acceptodds
Under review as a conference paper at ICLR 2027

MorITS: Morphology-Aware Modeling for Irregular Time Series Forecasting

Abstract

Irregular multivariate time series (IMTS) are prevalent in healthcare, biomechanics, climate science, and astronomy. Observations occur at irregular intervals within variables and at misaligned timestamps across variables. Existing methods typically treat timestamps and missingness masks only as point-level auxiliary information, without explicitly modeling patch-level observation morphology jointly formed by temporal distributions and missingness patterns. This makes it difficult for local semantic extraction and cross-patch interactions to adapt to different observation structures. To address this issue, we propose **MorITS**, a morphology-aware framework for IMTS forecasting. The Observation Morphology Modeling module uses localized Gaussian basis functions to jointly model all timestamps and missingness masks within each patch, constructing morphology representations independent of observed values. The Morphology-Guided Expert Mixture module adaptively extracts patch semantics through morphology-conditioned expert routing and semantic aggregation. The Semantic–Morphology Patch Interaction module combines semantic similarity with inter-patch morphology relations to guide cross-patch interaction. MorITS achieves state-of-the-art forecasting performance on five real-world datasets.

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

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_HBspOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In order to better leverage the sampling pattern information in patch-based IMTSF methods, the paper proposes MorITS, which contains a sampling-pattern-guided MoE patch encoder layer. Specifically, Gaussian basis functions are used to model morphology, explicitly encoding masks and timestamps to guide MoE routing. Comprehensive experiments demonstrate that MorITS is competitive against SOTA baselines.

Critical strengths
  1. Motivation is clear, closely correlated to IMTSF scenarios. The essence of "morphology" is actually sampling pattern, where different sampling rates within and among variables lead to varying observation densities and patterns.
  2. Comprehensive experiments. Although it is based on the existing code pipeline PyOmniTS to achieve such comparisons, it is not deniable that experiemental setups are sound.
Critical weaknesses

These are not critical weaknesses, but rather some suggestions for improvements:

  1. "Morphology" is not commonly seen in IMTSF task. Based on my understanding, it essentially refers to sampling pattern. Readers not familiar with IMTS may not be able to quickly understand the term, and the confusion may hinder the understanding of methodology.
  2. The experimental setups follow PyOmniTS, but its citation is missing. Again, some readers may not understand what PyOmniTS refers to.
Questions for authors

Fairly speaking, the methodology is not that surprising, and one question remains: Why MorITS chooses Gaussian basis functions to model morphology? Compared to other types of networks, is there anything special making it better suitted for such task? For example, TTCN is used in tPatchGNN to encode patches, so why not TTCN?

3 comments
AuthorsOpenReview ↗

Thanking Reviewer HBsp and Welcoming Further Discussion

Thank you for your careful assessment and constructive suggestions, and for taking the time to read and consider our response. We hope our explanations of observation morphology, the source and role of PyOmniTS, and the rationale for Gaussian basis functions, along with comparisons against alternative encoders including TTCN, have addressed your concerns. Should any questions remain or further explanation be helpful, we would welcome your feedback whenever convenient. We appreciate your time and thoughtful consideration.

Thanks for your explanations and experiments upon Gaussian basis. Since my score is already a positive one, I'll keep my score as it is.

AuthorsOpenReview ↗

Response to Reviewer HBsp's Follow-up

Thank you for your follow-up and positive assessment of our work. We appreciate your time and constructive suggestions. We will clarify the definition of observation morphology, add the PyOmniTS citation and repository link, and incorporate the additional encoder comparisons into the revised manuscript.

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

The paper studies irregular multivariate time-series forecasting (IMTSF), where observations are recorded at irregular intervals and different variables may be observed at different timestamps. The proposed framework, MorITS, contains three main modules, followed by cross-variable interaction and a forecasting decoder.

Critical strengths

The paper proposes a meaningful limitation in conventional patch-based forecasting: two patches with similar observed values may have different temporal observation structures.

The proposed method seems effective in selected datasets.

Critical weaknesses

The framework combines several established design patterns: Gaussian basis functions for representing temporal distributions; sparse mixture-of-experts routing; attention-based aggregation conditioned on an auxiliary representation; pairwise-distance biases added to attention scores. The potentially novel contribution is their integration around the idea of patch-level observation morphology. That is a valid direction, but the paper needs to make a particularly clear case for what is fundamentally new beyond this combination.

OMM normalizes timestamps within each patch to a relative interval and represents the temporal distribution using Gaussian basis functions at uniformly spaced centers. Absolute temporal scale is discarded by the normalization. Two patches can have similar normalized observation positions even if one spans a much shorter physical time interval than the other. That difference can matter in applications where the rate of change or elapsed time has physical significance. The semantic pathway includes time embeddings, which may preserve some temporal information, so this is not necessarily a flaw in the complete model. It does, however, limit what OMM itself captures.

Questions for authors

See Weakness.

3 comments
AuthorsOpenReview ↗

With Thanks to Reviewer xDRi: Following Up on Our Response

We appreciate your thoughtful comments on the methodological contribution and temporal-scale design, as well as your time in considering our response. We have clarified the connections among the three modules and the complementary roles of relative and absolute time, supported by an additional temporal encoding experiment. We hope these explanations have addressed your concerns. If any points remain unclear or need further discussion, we would be grateful if you could let us know at your convenience. Thank you again for your careful consideration.

Thank you for the detailed response and the additional temporal-encoding experiment. I appreciate the effort you have put into clarifying the design. After reconsidering your replies, however, I will keep my rating for the reasons below.

On W2 (temporal scale) — this part of the response is constructive. The additional experiment comparing relative vs. absolute time in OMM is informative, and the role separation — OMM captures shared morphology on a relative scale while the semantic pathway preserves absolute temporal meaning — is a reasonable design rationale. This addresses my secondary concern, which I had already flagged as "not necessarily a flaw." (One minor note: the performance gaps are small and of the same order as the reported standard deviations, so a significance test would make the claim more robust; I do not treat this as blocking.)

On W1 (methodological novelty) — my primary concern remains. The response re-describes how the three modules are connected (OMM → MGEM → SMPI), but it does not establish what is fundamentally new beyond combining established components.

AuthorsOpenReview ↗

Response to Reviewer xDRi's Follow-up

Thank you for your response. We are pleased that these clarifications have addressed your concern about temporal scale in W2. Regarding your remaining concern about methodological novelty in W1, we would like to further clarify how our components differ from existing approaches in terms of how they are used, how they interact, and which problems they address.

Gaussian basis functions for representing temporal distributions. Existing Gaussian kernel-based content aggregation typically normalizes Gaussian weights before computing a weighted sum of values or features, as in IP-Net's smooth and non-smooth interpolation branches [1] and KAFNet's temporal kernel aggregation [2]. OMM retains the standard Gaussian kernel but accumulates responses only over valid observation positions, without multiplying by observed values, and normalizes by the fixed K rather than the sum of kernel weights. This preserves relative observation density, while responses at multiple centers characterize temporal distribution and local clustering patterns, meeting the need to jointly describe observation morphology.

Sparse mixture-of-experts routing. Existing time series MoE methods typically assign experts based on encoded token representations and then transform these token representations through expert networks, as in Time-MoE [3] and Moirai-MoE [4]. However, different combinations of morphology attributes entail different semantic extraction requirements; MGEM therefore uses independently parameterized experts to learn complementary extraction responses. The morphology representation is further mapped to metric weights that adjust the importance of each dimension when matching the semantic summary to expert prototypes, allowing expert selection to adapt to the current patch's observation morphology.

Attention-based aggregation conditioned on an auxiliary representation. Conditioned attention aggregates features relevant to a given condition: DA-RNN [5] selects historical information conditioned on the decoder state, while mTAN [6] aggregates observations conditioned on reference time. MGEM instead generates the Query from the overall observation morphology of the patch being aggregated, with observation-level semantic representations providing the Keys and Values. This is motivated by the fact that an observation, even with the same value and timestamp, may have a different semantic contribution when its surrounding observations are distributed differently. Aggregation therefore needs to adapt to the overall morphology of the current patch.

Pairwise-distance biases added to attention scores. Existing relative-position biases adjust attention based on positional relationships to supplement semantic matching with positional information, as in Earthformer [7] and ALiBi [8]. SMPI instead computes negative squared Euclidean distances between patch morphology representations in a learnable projection space and adds the resulting morphology relation scores to the semantic matching scores. Even with the same positional relationship, different observation layouts can produce different morphology relation scores. This distinguishes patches with similar semantics but different morphologies, jointly using semantic similarity and morphology relations to guide cross-patch interaction.

Although these basic components have precedents, MorITS introduces task-specific designs in the computations and coordination of observation statistics, expert matching, semantic aggregation, and patch interaction to address morphology dependence. These task-driven designs are the methodological contribution we hope to further clarify through the comparisons above.


References

[1] Satya Narayan Shukla and Benjamin M. Marlin. Interpolation-Prediction Networks for Irregularly Sampled Time Series. ICLR, 2019.

[2] Ziyu Zhou et al. Revitalizing Canonical Pre-Alignment for Irregular Multivariate Time Series Forecasting. AAAI, 2026.

[3] Xiaoming Shi et al. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. ICLR, 2025.

[4] Xu Liu et al. Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts. ICML, 2025.

[5] Yao Qin et al. A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction. IJCAI, 2017.

[6] Satya Narayan Shukla and Benjamin M. Marlin. Multi-Time Attention Networks for Irregularly Sampled Time Series. ICLR, 2021.

[7] Zhihan Gao et al. Earthformer: Exploring Space-Time Transformers for Earth System Forecasting. NeurIPS, 2022.

[8] Ofir Press et al. Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation. ICLR, 2022.

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

This paper proposes MorITS, a morphology-aware framework for irregular multivariate time series forecasting. It introduces an Observation Morphology Modeling module that uses localized Gaussian basis functions to capture temporal distributions and missingness patterns within each patch. A Morphology-Guided Expert Mixture module then leverages these representations to guide expert routing and semantic aggregation. Additionally, a Semantic–Morphology Patch Interaction module combines semantic similarity and morphology relations to model interactions across patches.

Critical strengths
  1. The overall framework is clearly presented and easy to follow.
  2. MorITS achieves state-of-the-art forecasting performance on five real-world datasets, demonstrating the effectiveness of the proposed approach.
  3. The experimental evaluation is comprehensive, including comparisons with 13 baselines, ablation studies, mechanism analysis, efficiency analysis, and hyperparameter sensitivity analysis.
Critical weaknesses
  1. Although Table 2 demonstrates the effectiveness of OMM through ablation studies, it remains unclear whether the proposed Gaussian-based morphology encoding is more effective than simpler alternatives. Could the authors compare it with other encoding strategies, such as s an MLP-based encoder, while keeping the remaining architecture unchanged? Such comparisons would help better justify the proposed design.
  2. It would be helpful to compare the current approach with alternative patching strategies, such as variable-wise patching, fixed-duration patching, or adaptive patching. Such comparisons would help clarify the impact of patch construction on the proposed morphology representations and forecasting performance.
  3. Although the experiments cover five real-world datasets from different domains, it remains unclear how the proposed method performs under different levels of observation irregularity. Such experiments would provide more direct evidence that the proposed morphology-aware design effectively addresses the challenges of irregular multivariate time series.
Questions for authors
  1. Although Table 2 demonstrates the effectiveness of OMM through ablation studies, it remains unclear whether the proposed Gaussian-based morphology encoding is more effective than simpler alternatives. Could the authors compare it with other encoding strategies, such as s an MLP-based encoder, while keeping the remaining architecture unchanged? Such comparisons would help better justify the proposed design.
  2. It would be helpful to compare the current approach with alternative patching strategies, such as variable-wise patching, fixed-duration patching, or adaptive patching. Such comparisons would help clarify the impact of patch construction on the proposed morphology representations and forecasting performance.
  3. Although the experiments cover five real-world datasets from different domains, it remains unclear how the proposed method performs under different levels of observation irregularity. Such experiments would provide more direct evidence that the proposed morphology-aware design effectively addresses the challenges of irregular multivariate time series.
1 comment
AuthorsOpenReview ↗

Appreciation and Follow-up Regarding Our Response to Reviewer XAUh

Thank you for your careful review and constructive assessment of our work, and for taking the time to consider our response. We hope the additional comparisons of morphology encoders and patching strategies, together with experiments at different observation sparsity levels, have addressed your concerns. At your convenience, we would appreciate hearing whether any points remain unclear or whether further clarification would be helpful. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.