acceptodds
Under review as a conference paper at ICLR 2027

PARSI: Parallel Temporal Modeling with Selective Interaction for Irregular Multivariate Time Series Forecasting

Abstract

Irregular multivariate time series (IMTS) arise when multiple variables are measured at different times and with varying gaps between observations, as commonly occurs in healthcare and finance. These irregularities make it difficult to model how individual variables evolve and how they relate to one another, giving rise to two key challenges: (1) preserving useful short- and long-range temporal patterns that a single temporal resolution or progressive aggregation may obscure, and (2) capturing changing cross-variable dependencies without computing full attention for every variable representation. To address these challenges, we propose PARSI, a forecasting framework that combines parallel temporal modeling with selective cross-variable interaction. First, PARSI encodes different lengths of each variable’s history in parallel and adaptively combines the resulting representations, preserving complementary temporal information. Second, it estimates whether cross-variable dependencies are concentrated on particular variables or distributed broadly, and uses this estimate to select representations for full cross-variable attention. The remaining representations receive a shared global summary, reducing computation while retaining access to information from other variables. Extensive experiments on real-world IMTS forecasting datasets demonstrate that PARSI achieves up to 10.1% relative improvement in forecasting accuracy over representative state-of-the-art baselines while reducing inference time by 25.0%.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Multivariate Time Series, Time Series Forecasting, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tCHuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PARSI for irregular multivariate time series forecasting. It uses parallel temporal branches to capture patterns from different historical ranges and an entropy-based selective attention mechanism to model cross-variable dependencies. Experiments on several IMTS benchmarks evaluate forecasting accuracy and computational efficiency.

Critical strengths
  1. The paper is generally clear and easy to follow.
  2. The experiments cover multiple IMTS datasets and recent baselines, with results reported over five runs.
Critical weaknesses
  1. The novelty is somewhat limited. Parallel temporal modeling and adaptive fusion have been explored in previous multi-scale forecasting methods. The selective interaction module is also closely related to ProbSparse attention. The differences from recent IMTS methods such as APN, Hi-Patch, and ReIMTS could be explained more clearly.

  2. The ablation study only tests what happens when the interaction module is removed. It does not compare entropy-based selection with random selection or full attention. These comparisons are needed to show whether the selection strategy actually helps.

  3. In Supplementary Table 16, selective attention is slower than full attention for up to 1024 variables. Table 17 also shows that PARSI is much slower than t-PatchGNN at inference on several datasets. The efficiency advantage is therefore less clear for the datasets used in this paper.

  4. Removing the cross-variable interaction module makes little difference on CESNET and ClusterTrace (Table 2). The changes are also small compared with the reported standard deviations, so the benefit of this module on these datasets is not convincing.

  5. The temporal ablation compares different ranges, but does not compare parallel modeling with progressive multi-scale modeling. Such a comparison would help show whether using parallel branches is better than existing approaches.

Questions for authors
  1. Does entropy-based query selection outperform random selection or full attention under the same settings?
  2. Why is selective attention slower than full attention at the variable sizes used in most experiments?
  3. What specific advantage does parallel temporal modeling offer over existing multi-scale IMTS methods?
5 comments
AuthorsOpenReview ↗

Response to W1–W2, Q1

Response to W1: Novelty of the Proposed Framework

We agree that multi-scale modeling, adaptive fusion, and sparse attention are established ideas. PARSI’s contribution is not these mechanisms individually, but how it preserves different historical ranges before compression and allocates cross-variable computation using the resulting representations. Furthermore, our contribution is their problem-driven adaptation and organization for IMTS forecasting. This architectural approach to methodological contribution is also illustrated by STanHop [1] and Hi-Patch [2].

Relationship to ProbSparse. PARSI shares ProbSparse’s principle of prioritizing queries with concentrated attention, and we will make this connection explicit [3]. ProbSparse selects temporal queries using an approximation based on the maximum and mean attention logits. PARSI uses sampled attention entropy to select variable states after multi-range temporal fusion. Selected states receive individualized cross-variable aggregation, while the others receive shared global context. The distinction is therefore in the selection criterion and its application to temporally summarized variable states, rather than a new general sparse-attention principle.

Differences from APN, Hi-Patch, and ReIMTS. APN [4] learns adaptive patch boundaries, while Hi-Patch and ReIMTS organize temporal representations hierarchically or recursively [5]. PARSI instead constructs independent temporal views directly from the same observation embeddings. Longer-range branches therefore do not depend on summaries already compressed by shorter-range branches. Feature-wise adaptive fusion then combines these complementary views into a representation for each variable. In our controlled comparison with progressive encoding, keeping adaptive fusion unchanged, parallel encoding reduces test MAE, test MSE, and inference latency by up to 26.97%, 37.46%, and 26.12%, respectively, across the evaluated settings.

Response to W2 and Q1: Effectiveness of Entropy-Based Query Selection

We thank the reviewer for this suggestion. We have added controlled comparisons with random selection and full attention across all 11 datasets, keeping the other PARSI components unchanged and using the same query budget for entropy-based and random selection.

When only a subset of queries is selected, entropy-based selection achieves lower mean MAE and MSE than random selection, supporting the benefit of informed selection. It also maintains comparable or better accuracy than full attention. On datasets where the budget covers all queries, the strategies reduce to full attention and produce identical results. These comparisons directly evaluate the selection strategy rather than only the presence of cross-variable interaction.

Note that entropy-based selection is one component of PARSI, not its sole contribution. As shown in Section 4.2 (Table 1) and runtime measurements in Section 4.4 (Figure 4), the careful design of PARSI’s architecture achieves better forecasting accuracy than several attention-based baselines while remaining computationally efficient.

Table 1. Forecasting performance of different cross-variable attention strategies (MAE/MSE, mean ± SD)

DatasetMetricEntropy-basedRandomFull
FNSPIDMAE0.0932±0.00200.0932±0.00200.0932±0.0020
MSE0.0284±0.00040.0284±0.00040.0284±0.0004
CESNETMAE0.6041±0.00200.6041±0.00200.6041±0.0020
MSE0.6748±0.00110.6748±0.00110.6748±0.0011
ClusterTraceMAE0.7548±0.00500.7548±0.00500.7548±0.0050
MSE1.3657±0.02431.3657±0.02431.3657±0.0243
USHCNMAE0.2645±0.00930.2645±0.00930.2645±0.0093
MSE0.1590±0.00360.1590±0.00360.1590±0.0036
BaodingMAE0.3435±0.00220.3465±0.00330.3466±0.0040
MSE0.2570±0.00210.2579±0.00230.2597±0.0018
ZhuzhouMAE0.3619±0.00210.3623±0.00220.3623±0.0029
MSE0.4045±0.00330.4058±0.00240.4026±0.0028
StudentLifeMAE0.5425±0.00350.5425±0.00350.5425±0.0035
MSE0.5848±0.00350.5848±0.00350.5848±0.0035
EPA-AirMAE0.4922±0.00690.4922±0.00690.4922±0.0069
MSE0.4245±0.00730.4245±0.00730.4245±0.0073
RepoHealthMAE0.6808±0.04600.6808±0.04600.6808±0.0460
MSE1.4938±0.15311.4938±0.15311.4938±0.1531
P12MAE0.3838±0.01490.3938±0.01450.3903±0.0110
MSE0.3239±0.01380.3303±0.01260.3301±0.0111
HumanActivityMAE0.1240±0.00010.1240±0.00010.1240±0.0001
MSE0.0460±0.00060.0460±0.00060.0460±0.0006
AuthorsOpenReview ↗

Response to W3–W4, Q2

Response to W3 and Q2: Computational Efficiency of Selective Attention

We thank the reviewer for highlighting this issue. At smaller variable counts, full attention is already inexpensive, so the additional query-selection overhead outweighs the computational savings. However, as Table 3 shows, The advantage of selective attention emerges with large number of variables. Specifically, our controlled comparison shows that its computation time grows more slowly than full attention, yielding a 62.18% reduction in attention-core time at 4,096 variables. This is important for high-dimensional forecasting applications, such as large sensor networks and infrastructure monitoring.

The comparison with t-PatchGNN concerns the cost of the complete architecture, rather than the attention module alone. PARSI prioritizes forecasting accuracy through parallel temporal modeling and adaptive fusion, at the cost of higher inference latency. But our proposed method PARSI achieves lower MSE than t-PatchGNN on 8 of 9 evaluated datasets, and our ablations support the contribution of these architectural choices.

Thus, the longer runtime should not negate the demonstrated forecasting benefits. PARSI offers an accuracy–cost trade-off, with selective attention helping control cross-variable computation as the number of variables grows.

Table 2. Forecasting accuracy and inference latency comparison with t-PatchGNN

DatasetPARSI MSE ↓t-PatchGNN MSE ↓PARSI inference (ms/step) ↓t-PatchGNN inference (ms/step) ↓
FNSPID0.02810.03887.811.71
CESNET0.67420.69148.971.69
ClusterTrace1.36551.874524.751.64
USHCN0.15060.177239.484.02

Table 3. Scalability comparison of three cross-variable interaction strategies (ms/forward, mean ± SD)

Variables NNSelective cross-variable attentionFull cross-variable attentionRandom state selection
1280.2811 ± 0.00430.1525 ± 0.00220.1710 ± 0.0007
2560.2844 ± 0.00170.1448 ± 0.00300.1702 ± 0.0005
5120.2801 ± 0.00510.1450 ± 0.00070.1739 ± 0.0051
10240.2828 ± 0.00140.1489 ± 0.00070.1734± 0.0007
20480.2910 ± 0.00150.2866 ± 0.00020.2084 ± 0.0023
40960.4112 ± 0.00131.0872 ± 0.00100.3674 ± 0.0005

Response to W4: Effectiveness of Cross-Variable Interaction

We thank the reviewer for this observation. Table 2 of the main paper evaluates the incremental contribution of each component within PARSI, rather than expecting every component to contribute equally on every dataset. The small gains on CESNET and ClusterTrace indicate a limited additional benefit from cross-variable interaction in these settings, not that the module is ineffective overall. For example, its removal has a larger effect on USHCN, particularly on MAE.

This dataset-dependent contribution is consistent with the relatively weak cross-variable correlations measured on CESNET and ClusterTrace. When cross-variable information provides limited additional predictive value, PARSI can derive most of its benefit from temporal modeling and adaptive fusion.

Thus, the ablations therefore clarify the complementary roles of PARSI’s components. Cross-variable interaction contributes differently across datasets, while the complete architecture achieves the lowest mean errors across the four datasets in Table 2 of the main paper. We will revise the discussion to distinguish this overall performance.

AuthorsOpenReview ↗

Response to W5 and Q3 : Advantages of Parallel Temporal Modeling

We thank the reviewer for suggesting this direct comparison. The main advantage of parallel temporal modeling is that each temporal branch independently accesses the original observations, preserving information across different temporal ranges before adaptive fusion. In contrast, progressive multi-scale methods such as Hi-Patch [2] and ReIMTS [5] construct representations sequentially, where later stages depend on previously aggregated information. This may compress useful temporal details before their relevance to forecasting is determined.

To isolate this design choice, we compare parallel and progressive temporal modeling within PARSI across 11 datasets and five paired seeds. As shown in Table 4 and 5, the results demonstrate the benefits of parallel temporal modeling under controlled conditions. Parallel encoding achieves lower MAE on 10 of 11 datasets and lower MSE on 7 of 11 datasets, while reducing inference latency on all 11 datasets.

Table 4. Forecasting Accuracy of Parallel and Progressive Temporal Modeling

DatasetParallel MAE ↓Progressive MAE ↓Parallel MSE ↓Progressive MSE ↓
FNSPID0.092053 ± 0.0025150.092221 ± 0.0020110.028007 ± 0.0004070.028008 ± 0.000591
StudentLife0.541387 ± 0.0009330.542591 ± 0.0015820.583221 ± 0.0012670.584695 ± 0.001152
EPA-Air0.481542 ± 0.0050160.485880 ± 0.0146270.413304 ± 0.0097730.417551 ± 0.020137
CESNET0.608555 ± 0.0007500.608712 ± 0.0010790.682584 ± 0.0006310.682034 ± 0.001351
ClusterTrace0.753991 ± 0.0036250.757944 ± 0.0032891.344334 ± 0.0157571.346721 ± 0.011809
RepoHealth0.674452 ± 0.0297150.627628 ± 0.0213481.520127 ± 0.0846611.336648 ± 0.090141
USHCN0.263026 ± 0.0060870.264626 ± 0.0103190.157706 ± 0.0153500.168104 ± 0.013427
P120.382088 ± 0.0068140.523227 ± 0.0867150.320642 ± 0.0051680.512712 ± 0.147759
HumanActivity0.122452 ± 0.0007650.122731 ± 0.0007540.045902 ± 0.0005020.044680 ± 0.000772
Baoding0.419466 ± 0.0014320.419524 ± 0.0013150.351095 ± 0.0004700.350945 ± 0.000421
Zhuzhou0.445357 ± 0.0020190.445816 ± 0.0020550.500849 ± 0.0021840.501106 ± 0.002000

Table 5. Inference Latency of Parallel and Progressive Temporal Modeling (ms/batch, mean ± SD)

DatasetParallel ↓Progressive ↓
FNSPID7.121 ± 0.1817.714 ± 0.259
StudentLife8.229 ± 0.44411.139 ± 1.159
EPA-Air25.443 ± 2.39526.687 ± 4.804
CESNET9.492 ± 1.05110.195 ± 0.920
ClusterTrace23.629 ± 0.54326.178 ± 3.866
RepoHealth11.179 ± 0.17812.194 ± 0.362
USHCN48.016 ± 4.10548.184 ± 4.732
P128.605 ± 0.1179.993 ± 1.338
HumanActivity16.776 ± 2.14017.411 ± 2.122
Baoding13.373 ± 0.17214.155 ± 0.110
Zhuzhou12.086 ± 0.50312.434 ± 0.102

Reference

[1] Wu, Y.-H., et al. STAnHop: Sparse Tandem Hopfield Model for Memory-Enhanced Time Series Prediction. ICLR, 2024.

[2] Luo, Y., et al. Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series. ICML, 2025.

[3] Zhou, H., et al. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. AAAI, 2021.

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

[5] Li, B., et al. Learning Recursive Multi-Scale Representations for Irregular Multivariate Time Series Forecasting. ICLR, 2026.

Reviewer_f34ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PARSI, a forecasting framework that combines parallel temporal modeling with selective cross-variable interaction. Existing irregular multivariate time series models struggle to preserve useful short- and long-range temporal patterns and capture changing cross-variable dependencies without computing full attention. To address this, PARSI models multiple temporal ranges of each variable's history in parallel and adaptively combines the resulting representations, preserving complementary temporal information. PARSI then estimates whether cross-variable dependencies are concentrated on specific variables or distributed broadly, and uses this estimate to select representations for full cross-variable attention. Extensive experiments on multiple IMTS forecasting datasets demonstrate that PARSI achieves competitive forecasting performance while reducing computational costs.

Critical strengths
  1. Preserving useful short- and long-range temporal patterns and avoiding the enormous computational cost of computing full attention for every variable representation are two key critical challenges in IMTS tasks. The authors propose two techniques to address these two problems, respectively, with a very clear modeling motivation.
  2. The authors compared PARSI with multiple different baselines across several different IMTS benchmarks, demonstrating its competitive performance. Additionally, efficiency comparison experiments further demonstrate that PARSI can effectively reduce the inference time. Ablation studies also confirm the necessity of multi-range temporal modeling and the complementary contribution of selective cross-variable interaction.
Critical weaknesses
  1. Limited novelty. This paper aims to address two challenges in current IMTS tasks: modeling short- and long-range temporal patterns and capturing changing cross-variable dependencies without computing full attention. Although the authors propose two solutions, the overall method appears to be a simple combination of two existing technical components rather than a tightly integrated framework. Furthermore, multi-temporal windows and attention mechanism for variable interaction modeling are not particularly novel in themselves, and the paper does not sufficiently clarify what fundamentally distinguishes its approach from existing methods. This limits the methodological novelty of this work.
  2. In the Selectively Cross-Variable Interaction section, the authors randomly sample SS keys and use them to determine whether each variable query should be modeled using full attention or mean context to capture variable interactions. However, is there a theoretical justification for using attention entropy as the selection criterion, or is this merely a heuristic approach choice? The authors need to provide a detailed explanation and justification for this design.
  3. The main text presents experimental results on only four datasets but the supplementary materials report results on eleven datasets. The authors should provide supplementary explanations in the Experimental Setup section to ensure consistency between the main text and the supplementary material and to improve the clarity of the experimental evaluation.
  4. In the Implementational Details section of the Supplementary Materials, the values for the Temporal Ranges parameter varies substantially across datasets. Is this the result determined by manual tuning for each dataset? If so, what search ranges and selection criteria were used? Or were the values chosen based on the data characteristics of each dataset? If so, the authors should provide an analysis to justify these choices. Such substantial variation in hyperparameter settings raises concerns about the generalization across different datasets.
  5. The bold formatting in several tables in the supplementary materials appears to be incorrect, including Tables 11 and 14. The authors should carefully check these and correct these minor errors.
Questions for authors
  1. In Figure 2, the performance of multi-temporal ranges is inferior to that of the medium or long range branches on some datasets. The authors should provide a detailed explanation for this observation. Does this indicate that multi-temporal ranges are only useful on certain datasets rather than being universally effective? Why does its performance degrade on some datasets, and is this related to the specific characteristics of those datasets?
  2. In Figure 4, PARSI incurs higher costs in both training and inference time than some competing methods, with both costs even being several times higher than those of the iPatchGNN and HyperIMTS. Please provide ablation experiments to identify the specific sources of these costs and conduct a detailed analysis of potential strategies to improve this.
  3. Selectively Cross-Variable Interaction section lacks a performance comparison and an efficiency analysis between selective attention and full attention in the experiments. If selective attention results in a substantial performance degradation compared to full attention, the effectiveness of selective attention would need to be reconsidered.
  4. The number of sampled keys SS and lowest-entropy queries KsK_s are two key parameters in the Selective Cross-Variable Interaction module; however, the paper lacks sensitivity analysis experiments for these two parameters and an evaluation of how their values affect model efficiency. The supplementary materials use a unified value for SS and KsK_s, but was this choice based on random selection or determined through experimental validation? The authors should explain this choice or provide sensitivity analysis to justify it.
5 comments
AuthorsOpenReview ↗

Response to W1-W5: Novelty, entropy rationale, evaluation and tuning

Response to W1: Novelty of the Proposed Framework

We thank the reviewer for the comment. (1) PARSI goes beyond placing two existing modules in sequence. It adapts and organizes them to address three linked challenges in IMTS forecasting: preserving temporal patterns under irregular sampling, accommodating different temporal needs across variables, and avoiding unnecessary cross-variable computation. (2) We do not claim temporal windows or attention mechanisms themselves as novel. Our contribution is their problem-driven adaptation and organization for IMTS forecasting. This architectural approach to methodological contribution is also illustrated by STanHop [1] and Hi-Patch [2]. (3) PARSI differs in when temporal information is combined and where detailed interaction is performed. It learns temporal views independently before adaptive fusion, rather than through progressive aggregation. Moreover, it uses the fused states to selectively allocate full attention rather than applying it to every state. The practical value of these design choices is supported by the ablations and controlled accuracy–efficiency comparisons in Sections 4.2–4.4 of the main paper (Tables 1–2 and Figures 2 and 4).

Response to W2: Theoretical Motivation of Entropy-Based Selection

We thank the reviewer for the comment. Our entropy-based selection is a theoretically motivated heuristic, grounded in the approximation error between individualized aggregation and mean context. For attention weights aia_{i}, the triangle inequality and Pinsker’s inequality give

∥∑j=1Naijvj−v‾∥2≤M2(ln⁡N−H(ai)),\left\lVert \sum_{j = 1}^{N}a_{\text{ij}}v_{j} - \overline{v} \right\rVert_{2} \leq M\sqrt{2\left( \ln N - H\left( a_{i} \right) \right)},

where v‾=1N∑j=1Nvj\overline{v} = \frac{1}{N}\sum_{j = 1}^{N}v_{j}, M=max⁡j∥vj−v‾∥2M = \max_{j}\lVert v_j - \overline{v} \rVert_2, and H(ai)H\left( a_{i} \right) denotes attention entropy.

When entropy approaches its maximum, mean context closely approximates individualized aggregation. Lower entropy permits a larger approximation error, motivating the allocation of individualized aggregation to low-entropy states.

Response to W3: Clarification of Experimental Evaluation Scope

Thank you for pointing out that the scope of the evaluation could be clearer. The main paper presents four representative datasets. The supplement reports 11 evaluation settings in total, including the two ASeer subsets. We will add the following text to the Experimental Setup (Section 4.1): “We evaluate PARSI on 11 settings in total, comprising six TIME-IMM datasets, three traditional IMTS benchmarks, and the Baoding and Zhuzhou subsets of ASeer. Table 1 presents results on four representative datasets: FNSPID, CESNET, ClusterTrace, and USHCN. Supplementary Tables 7–10 report the complete forecasting results for all 11 settings.”

Response to W4: Temporal Range Configuration and Cross-Dataset Generalization

We thank the reviewer for the comment. The temporal-range configurations were selected through dataset-specific hyperparameter tuning. Historical length and observation patterns guided the initial candidates. Validation performance determined the final selection; the settings were not determined by data characteristics alone.

(1) Search and selection. We evaluated actual candidate configurations dataset-specific sets of temporal window count triplets using a grid. We selected the configuration with the lowest mean validation MSE. We will add these details alongside the final settings in Table 4. These settings denote window counts, not absolute temporal durations.

(2) Cross-dataset applicability. PARSI uses the same three-branch design and window construction mechanism on all 11 datasets. Only the temporal window counts are adjusted. This demonstrates applicability across datasets with validation-based tuning. To directly assess sensitivity to these choices, we examined window-count searches across eleven datasets and found that validation loss varied to different degrees across the tested configurations..

Thank you for flagging the boldface in Supplementary Tables 11 and 14. We will correct it.

AuthorsOpenReview ↗

Response to Q1: Explaining Performance Differences Between Single-Range and Multi-Range Modeling

Figure 2 shows that the benefit of combining ranges is dataset-dependent, not that multi-range modeling is generally ineffective. Combining ranges helps when they capture complementary predictive patterns. When one range already captures the most useful information, additional ranges may introduce redundancy or noise and slightly reduce accuracy. Differences in observation density and temporal persistence can influence this balance.

Our motivation is therefore not that more ranges are always better, but that no single range consistently suits all datasets. PARSI preserves different temporal views and learns their relative contributions instead of relying on one range alone. This design outperforms the best single-range variant on a majority of datasets in Table 1. Figure 2 here only identifies where additional ranges offer less benefit.

Table 1. Forecasting performance under different temporal range configurations.

DatasetShort MAEShort MSEMedium MAEMedium MSELong MAELong MSEMulti MAEMulti MSE
P120.3720 ± 0.00080.3133 ± 0.00120.3704 ± 0.00280.3106 ± 0.00120.3730 ± 0.00110.3164 ± 0.00120.3642 ± 0.00170.3071 ± 0.0004
HumanActivity0.1294 ± 0.00110.0483 ± 0.00040.1250 ± 0.00180.0465 ± 0.00070.1214 ± 0.00080.0470 ± 0.00060.1200 ± 0.00150.0447 ± 0.0009
USHCN0.2725 ± 0.02780.1781 ± 0.04290.2661 ± 0.01950.1497 ± 0.00800.2730 ± 0.00890.1591 ± 0.01270.2482 ± 0.01390.1506 ± 0.0093
EPA-Air0.5077 ± 0.00780.4535 ± 0.01750.4852 ± 0.01130.4130 ± 0.01550.4770 ± 0.00300.3998 ± 0.00390.4833 ± 0.00250.4099 ± 0.0056
CESNET0.6066 ± 0.00250.6783 ± 0.00230.6051 ± 0.00170.6768 ± 0.00120.6058 ± 0.00110.6783 ± 0.00100.6033 ± 0.00220.6742 ± 0.0017
ClusterTrace0.7984 ± 0.00641.4952 ± 0.02860.7876 ± 0.00261.4667 ± 0.02180.7765 ± 0.00341.4407 ± 0.02850.7557 ± 0.00551.3655 ± 0.0270
RepoHealth0.6834 ± 0.03931.5703 ± 0.13920.7089 ± 0.03951.6225 ± 0.13990.7184 ± 0.02211.6954 ± 0.12230.6746 ± 0.02671.5043 ± 0.0949
Baoding0.3450 ± 0.00400.2609 ± 0.00120.3470 ± 0.00470.2617 ± 0.00140.3460 ± 0.00290.2620 ± 0.00150.3432 ± 0.00180.2573 ± 0.0018
Zhuzhou0.3631 ± 0.00140.4101 ± 0.00410.3635 ± 0.00250.4058 ± 0.00710.3629 ± 0.00230.4092 ± 0.00420.3628 ± 0.00200.4071 ± 0.0029
FNSPID0.0942 ± 0.00130.0291 ± 0.00050.0919 ± 0.00270.0276 ± 0.00020.0913 ± 0.00240.0277 ± 0.00050.0922 ± 0.00270.0281 ± 0.0005
StudentLife0.5444 ± 0.00420.5855 ± 0.00480.5414 ± 0.00100.5814 ± 0.00060.5436 ± 0.00390.5839 ± 0.00370.5420 ± 0.00270.5829 ± 0.0010
AuthorsOpenReview ↗

Response to Q2: Computational Cost Analysis

We thank the reviewer for this comment. PARSI’s higher runtime mainly reflects its observation-level encoding and independent processing of three temporal ranges before adaptive fusion. Our component-wise profiling (Table 2) shows that these stages account for approximately 85–96% of the measured forward-component time. Backpropagation adds further training cost, while selective cross-variable interaction and decoding contribute relatively little.

These findings suggest possible improvements: fusing observation-embedding operations to reduce execution overhead, and batching similar-length temporal windows to improve parallelism and reduce padding. These optimizations could accelerate training and inference while preserving the multi-range design.

Although slower than t-PatchGNN and HyperIMTS, PARSI achieves better forecasting accuracy (Table 3) and generally less GPU memory consumption (Table 4), offering an accuracy–runtime trade-off rather than universal speed superiority.

Table 2. Component-wise training-step time breakdown across 11 datasets (ms)

DatasetObservation encodingTemporal modeling & fusionVariable interactionDecoderBackpropagationOptimizer
FNSPID6.7298.7302.0150.63013.7531.580
StudentLife8.7769.3882.2450.69516.0701.642
EPA-Air24.5325.5961.3420.41014.9980.920
CESNET7.7239.4632.2500.70316.2981.628
ClusterTrace19.7366.5521.5510.49016.0191.137
RepoHealth8.4078.9622.1750.66715.6421.525
USHCN37.3205.1151.2280.36416.4150.987
P129.3438.7692.1160.63715.7921.339
HumanActivity18.7696.9361.5420.50516.1791.212
Baoding16.0117.3691.7530.58916.6601.379
Zhuzhou14.2397.8281.8580.60016.3031.418

Table 3. Forecasting Accuracy Comparison.

DatasetMethodMAEMSE
FNSPIDPARSI0.0922 ± 0.00270.0281 ± 0.0005
FNSPIDt-PatchGNN0.1214 ± 0.00890.0388 ± 0.0038
FNSPIDHyperIMTS0.1139 ± 0.00510.0367 ± 0.0030
CESNETPARSI0.6033 ± 0.00220.6742 ± 0.0017
CESNETt-PatchGNN0.6136 ± 0.00180.6914 ± 0.0025
CESNETHyperIMTS0.6146 ± 0.00170.6911 ± 0.0006
ClusterTracePARSI0.7557 ± 0.00551.3655 ± 0.0270
ClusterTracet-PatchGNN0.8858 ± 0.03371.8745 ± 0.1366
ClusterTraceHyperIMTS0.8993 ± 0.01811.7419 ± 0.0786
USHCNPARSI0.2482 ± 0.01390.1506 ± 0.0093
USHCNt-PatchGNN0.2877 ± 0.02450.1772 ± 0.0243
USHCNHyperIMTS0.2811 ± 0.02650.1775 ± 0.0121

Table 4. Computational Cost Comparison.

DatasetMethodPeak GPU Memory (GiB)Inference Latency (ms/step)
FNSPIDPARSI0.02617.8147
FNSPIDt-PatchGNN0.03011.7137
FNSPIDHyperIMTS0.12185.2456
CESNETPARSI0.05418.9717
CESNETt-PatchGNN0.03181.6861
CESNETHyperIMTS0.24145.3502
ClusterTracePARSI0.108424.7474
ClusterTracet-PatchGNN0.08401.6428
ClusterTraceHyperIMTS0.18735.2813
USHCNPARSI0.059439.4796
USHCNt-PatchGNN1.10214.0228
USHCNHyperIMTS0.38754.3598
Reviewer_bLmZOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PARSI, a framework for irregular multivariate time series forecasting, aiming to model temporal dependencies across different historical ranges while reducing the computational cost of cross-variable interactions.

Specifically, PARSI first employs multiple parallel temporal branches to perform local attention modeling within different temporal window ranges and uses feature-wise adaptive fusion to obtain variable-level representations. Subsequently, the model selects certain variable states for full cross-variable attention based on sampled attention entropy, while the unselected states use a shared global mean representation. Finally, a query-conditioned decoder generates predictions based on future timestamps.

Critical strengths

1: The motivation of the paper is relatively clear. It considers both the preservation of information across different temporal ranges and the computational cost of cross-variable interactions, and the overall framework is easy to understand.

2: The method can directly handle irregular and asynchronous observation timestamps without forcing all variables to align to a common time grid.

Critical weaknesses

W1. The design of selective cross-variable interaction lacks direct and sufficient validation.

The paper presents entropy-based variable state selection as one of its main contributions, but Table 2 only reports the results of removing the entire cross-variable interaction module, without directly comparing different interaction and selection strategies.

In particular, removing cross-variable interaction leads to only very limited performance degradation on FNSPID, CESNET, and ClusterTrace.

W2. The core claims regarding computational efficiency require further experiments.

Although the paper emphasizes reducing computational costs through selective attention, it does not directly compare the forecasting accuracy, inference time, and memory usage of PARSI using full cross-variable attention versus the proposed selective mechanism.

W3. The novelty of the core method is incremental.

This paper mainly combines parallel multi-granularity temporal modeling, adaptive feature fusion, and selective cross-variable attention. The overall novelty is relatively incremental.

W4. Lack of an illustrative motivation figure.

Questions for authors

Q1. Could the authors provide direct comparisons between selective cross-variable interaction and full cross-variable attention, random state selection, and using only the global mean context? In particular, I would like to see the differences in forecasting performance, runtime, and memory usage among these designs.

Q2. Why are variable states with lower sampled attention entropy selected for full attention computation? Is there experimental evidence showing that low-entropy states can obtain greater predictive benefits from individualized cross-variable aggregation compared to random selection?

Q3. I suggest that the authors report the number of variables, the selection budget coefficient, and the actual selection ratio for each dataset. In particular, for low-dimensional datasets, could the selection budget cover all or most variables, causing the selective mechanism to degenerate into full attention?

Q4. I suggest that the authors further report the computational complexity and actual runtime under different numbers of variables, as well as the accuracy–efficiency trade-off between full attention and selective attention.

7 comments
AuthorsOpenReview ↗

Response to Q1

Thank you for the suggestion. We report the comparisons in Tables 1–4. To isolate the contribution of selective cross-variable interaction, we replace only this component while keeping PARSI’s parallel temporal modeling, adaptive fusion, and forecasting decoder unchanged.

Table 1 shows that our method achieves the best or tied-best MAE and MSE on 6/11 datasets. Global mean context performs better on some datasets, possibly because averaging preserves shared patterns while reducing noise. When particular states carry more useful information, however, averaging can dilute it. Our improvements over all alternatives on ASeer (Baoding) and P12 support the benefit of informed selection, without implying that it is universally preferable.

This benefit comes with limited computational overhead. Table 2 shows that selection adds less than 0.25 ms on average relative to each alternative. Although selection introduces additional work at small scale, restricting interactions becomes beneficial as the variable count grows, reducing attention-core computation time by 62.18% at 4,096 variables relative to full attention (Table 3). The memory results show a similar trade-off, with small absolute increases over full attention on seven datasets but reductions on four, reaching 89.09% (Table 4). Random selection remains faster but is not consistently more memory-efficient, while global averaging is both faster and smaller. Neither simpler alternative, however, consistently matches our forecasting accuracy.

Together, these results support PARSI’s combination of temporal representation learning and selective interaction, rather than a claim that selection always performs best. Parallel branches preserve different historical ranges, adaptive fusion combines them, and selective interaction determines which resulting states receive individualized cross-variable context. Because all compared variants retain PARSI’s temporal modeling, the results also clarify when simpler interaction suffices within the framework. Automatically choosing among interaction designs is a useful extension, but requires a separately validated selection rule balancing accuracy and computational cost, beyond the controlled comparison here.

Table 1. Forecasting performance

DatasetMetricSelectiveFullRandomGlobal mean
FNSPIDMAE0.0932±0.00200.0932±0.00200.0932±0.00200.0929±0.0020
MSE0.0284±0.00040.0284±0.00040.0284±0.00040.0284±0.0005
CESNETMAE0.6041±0.00200.6041±0.00200.6041±0.00200.6070±0.0017
MSE0.6748±0.00110.6748±0.00110.6748±0.00110.6790±0.0016
ClusterTraceMAE0.7548±0.00500.7548±0.00500.7548±0.00500.7668±0.0239
MSE1.3657±0.02431.3657±0.02431.3657±0.02431.3983±0.0712
USHCNMAE0.2645±0.00930.2645±0.00930.2645±0.00930.2578±0.0190
MSE0.1590±0.00360.1590±0.00360.1590±0.00360.1580±0.0118
ASeer (Baoding)MAE0.3435±0.00220.3466±0.00400.3465±0.00330.3453±0.0014
MSE0.2570±0.00210.2597±0.00180.2579±0.00230.2594±0.0018
ASeer (Zhuzhou)MAE0.3619±0.00210.3623±0.00290.3623±0.00220.3628±0.0025
MSE0.4045±0.00330.4026±0.00280.4058±0.00240.4076±0.0023
StudentLifeMAE0.5425±0.00350.5425±0.00350.5425±0.00350.5436±0.0033
MSE0.5848±0.00350.5848±0.00350.5848±0.00350.5848±0.0040
EPA-AirMAE0.4922±0.00690.4922±0.00690.4922±0.00690.4908±0.0062
MSE0.4245±0.00730.4245±0.00730.4245±0.00730.4216±0.0088
RepoHealthMAE0.6808±0.04600.6808±0.04600.6808±0.04600.6930±0.0414
MSE1.4938±0.15311.4938±0.15311.4938±0.15311.5764±0.1358
P12MAE0.3838±0.01490.3903±0.01100.3938±0.01450.3857±0.0145
MSE0.3239±0.01380.3301±0.01110.3303±0.01260.3267±0.0149
HumanActivityMAE0.1240±0.00010.1240±0.00010.1240±0.00010.1236±0.0008
MSE0.0460±0.00060.0460±0.00060.0460±0.00060.0461±0.0006
AuthorsOpenReview ↗

Q1 continued — Latency and memory

Table 2. Inference latency (ms, mean ± SD)

DatasetSelective cross-variable attentionFull cross-variable attentionRandom state selectionGlobal mean context
FNSPID4.5133 ± 0.09054.2867 ± 0.13654.4010 ± 0.07714.2866 ± 0.0952
StudentLife4.3698 ± 0.02234.1679 ± 0.03314.2566 ± 0.04314.1562 ± 0.0341
EPA-Air4.3343 ± 0.04164.1792 ± 0.01914.2422 ± 0.04004.1238 ± 0.0260
CESNET4.3652 ± 0.02724.1640 ± 0.01304.2516 ± 0.02034.1278 ± 0.0154
ClusterTrace4.4509 ± 0.05044.2530 ± 0.03324.3373 ± 0.02714.1971 ± 0.0487
RepoHealth4.4100 ± 0.02574.2313 ± 0.03944.3137 ± 0.04094.1690 ± 0.0483
USHCN4.4985 ± 0.04454.3051 ± 0.05364.3816 ± 0.02534.2590 ± 0.0523
P124.2419 ± 0.02124.0707 ± 0.03204.1558 ± 0.03903.9806 ± 0.0382
HumanActivity4.6320 ± 0.11444.4274 ± 0.08824.4952 ± 0.08894.3710 ± 0.1189
Baoding4.4452 ± 0.02784.2783 ± 0.03144.3504 ± 0.01704.1953 ± 0.0288
Zhuzhou4.6548 ± 0.12794.4596 ± 0.10984.5329 ± 0.12994.3788 ± 0.1142

Table 3. Scalability comparison of four cross-variable interaction strategies (ms/forward, mean ± SD)

Variables NNSelective cross-variable attentionFull cross-variable attentionRandom state selectionGlobal mean context
1280.2811 ± 0.00430.1525 ± 0.00220.1710 ± 0.00070.0216 ± 0.0001
2560.2844 ± 0.00170.1448 ± 0.00300.1702 ± 0.00050.0216 ± 0.0001
5120.2801 ± 0.00510.1450 ± 0.00070.1739 ± 0.00510.0216 ± 0.0002
10240.2828 ± 0.00140.1489 ± 0.00070.1734± 0.00070.0227 ± 0.0015
20480.2910 ± 0.00150.2866 ± 0.00020.2084 ± 0.00230.0270 ± 0.0030
40960.4112 ± 0.00131.0872 ± 0.00100.3674 ± 0.00050.0298 ± 0.0052

Table 4. Additional peak GPU memory within the attention core (MiB)

DatasetSelective cross-variable attentionFull cross-variable attentionRandom state selectionGlobal mean context
FNSPID0.024900.022460.034180.01025
StudentLife0.030760.028320.040530.00977
EPA-Air0.020020.017580.028320.00977
CESNET0.075680.071290.105960.03223
ClusterTrace0.062010.058590.084470.02344
RepoHealth0.035640.032230.045410.01074
USHCN0.029790.020510.029300.00488
P120.762210.843750.595700.10840
HumanActivity0.222660.233400.328610.10156
Baoding0.250001.192380.203610.06494
Zhuzhou0.673346.168950.507320.15186
AuthorsOpenReview ↗

Response to Q2: The Rationale and Effectiveness of Low-Entropy Variable State Selection

We select states with lower sampled attention entropy because their attention is concentrated on particular variables, suggesting that some cross-variable relationships are especially relevant to these states. Computing full attention for them allows these relationships to be evaluated using all variables, rather than only the sampled variables used to estimate entropy.

Table 5 shows this prioritization benefits prediction, as Low-entropy selection produces larger forecasting gains than random selection on 9 of 11 datasets.

Table 5. Test MSE reduction relative to mean-only context with fixed checkpoints and matched query budgets (Ks<NK_s < N).

DatasetLow entropyRandom
FNSPID-0.000090-0.000066
StudentLife+0.000296+0.000152
EPA-Air+0.003569+0.001645
CESNET+0.006346+0.004733
ClusterTrace+0.002377+0.001162
RepoHealth-0.001777-0.022676
USHCN+0.000052-0.000287
P12+0.000714+0.000188
HumanActivity-0.000166-0.000069
Baoding+0.001287+0.001249
Zhuzhou+0.002944+0.000143

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.