-conditioned retrieval. The approach distributes a limited patch budget among variables based on their sampling patterns and represents each patch using coarse, boundary, and transition information. For each forecasting query, the model identifies and combines relevant historical information based on the target variable and prediction time. Experiments on four benchmark datasets against sixteen baseline methods show that SPAQR achieves the lowest mean MSE and MAE across all datasets. The ablation studies further support the effectiveness of its main components.

Critical strengths

1- The paper focuses on an important challenge in irregular time-series forecasting, where variables often have different sampling patterns and assigning the same number of patches to each variable may not be effective. Allocating a limited number of patches based on these sampling differences is an interesting and practical idea.

2- The idea of capturing different types of temporal information within each patch and selecting the relevant information based on the target variable and forecasting time is interesting. This allows the model to use the same historical data differently for different forecasting queries, rather than relying on a single fixed representation.

3- The approach is evaluated on four irregular time-series benchmarks against sixteen baseline methods, including several recent approaches. SPAQR achieves the lowest average forecasting errors across all four datasets. The ablation and sensitivity studies further demonstrate the effectiveness of the proposed framework and its main components.

Critical weaknesses

1- The proposed sampling-aware patch allocation is interesting, but the scoring function in Eq. (5) could be better justified. Although the authors use equal weights for observation count, temporal span, and sampling gaps, these factors have different numerical ranges and may not contribute equally to the final allocation score. I would suggest including an ablation study to examine the effect of each term individually and in combination. This would help explain which factors are most important for improving forecasting performance and whether all three are needed.

2- Query-conditioned retrieval is one of the main contributions of the paper. However, the proposed weighting mechanism in Eqs. (12)–(14) appears similar to standard attention, where relevant information is selected based on a query. It is therefore not clear what additional benefit the proposed retrieval mechanism provides over conventional cross-attention. Also, the ablation without query-conditioned retrieval still uses query-dependent weighting across patches, so it does not fully remove query conditioning. I would suggest comparing the proposed method with standard cross-attention using the same patch representations and experimental settings. This would help demonstrate the specific advantage of the proposed retrieval mechanism.

3- Can the authors provide a more detailed analysis of forecasting performance under different sampling densities and irregularity levels to directly demonstrate the advantages of sampling-aware allocation?

Questions for authors

Please address my all concerns raised in the weaknesses section.

5 comments

We thank the reviewer for recognizing the motivation for sampling-aware allocation and query-dependent use of historical evidence. Following your suggestions, we conducted three sets of additional experiments and clarify what each comparison establishes below.

W1: Rationale and contributions of the allocation-score terms.

1. Equal coefficients do not imply equal numerical contributions.

We agree. The coefficients in Eq. (5) specify a default weighting. The normalization and cap control the scales of the three terms, but do not equalize their numerical ranges or contributions.

2. Each term expresses a different split preference.

The three terms in Eq. (5) capture different aspects of sampling:

  • Mass favors balanced observation counts on the two sides of a split.
  • Span favors cuts that reduce squared temporal extent.
  • Gap favors boundaries at unusually large internal gaps relative to the variable’s typical gap.

3. The full combination performs best in the score-term ablation.

We evaluated all seven nonempty combinations. The support budget, per-variable cap, multi-scale support encoding, query-conditioned evidence retrieval, and training settings were unchanged. The full combination achieves the lowest mean MSE and MAE on all four datasets.

Allocation scoreHumanActivity MSEHumanActivity MAEUSHCN MSEUSHCN MAEPhysioNet’12 MSEPhysioNet’12 MAEMIMIC MSEMIMIC MAE
Mass0.0427 ± 0.00040.1182 ± 0.00110.1709 ± 0.00260.2705 ± 0.01620.3015 ± 0.00100.3608 ± 0.00200.4125 ± 0.00550.3832 ± 0.0018
Span0.0420 ± 0.00040.1175 ± 0.00070.1770 ± 0.01050.2713 ± 0.00700.2999 ± 0.00170.3599 ± 0.00090.4112 ± 0.00720.3815 ± 0.0021
Gap0.0477 ± 0.00650.1315 ± 0.01370.1871 ± 0.00890.2758 ± 0.00380.3001 ± 0.00220.3597 ± 0.00190.4130 ± 0.00560.3841 ± 0.0026
Mass + span0.0424 ± 0.00060.1180 ± 0.00140.1864 ± 0.00640.2805 ± 0.01170.2972 ± 0.00210.3561 ± 0.00490.4239 ± 0.00450.3809 ± 0.0027
Mass + gap0.0435 ± 0.00070.1208 ± 0.00220.1831 ± 0.00330.2781 ± 0.01240.3003 ± 0.00080.3616 ± 0.00230.4110 ± 0.00430.3817 ± 0.0023
Span + gap0.0451 ± 0.00500.1260 ± 0.01240.1693 ± 0.00820.2797 ± 0.02760.3010 ± 0.00110.3598 ± 0.00180.4098 ± 0.00340.3851 ± 0.0054
Mass + span + gap (SPAQR)0.0409 ± 0.00070.1054 ± 0.00090.1582 ± 0.00090.2549 ± 0.00560.2899 ± 0.00160.3489 ± 0.00180.3956 ± 0.00180.3780 ± 0.0023

W2: Comparison with standard cross-attention and interpretation of the existing ablation.

1. The design concerns the compatibility score.

We agree that SPAQR's query-conditioned evidence retrieval is attention-based. Its score combines a coarse-support content term shared by all four states with state-specific content, temporal anchors, geometry, and target–source variable biases.

Using the notation of Eqs. (12)–(13), let aj,n,pa_{j,n,p} denote the coarse-support content score plus bj,n,psupb^{\mathrm{sup}}_{j,n,p}. Then, over valid support–state pairs,

βj,n,pαj,n,p,r=softmax⁡(n,p,r)(aj,n,p+ℓj,n,p,r).\beta_{j,n,p}\alpha_{j,n,p,r}=\operatorname{softmax}_{(n,p,r)}\left(a_{j,n,p}+\ell_{j,n,p,r}\right).

The factor 1/41/4 in Eq. (13) cancels because every support has four states. This equivalence makes the relationship to standard attention explicit: our contribution is the construction of the support–state compatibility score, not a new normalization operation.

2. The existing ablation simplifies state reading.

The w/o query-conditioned retrieval does not remove all query conditioning. It uniformly averages the four states and removes the state-evidence term from the support score, while retaining query-dependent support weighting.

3. New cross-attention controls retain the four-state representation.

We replaced only the final joint support–state retrieval with standard dot-product cross-attention over all four states of every valid support. We also evaluated cross-attention + metadata, which adds the six-dimensional support geometry, source-variable identity, and state identity to the key features. The encoder architecture and training settings were unchanged, and each model was trained end to end.

Final retrievalHumanActivity MSEHumanActivity MAEUSHCN MSEUSHCN MAEPhysioNet’12 MSEPhysioNet’12 MAEMIMIC MSEMIMIC MAE
Standard cross-attention0.0585 ± 0.01220.1587 ± 0.02440.1997 ± 0.01910.2834 ± 0.00460.3125 ± 0.00150.3738 ± 0.00400.4174 ± 0.00420.3863 ± 0.0055
Cross-attention + metadata0.0506 ± 0.00410.1416 ± 0.01040.1869 ± 0.00530.2771 ± 0.00620.3124 ± 0.00170.3719 ± 0.00240.4274 ± 0.00480.3856 ± 0.0028
SPAQR0.0409 ± 0.00070.1054 ± 0.00090.1582 ± 0.00090.2549 ± 0.00560.2899 ± 0.00160.3489 ± 0.00180.3956 ± 0.00180.3780 ± 0.0023

W3: Performance across sampling densities and irregularity levels.

1. Controlled changes to test-time sampling.

We compared sampling-aware patch allocation with Equal time from Appendix B.3, retaining the same encoding and retrieval architecture. Both models were trained on the original training data and evaluated without retraining. For each test sample, they received the same retained observations; forecast times and targets were unchanged. We also preserved the first and last observation of each nonempty variable.

2. Density, internal gaps, and variable imbalance.

  • Density (50% nominal retention): Uniformly thin observations within each variable. Endpoint preservation and rounding may increase the realized retention fraction.
  • Internal gaps (strength 1.0): Concentrate removals near a randomly chosen interior time, keeping the same per-variable retained counts as uniform thinning.
  • Variable imbalance (strength 1.0): Vary retention quotas across variables while keeping the same total retained count per sample as uniform thinning.

3. Results under sampling perturbations.

The table reports MSE and MAE for the original history and three representative perturbations. Across the three reported perturbations, SPAQR achieves lower mean MSE on all four datasets and lower mean MAE in 10 of 12 comparisons.

ConditionAllocationHumanActivity MSEHumanActivity MAEUSHCN MSEUSHCN MAEPhysioNet’12 MSEPhysioNet’12 MAEMIMIC MSEMIMIC MAE
Original historySPAQR0.0409 ± 0.00070.1054 ± 0.00090.1582 ± 0.00090.2549 ± 0.00560.2899 ± 0.00160.3489 ± 0.00180.3956 ± 0.00180.3780 ± 0.0023
Original historyEqual time0.0419 ± 0.00040.1179 ± 0.00080.1979 ± 0.02250.2745 ± 0.00770.3030 ± 0.00250.3640 ± 0.00490.4177 ± 0.00370.3861 ± 0.0049
Uniform retentionSPAQR0.0493 ± 0.00130.1392 ± 0.00270.1814 ± 0.01390.2732 ± 0.00840.3076 ± 0.00330.3661 ± 0.00270.4130 ± 0.00460.3884 ± 0.0055
Uniform retentionEqual time0.0508 ± 0.00400.1375 ± 0.00560.1964 ± 0.01560.2746 ± 0.00640.3086 ± 0.00150.3676 ± 0.00230.4142 ± 0.00420.3896 ± 0.0053
Gap concentrationSPAQR0.0525 ± 0.00110.1436 ± 0.00350.1913 ± 0.01700.2785 ± 0.01350.3097 ± 0.00290.3660 ± 0.00240.4248 ± 0.00460.3913 ± 0.0053
Gap concentrationEqual time0.0533 ± 0.00510.1436 ± 0.00740.2111 ± 0.02320.2862 ± 0.01130.3131 ± 0.00250.3712 ± 0.00320.4258 ± 0.00330.3925 ± 0.0065
Variable imbalanceSPAQR0.0661 ± 0.00320.1635 ± 0.00340.1936 ± 0.02150.2853 ± 0.01440.3179 ± 0.00480.3709 ± 0.00320.4246 ± 0.00440.3890 ± 0.0051
Variable imbalanceEqual time0.0693 ± 0.02130.1659 ± 0.01980.2167 ± 0.02660.2893 ± 0.00830.3211 ± 0.00290.3774 ± 0.00310.4287 ± 0.00360.3936 ± 0.0069
Reviewer_ebvBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPAQR, a framework for irregular multivariate time series (IMTS) forecasting that combines sampling-aware patch allocation, multi-scale support encoding, and query-conditioned evidence retrieval.

Specifically, SPAQR allocates a fixed patch budget across variables and temporal segments using a greedy splitting strategy based on observation density, temporal span, and internal gaps. Each resulting patch is then encoded into four complementary representations capturing coarse context, boundary information, and temporal transitions. Finally, a query-conditioned retrieval module uses the target variable and forecasting timestamp to aggregate relevant historical evidence.

Experiments on four IMTS benchmarks demonstrate competitive forecasting performance against sixteen baselines. The paper also provides ablation studies and sensitivity analyses to examine the proposed components.

Critical strengths
  • I find the idea of decomposing historical information into distinct representations and letting each forecast query determine their contributions appealing. Broad context, boundary conditions, and temporal changes may play different roles in forecasting; retaining them separately is a worthwhile direction.
  • The design addresses irregular sampling explicitly. Patch allocation accounts for differences in sampling patterns across variables, while the encoder uses elapsed time. The allocation, encoding, and retrieval stages have clear roles.
  • The reported results are consistent across the four benchmarks. The full model achieves strong average performance, and the existing ablations show degradation when its components are simplified, providing initial support for the overall approach.
Critical weaknesses
  • The code repository is inaccessible, limiting verification of the implementation and experimental setup. I currently cannot access the repository linked in the paper. This prevents me from checking the state construction, query scoring, and gated updates against the method description and experimental configurations. I hope the authors can restore access during rebuttal or provide an anonymous code supplement with the necessary execution instructions.
  • The rationale and individual contributions of the four states remain insufficiently established. Section 4.3 describes their construction, but the distinct value of the left and right boundary states for future prediction is unclear, particularly given that the transition state already includes endpoint information. The coarse-only ablation reduces both the number of states and the types of information retained, so it does not isolate the benefit of this specific semantic decomposition from increased representation capacity. I would appreciate a clearer justification and any existing analyses or ablations that directly support this choice.
Questions for authors
  1. Why are these four states the appropriate decomposition? Please clarify the distinct roles of the left and right boundary states in future prediction and how they complement the endpoint information already present in the transition state.

  2. Do any existing ablations or analyses distinguish the benefit of the specific state roles from that of additional representation capacity or explicit endpoint and slope features? Please provide the most direct available evidence and clarify what the current experiments establish.

  3. Could you restore access to the anonymous repository during rebuttal, or provide an anonymous code supplement? Please include the configurations and execution instructions corresponding to the main results, and identify the code implementing state construction and query-conditioned retrieval.

3 comments

We thank the reviewer for recognizing the motivation for retaining distinct support states and the consistency of the experimental results.

W1 & Q3: Anonymous code access and implementation details.

We apologize for the access issue. The anonymous repository is now accessible.

The relevant implementations are in layers/SPAQR/support_unfolding.py:

  • State construction: MultiScaleSupportFieldEncoder, using TimeScaleStateScan.
  • Query-conditioned evidence retrieval: QueryConditionedSupportUnfoldingDecoder.
  • Gated updates: GatedStateUpdate.

The scripts below contain the dataset-specific configurations. Follow the environment setup in README.md, then run the appropriate script from the repository root:

bash scripts/SPAQR/HumanActivity.sh
bash scripts/SPAQR/USHCN.sh
bash scripts/SPAQR/P12.sh
bash scripts/SPAQR/MIMIC_III.sh

The three unrestricted datasets are downloaded and preprocessed automatically. MIMIC-III requires separately authorized data access.

W2 & Q1: Roles of the four support states.

The four states provide broad context, boundary conditions, and local change from the same historical support. They are constructed at two learned time scales (Section 4.3):

  1. Coarse state: broad context. It combines the two coarse directional scans with support geometry and a variable embedding.

  2. Left-boundary state: earlier local conditions. It combines the fine and coarse backward scans ending at the support start. This gives the query an earlier reference condition for interpreting within-patch change.

  3. Right-boundary state: recent local conditions. It combines the fine and coarse forward scans ending at the support end. The query can thus read the most recent conditions within that support for future prediction.

  4. Transition state: within-patch change. It combines both scales and directions with explicit dynamics, including the first and last values, their difference, and a signed log-slope.

Complementarity with the transition state. The transition state combines endpoint values with other dynamic information. Separate boundary states let a query read either boundary condition directly, without recovering it from that combined representation.

Their temporal anchors also differ (Section 4.4). The left-boundary and right-boundary states use the support start and end, respectively; the coarse and transition states use the support centre. All four states use only observations within the historical support.

W2 & Q2: Evidence from the existing ablations and new controls.

We have added two cross-attention controls to examine retrieval while retaining the four-state construction and explicit endpoint and slope features. We distinguish this evidence from the coarse-only representation ablation below.

1. New controls: same four-state encoder architecture and features, different retrieval.

Both controls attend to all four dd-dimensional states of every valid support. Standard cross-attention replaces only the final joint support–state retrieval; the second control additionally receives support geometry, source-variable identity, and state identity as key features. The results show that SPAQR has lower mean MSE and MAE than both controls on all four datasets.

Final retrievalHumanActivity MSEHumanActivity MAEUSHCN MSEUSHCN MAEPhysioNet’12 MSEPhysioNet’12 MAEMIMIC MSEMIMIC MAE
Standard cross-attention0.0585 ± 0.01220.1587 ± 0.02440.1997 ± 0.01910.2834 ± 0.00460.3125 ± 0.00150.3738 ± 0.00400.4174 ± 0.00420.3863 ± 0.0055
Cross-attention + metadata0.0506 ± 0.00410.1416 ± 0.01040.1869 ± 0.00530.2771 ± 0.00620.3124 ± 0.00170.3719 ± 0.00240.4274 ± 0.00480.3856 ± 0.0028
SPAQR0.0409 ± 0.00070.1054 ± 0.00090.1582 ± 0.00090.2549 ± 0.00560.2899 ± 0.00160.3489 ± 0.00180.3956 ± 0.00180.3780 ± 0.0023

2. Existing ablation: uniform reading of the four states.

The w/o query-conditioned retrieval variant retains all four states and their explicit dynamics, averages the states uniformly, and removes the state-evidence term from support scores. Query-dependent support weighting remains. Full SPAQR's lower mean MSE and MAE across all four datasets provide further evidence for query-conditioned state reading (Tables 2–3).

3. Existing ablation: the full support representation as a whole.

The w/o multi-scale encoding variant retains only the coarse state, changing the state count, fine-scale encoding, and explicit dynamics together. Its higher mean errors support the full representation as a whole (Tables 2–3). This comparison does not separate the contribution of each state role.

Reviewer_ebvBOpenReview ↗

Follow up Comments

Thank the authors for the detailed response and additional experiments. These efforts have addressed my main concerns. I will maintain my original positive rating.

Thank you for your positive feedback and continued support. We are glad that our response and additional experiments have addressed your main concerns.

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

In order to address the different sampling rate among different varaibles in IMTS datasets when using patches, the paper proposes SPAQR, covering sampling-aware patching and query-conditioned retrieval. Comprehensive experiments demonstrate state-of-the-art performance against baselines.

Critical strengths
  1. The paper is overall well-motivated. Existing methods are indeed short at considering the difference in sampling rate among varaibles.
  2. Comprehensive experiments. Although the code pipeline is directly taken from existing papers, the benchmarking results are likely to be replicated by authors themselves, which still need considerable time and effort.
Critical weaknesses
  1. My biggest concern is about the rationality of methodology, specifically sampling-aware patching. Judging from Figure 2 (a), the method seems to allocate the patches with varing lengths. Since both PrimeNet and ReIMTS adopt the fix-time-length approach to retain the sampling density, and the varying-length approach in this paper lacks proper theoretical justifications, I'm afraid the method itself is not persuasive enough.
  2. Minor problems: Some notation is inconsistent or unclear, like “support” and “patch” are used somewhat interchangeably, “Physio’12” and “PhysioNet’12” are inconsistent.
Questions for authors

See weaknesses.

3 comments

We thank the reviewer for recognizing the motivation and empirical evaluation of SPAQR.

W1: Rationale for sampling-aware patch allocation.

We appreciate this concern. Allowing support durations to vary does not discard the original timestamps or observation counts. Appendix B.3 compares SPAQR with Equal time, keeping the remaining components unchanged. SPAQR achieves lower mean MSE and MAE on all four datasets (Tables 6–7), supporting the effectiveness of sampling-aware patch allocation. We explain the rationale in three parts below:

1. Preserving sampling information.

Sampling-aware patch allocation retains the original observations and their timestamps, assigning each observation to exactly one support (Section 4.2). Sampling information remains available to the subsequent stages:

  • Multi-scale support encoding uses absolute and support-relative time features and accounts for elapsed time in the state scan (Eqs. 7–8).
  • Support geometry records the mean, first, and last timestamps, together with the observation count, span, and largest internal gap. The last three entries are log-transformed (Eq. 6).

Because observation count is retained together with temporal span, the model can distinguish dense supports from sparse ones even when their durations differ. The original timestamps and largest internal gap also retain information about uneven spacing within each support.

2. Allocating representation capacity under a fixed budget.

Fixed-duration patches can contain very different numbers of observations across variables. Giving every variable the same number of patches may therefore assign too little capacity to dense variables and more than needed to sparse ones.

SPAQR allocates supports under a shared per-sample budget cap BB. Every non-empty variable first receives one support. The remaining slots are assigned greedily across all variables and feasible cuts, until the budget is reached or no feasible split remains (Section 4.2).

APN [1] also learns adaptive temporal windows for patch aggregation, as discussed in Section 2.2. SPAQR specifically addresses how a fixed per-sample patch budget is allocated across variables according to their sampling patterns.

3. Rationale for the split score.

The three terms in Eq. (5) use different sampling properties to guide each split:

  • Observation count: favors balanced counts when splitting a given support. The identity (mℓ+mr)2−mℓ2−mr2=2mℓmr(m_\ell+m_r)^2-m_\ell^2-m_r^2=2m_\ell m_r explains the numerator of this term. Eq. (5) scales it by λm/m02\lambda_m/m_0^2, where m0m_0 is the mass-score scale.
  • Temporal span: favors cuts that reduce an excessively long temporal span.
  • Internal gap: favors placing a boundary at a large internal gap.

At each step, SPAQR selects the feasible cut with the largest Gn,a,c,bG_{n,a,c,b} under the budget and per-variable cap in Eq. (4). This produces a sample-specific hard partition using only timestamps and observation masks.

W2: Terminology and dataset naming.

In Section 4.1, each allocated patch is termed a support. The two words refer to the same unit: patch emphasizes the partition of observations, whereas support emphasizes the fixed observation set used for encoding and query-conditioned evidence retrieval. Its membership and boundaries remain unchanged across forecast queries; the four states are representations encoded from that support.

The inconsistent dataset naming was our oversight. We have standardized the dataset name to PhysioNet’12 throughout the manuscript. The revised content has been highlighted in red in the updated manuscript.

[1] Liu, X., et al; "Rethinking Irregular Time Series Forecasting: A Simple Yet Effective Baseline"; AAAI 2026.

Thanks for your explanations, which solve my concerns. I will raise my score accordingly.

We thank the reviewer for carefully considering our response and reassessing our work. We are glad that our clarifications helped address the concerns raised.

Reviewer_t1UFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPAQR, a forecasting framework for irregular multivariate time series (IMTS) that decouples representation construction from query-specific evidence reading. It introduces a sampling-aware patch allocation scheme that greedily distributes a fixed support (patch) budget across variables based on observation density, temporal span, and internal gaps, then encodes each support at two learned time scales into four complementary states (coarse, left/right boundary, transition). A query-conditioned retrieval mechanism attends over supports and their internal states using target variable/time and geometry-aware biases. Across four benchmarks (HumanActivity, USHCN, PhysioNet’12, MIMIC-III) and sixteen baselines, SPAQR achieves the best mean MSE/MAE, with ablations showing each stage contributes to performance.

Critical strengths
  1. The approach is architecturally simple enough to be reusable or integrated with other backbones yet yields consistent gains across diverse domains.

  2. Comprehensive comparisons across four standard IMTS datasets and sixteen representative baselines spanning recurrent, continuous-time, graph, and (adaptive) patch-based families.

Critical weaknesses
  1. Query-conditioned retrieval uses bias terms for geometry and variable-pair relevance but lacks deeper structural modeling of cross-variable dependencies (e.g., explicit graphs/hypergraphs), which could matter in highly asynchronous settings.

  2. Several strong recent IMTS baselines are absent in the main results, notably HyperIMTS and MissHyper (hypergraph family) and recent foundation-model baselines tailored to irregular data (e.g., QiYao-I) under fine-tuning or strong zero/few-shot protocols. This somewhat weakens the universality of the SOTA claim.

Questions for authors

Why were HyperIMTS and MissHyper omitted from the main results? If available, please add comparisons or at least discuss expected performance trade-offs.

1 comment

We thank the reviewer for recognizing SPAQR’s architectural simplicity and empirical performance.

W1: Cross-variable dependencies under asynchronous sampling.

We appreciate this concern. SPAQR combines cross-variable information through query updates that read supports across all variables. We clarify the role of the biases, the handling of asynchronous observations, and the comparison with HyperIMTS below:

  1. Rationale for geometry and variable-pair biases. Supports with similar content can differ in when they were observed, how they were sampled, and their relevance to the target variable. The geometry bias explicitly uses query-to-support temporal relations and the log-transformed count, span, and maximum-gap entries of the support geometry vector. The variable-pair bias learns target–source variable relevance. These biases complement content similarity when selecting evidence for a particular forecast (Section 4.4).

  2. Handling highly asynchronous observations. SPAQR retains each variable's original timestamps and uses elapsed time in multi-scale support encoding (Section 4.3). Each target variable–time query can then attend to the coarse states of all valid supports across variables. Evidence from different variables can therefore contribute to the same forecast without requiring observations to share a timestamp.

  3. Comparison with HyperIMTS. HyperIMTS propagates information through explicit temporal and variable hyperedges. In SPAQR, stacked coarse-global query layers read coarse states across variables and update the query. The context collected in one layer guides the next layer's attention, allowing cross-variable information to influence subsequent retrieval. The updated query then controls the final joint support–state weighting for the target variable and forecast time. The encoded support representation remains shared across queries (Section 4.4).

W2 & Q1: Missing baselines.

We agree that these are relevant baselines. We have added a comparison with HyperIMTS and clarify the status of the other two methods below.

We evaluated the HyperIMTS [1] implementation from the official PyOmniTS GitHub repository using the same data splits, preprocessing, history and forecast windows, and metric computation as SPAQR.

MethodHumanActivity MSEHumanActivity MAEUSHCN MSEUSHCN MAEPhysioNet’12 MSEPhysioNet’12 MAEMIMIC MSEMIMIC MAE
HyperIMTS0.0414 ± 0.00020.1144 ± 0.00090.1916 ± 0.03650.2688 ± 0.00850.2992 ± 0.00050.3593 ± 0.00080.3966 ± 0.01460.3738 ± 0.0079
SPAQR0.0409 ± 0.00070.1054 ± 0.00090.1582 ± 0.00090.2549 ± 0.00560.2899 ± 0.00160.3489 ± 0.00180.3956 ± 0.00180.3780 ± 0.0023

For MissHyper [2], we could not locate a public implementation in the paper or through GitHub searches, so we have not reproduced it under our evaluation protocol.

For QiYao-I [3], the anonymous repository provides core model code, but its README states that pretrained checkpoints will be released upon acceptance. We therefore cannot yet evaluate the reported pretrained model under our protocol.

[1] Li, B., et al; "HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting"; ICML 2025.

[2] Ma, M., and Tan, Q.; "MissHyper: Restoring Clinical Synchronicity in Missingness-Guided Hypergraph Forecasting"; arXiv:2607.21922, 2026.

[3] Wang, L., et al; "QiYao-I: A Manifold Based Foundation Model for Irregular Multivariate Time Series Forecasting"; arXiv:2610.06936, 2026.

Comments on the paper

1 comment

Dear Reviewers t1UF, RsBJ, ebvB, and v6Yb,

The authors have responded to your questions. Please review their responses and consider whether your concerns have been addressed. You may also read the other reviewers' comments and the revised manuscript and share any further feedback.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.