acceptodds
Under review as a conference paper at ICLR 2027

Repackaging Temporal Evidence: A Unifying Interface for Temporal Prediction

Abstract

Temporal tabular prediction and irregular time-series prediction often arise from similar longitudinal records, but they expose different evidence at query time. Temporal tables usually present a timestamped covariate row and evaluate generalization to future periods, whereas irregular time-series protocols present entity-specific observations and query values at selected times. We study this gap as a protocol issue rather than a modality boundary. We introduce a unified prediction interface that specifies the query-time information set and the prediction rule applied to it, and use it to diagnose protocol mismatches in cross-domain reuse. The interface highlights two common failures: entity collapse, where a temporal table is forced into a single sequence across unrelated entities, and temporal relation mismatch, where tabulation hides the within-sequence structure expected by a time-series model. Empirically, we use the interface as a diagnostic tool in both directions. A retrieval-based tabular predictor can be evaluated on irregular time-series tasks once candidate sets are repackaged as temporally embedded, entity-specific contexts. Conversely, selected temporal tables admit time-series evaluations when recoverable entity histories or source sequences are restored before modeling. These results highlight repackaging temporal evidence as a way to enable cross-domain reuse, without implying that either model family subsumes the other.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Tabular Machine Learning, Temporal Prediction, Irregular Time Series, Structured Data

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_dz5HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors describe that the temporal tabular and irregular time-series data are currently considered different modalities but can support cross-domain model reuse when underlying information is available. Therefore, they propose a framework to repackage available temporal tabular tasks into irregular time-series tasks and vice-versa.

The authors highlight entity collapse and temporal relation mismatch when converting from temporal tabular to irregular time-series. By recovering entity-specific histories, the authors enable irregular time series models to operate on temporal tabular tasks. Temporal embeddings and entity specific context adapt the tabular model ModernNCA to be applied to irregular time-series datasets.

Critical strengths

S1 Interesting framework for temporal tabular and irregular time-series. It is addressing entity collapse and temporal relation mismatch compared to a naive cross-domain transfer.

S2 Practical demonstration of the proposed protocol through adapting ModernNCA to irregular time-series.

Critical weaknesses

Broader evaluation

While the authors acknowledge that they are currently limiting tabular-to-time-series recovery to EO and WE in section 6.2, the limited evaluation of tabular to time-series recovery also limits the contribution of this paper. The paper’s evaluation should be extended to broader irregular time-series and temporal tabular data to support its applicability across domains, ideally including cases where underlying histories are incomplete. That would provide a more rigorous evaluation of the framework and its applicability and boundaries. Without this broader evaluation or clearer criteria for when cross-domain model reuse is valid, the contribution appears modest.

Questions for authors

Q1 Could you please report the standard deviations and confidence intervals for the 15 random seeds of all reported results?

Q2 Could you please provide ModernNCA (+c) in table 3 as a control to the rationale provided in section 5.2?

Q3 Could you please include further datasets for evaluation and derive applicability criteria?

2 comments

Thank you for carefully reviewing our paper and for your constructive suggestions on the scope of the experiments and the reporting of results.

W1/Q3: While the authors acknowledge that they are currently limiting tabular-to-time-series recovery to EO and WE in section 6.2, the limited evaluation of tabular to time-series recovery also limits the contribution of this paper. Could you please include further datasets for evaluation and derive applicability criteria?

We acknowledge that we have used only EO and WE as representative cases to study the use of irregular time-series models on temporal tabular tasks, and that the empirical coverage in this direction is still limited. The two cases examine different protocol issues. WE examines entity collapse caused by interleaving records from different stations. EO examines temporal relation mismatch caused by row-level summaries that hide the source transaction sequences. Other temporal tabular tasks can be converted to the input format of irregular time-series models, but the resulting input may be a poor fit for these models and lead to weak performance. For how we identify this mismatch, please see our response to Q1 from Reviewer F1kq.

We believe this is largely a limitation of the datasets themselves. When information about the meaning of the records is missing or the entity information is unclear, it can be inherently difficult to convert a dataset into a form that an irregular time-series model can use. The three irregular time-series datasets in Table 3 examine adaptation in the reverse direction using the tabular model ModernNCA. The adapted context-based tabular model is competitive on both tabular and time-series tasks, which may suggest that this approach is a more general way to handle such tasks. We will further explore datasets from the BeyondArena benchmark.

Q1: Could you please report the standard deviations and confidence intervals for the 15 random seeds of all reported results?

We will add the standard deviations shown below. Please also see the significance test results in our response to W2 from Reviewer 4Whv.

MethodPhysioNet interpolationUSHCN interpolationPhysioNet extrapolationUSHCN extrapolation
CRU0.00314020.00081830.00381510.0060423
ACSSM0.00255190.00156140.00243240.0002117
Latent ODE0.00049810.00007810.00155900.0001572
Neural Flows0.00494970.00029740.00292770.0002500
mTAND0.00188600.00087140.00316140.0075999
ModernNCA (+ec)0.00005360.00019090.00001690.0001003

These results show that ModernNCA yields more stable predictions, and the appendix also shows that it has lower computational cost.

Q2: Could you please provide ModernNCA (+c) in table 3 as a control to the rationale provided in section 5.2?

Table 3 currently reports only (+ec) and does not include (+c). In temporal tabular tasks, the model has input features in addition to timestamps and can make predictions using those features alone, ignoring timestamps entirely. Thus, (+e) and (+m) are optional, and (+c) can be used on its own. In irregular time-series extrapolation and interpolation tasks, however, a tabular model needs the timestamp (or its encoding) as input to make a prediction, and often there are no other features. In this setting, (+c) would mean “using information from the historical window while ignoring timestamp information” and cannot be used on its own.

Thanks for providing the additional information and explanations.

However, given the limitation described in W1/Q3, I will keep my rating.

Reviewer_4WhvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers both datasets with timestamped items (e.g., TabRed's real estate price prediction) and time-series datasets (e.g., USHCN's daily climate record prediction). It compares time-series prediction models (that incorporate the series directly in their architecture) against series-independent models (that have no notion of a series in their architecture). It adapts a series-independent model (ModernNCA) to time-series prediction by adding a temporal embedding to encode the timestamp and incorporating temporally nearby series points via weighted voting. It adapts time-series prediction models (GRU-D, CRU, etc.) to timestamped item prediction by separating out per-entity histories. On some datasets, the adapted "wrong" approach achieves comparable performance to the "right" approach.

Critical strengths
  • Shows that for both time series prediction/interpolation and timestamped item prediction, both time-series models and series-independent models can and should be considered as baselines (even though historically these have often been thought as non-comparable tasks and thus the models inappropriate as baselines).
Critical weaknesses
  • The writing in the first 3 pages of the article is difficult to follow, as no concrete examples of the problems under consideration are given and all definitions of the problems and approaches are in abstract and often confusing terms. Giving concrete examples of tasks that count as "temporal tabular prediction" vs. "irregular time-series prediction" early on would have significantly aided readability. It seems like Figure 1 was supposed to help with this, but most of Figure 1 is also abstract and/or undefined, since none of the y-axis, 𝛙(t), yellow diamond, blue star, or green shading are explained.
  • Many of the observed performance differences are small, but no significance testing is included.
Questions for authors

Minor clarification:

  • What does PFN mean?

Minor writing suggestion (not a question):

  • Consider reordering Table 4 to put WE above EO since you talk about WE in the text before you talk about EO in the text.
2 comments

Thank you for your careful review of our paper and for your specific suggestions on the writing of the introduction and on statistical testing.

W1: The writing in the first 3 pages of the article is difficult to follow, as no concrete examples of the problems under consideration are given and all definitions of the problems and approaches are in abstract and often confusing terms. Giving concrete examples of tasks that count as "temporal tabular prediction" vs. "irregular time-series prediction" early on would have significantly aided readability.

Thank you for pointing out exactly where the paper was difficult to follow. The current introduction presents the unified interface first, while the concrete examples of station histories and transaction histories for coupon offers appear only in §6. The caption of Figure 1 also does not fully explain the visual symbols. We will add a concrete example of the input, query, and target early in the introduction. Using the same real-world task, we will show what evidence is available from a single row of covariates and what evidence is available from an entity-specific history before introducing the formal definitions. We will also define the y-axis, ψ(t), the yellow diamond, the blue star, and the green shading in Figure 1, and explain at first use that ψ(t) is the temporal embedding. This way, readers can first understand the concrete prediction problem and then see how different protocols organize the evidence.

W2: Many of the observed performance differences are small, but no significance testing is included.

Please see our response to Q1 from Reviewer dz5H. We will also add the following significance test results comparing ModernNCA (+ec) with the baselines in Table 3.

Taskvs. ACSSMvs. CRUvs. Latent ODEvs. Neural Flowsvs. mTAND
PhysioNet interpolation<0.001<0.001<0.0010.00145<0.001, mTAND performs better
USHCN interpolation0.00723<0.001<0.001<0.001<0.001
USHCN extrapolation<0.0010.00145<0.001<0.001<0.001

Minor clarification: What does PFN mean?

We apologize for the confusion. PFN stands for Prior-Data Fitted Network and refers to a class of models that learn through pretraining to use labeled examples in the context to make predictions for new queries. The TabPFN family is a representative example for tabular tasks. Its prediction procedure is similar to that of ModernNCA in that changing the context can change the model’s predictive behavior. With pretraining, it achieves stronger performance. We will add this explanation where PFN first appears in the main text.

Minor writing suggestion (not a question): Consider reordering Table 4 to put WE above EO since you talk about WE in the text before you talk about EO in the text.

Thank you for this suggestion! We will reorder Table 4 so that WE appears before EO, following the order of the discussion in §6.1 and §6.2.

Thanks for providing some significance testing. You should report what type of significance test you conducted. Significance tests are also needed for HyperIMTS, GraFITi, and Hi-Parch (the top performers in Table 3) and for everything in Tables 2 and 4.

Reviewer_F1kqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified interface for temporal tabular prediction and irregular timeseries prediction. This interface separates the query-time information sets during the query from the predictor, and uses the evidence-packing operators to describe the available information for each query. Based on this interface, the paper identifies two protocol mismatch problems, namely entity collapse and temporal relation mismatch. Subsequently, the article discusses the realignment methods from temporal tabular to irregular timeseries, and from timeseries to temporal tabular. Experiments on datasets such as TabReD, PhysioNet, USHCN, Human Activity, Weather, and Ecom-Offers show that appropriate evidence re-encapsulation can improve the performance of specific tasks.

Critical strengths

S1. This paper presents a unified protocol perspective, making the query-time information sets and evidence-packaging operators clearly visible, thereby facilitating the diagnosis of the cross-domain migration failure issues in temporal prediction.

S2. This paper studies the bidirectional conversion between temporal tabular and irregular timeseries under the unified protocol, and provides evidence indicating that restoring the specific history can improve the time prediction performance on selected benchmarks.

Critical weaknesses

W1. The proposed framework lacks significant innovation in terms of methodology. This paper builds upon existing research, including the time embedding method in temporal tabular (Cai & Ye, 2025a, b), temporal tabular prediction (Gorishniy et al., 2024; Ye et al., 2025), and irregular timeseries prediction (Che et al., 2018; Rubanova et al., 2019; Shukla & Marlin, 2021). The main contribution lies in integrating temporal tabular and irregular timeseries prediction into a unified evidence encapsulation interface and providing their bidirectional conversion. However, the paper does not explain what new insights the unified protocol framework can bring.

W2. The experiments failed to clearly demonstrate the impact of the protocol. In Ecom-Offers, the paper replaced the summary features with pre-offer transaction sequences, thereby changing the input information and its temporal structure. In ModernNCA (+ec), both time embedding and entity-specific context construction were applied simultaneously. Therefore, the reported benefits do not clearly indicate whether the improvement mainly comes from the protocol and query-time information sets, or other individual components. These influencing factors should be distinguished through comparative analysis.

Questions for authors

Q1. How are the two protocol mismatches identified in practice?

Q2. What rules guide the choice of protocol changes for a given task?

3 comments

Thank you for carefully reviewing our work and for your constructive suggestions on how to explain the protocol diagnostics more clearly and improve the controlled comparisons.

W1: The methodological novelty of the proposed framework is limited. The main contribution is to bring these techniques and related protocol changes together under a unified evidence-packaging interface. However, the paper does not clearly show what new diagnostic rules or insights the interface provides beyond describing these existing methods in a common form.

We thank the reviewer for asking us to clarify what the interface adds. Temporal embeddings, retrieval-based tabular predictors, and the irregular time-series architectures we use all come from prior work. Our intended contribution is a set of practical protocol diagnostics and their validation. Temporal tabular prediction and irregular time-series prediction were previously viewed as two fields with no interaction. We bring them together under one interface and, on this basis, clarify that temporal tabular methods are well suited to covariate-centric tasks and irregular time-series methods to history-centric tasks. This insight connects the two fields and clarifies the strengths of each. Methods from the two fields use different prediction procedures to solve the same task, which can inform future method design. Finally, we find that context-based tabular methods achieve competitive performance on both tabular classification and regression and time-series interpolation and extrapolation, indicating that this approach applies more broadly across task types and that its capabilities generalize more widely.

More specifically, our subsequent analysis on WE shows that even when station ID is provided as a feature, global row ordering still causes the temporal state to evolve across stations. What therefore needs to be recovered is the entity-specific history associated with that state. EO shows that ordering rows by coupon issue time does not expose the temporal relations within the transaction history behind each coupon. The input unit identified by the diagnostic is the source transaction sequence. For the context-based tabular model ModernNCA, temporal embedding determines how time is represented, and entity-specific context determines which observations can be used. These correspond to timestamp handling and the context window in time-series methods, respectively.

We will incorporate these insights more fully into the revision.

W2: The experiments do not clearly isolate the effect of protocol changes. In Ecom-Offers, the paper replaces summary features with pre-offer transaction sequences, which changes both the input information and its temporal structure.

Thank you for pointing out this distinction in attribution. We agree that recovering the original transaction sequences from summary features in EO both increases the available information and changes how it is organized. Table 4 therefore cannot, on its own, establish that all the gains come from temporal structure. Section 6.2 already describes EO in terms of the specific protocol change of recovering the source sequences. We will make this limit on the interpretation clearer. For EO, we plan to hold the pre-offer transactions, event features, and history window fixed and add a time-agnostic or aggregated control using the same underlying events to distinguish the effect of recovering information from that of temporal relations.

W2/Q2: In ModernNCA (+ec), temporal embedding and entity-specific context construction are also applied together. Do the authors have results that separately measure the effects of temporal embedding and context construction in ModernNCA (+ec)?

We apologize for the confusion. Please see our response to Q2 from Reviewer dz5H.

Q1: How are the two protocol mismatches identified in practice, and what rules guide the choice of protocol changes for a given task?

We thank the reviewer for this question. Our paper primarily aims to show the connections between the methods and ideas in the two fields. When deciding where each family of methods applies, we believe the main considerations are the semantics of the task, its specific form, and our understanding of which capabilities the task calls for. Temporal tabular methods are better suited to covariate-centric tasks, so they tend to perform better when there are many covariates. Irregular time-series methods are better suited to history-centric tasks, so they tend to perform better when the task depends more on contextual observations. Finally, we would like to emphasize that context-based tabular methods tend to perform well across task types and may well be a first choice for these problems.

To identify a mismatch in a TabReD dataset, we first specify the entity, time, and target of the query. For entity collapse, we check whether consecutive state updates in the input sequence cross between distinct entities that do not share a common state process. The globally ordered sequence of rows in WE crosses station boundaries, so we construct a separate history for each station. For temporal relation mismatch, we trace how each row was generated. If a row summarizes a source sequence and the prediction target belongs to that instance, we check whether the relevant relations lie within the source sequence. In EO, we therefore use the transaction history before each offer as the input to a sequence-to-label task.

Thank you for the detailed clarification.

For W1, the explanation of entity collapse in WE and the explanation of the mismatch in temporal relationships in EO are helpful in clarifying the value of the unified interface. However, these insights are still closely related to the two case studies. Apart from reorganizing the existing methods, the contribution remains limited.

For W2, I appreciate your pointing out that restoring the transaction sequence in EO would change the available information and its temporal structure. The proposed time-agnostic or aggregated control helps to isolate these effects. But for ModernNCA (+ec), the explanation that only (+c) alone is not applicable does not fully isolate the contributions of time embedding and entity-specific context. The novelty and the attribution of the exps remain unresolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.