acceptodds
Withdrawn from ICLR 2027

Before Relational Reasoning: Restoring Synchronicity in Event-Centric Irregular Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Event-centric models for irregular multivariate time series commonly embed observations independently before learning their relations. This computation order leaves synchronous evidence outside the states from which relational reasoning begins, even though measurements recorded at the same timestamp jointly describe the underlying system. We identify this architectural mismatch as pre-propagation synchronicity loss and challenge the assumption that event representations should be constructed independently of available snapshot context. We formulate pre-propagation synchronicity modeling as a representation interface that separates context construction from relational forecasting. The interface reconstructs permutation-invariant snapshot context and fuses it with event evidence using local event-layout support. Across three clinical benchmarks and four forecasting architectures, the design improves mean MSE and MAE in all 12 model–dataset pairs. Component ablations support the contributions of synchronous context, adaptive fusion, and event-layout support. The consistent gains across attention-based, graph-based, and hypergraph-based forecasters establish the empirical reach of the principle beyond a single architecture. These results identify representation ordering as an overlooked design axis for irregular time-series forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series, Event-Centric Modeling, Representation Learning, Relational Reasoning, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eyztOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a class of irregular multivariate time-series forecasting architectures that (1) represent each observed event independently and (2) integrate information across events through a relational mechanisms. The authors argue that this computation order overlooks context already available among measurements sharing the same recorded timestamp. Although the relational structure may connect co-timestamp events, their initial states do not yet depend on one another when the first relational transformation is applied. The authors call this mismatch pre-propagation synchronicity loss. To address this issue, the paper introduces pre-propagation synchronicity modeling, an architecture-agnostic representation interface that reconstructs the available co-timestamp context and incorporates it into each event state before relational propagation. The resulting information path is: (1) event representation, (2) synchronicity restoration, and (3) relational reasoning. Its hypergraph realization, built on HyperIMTS, is called MissHyper. Experiments on three clinical datasets report improved mean forecasting errors across the evaluated architectures, with MissHyper obtaining the lowest mean errors among the tested configurations.

Critical strengths
  • The paper identifies an underexamined design choice in irregular time-series models using relational forecasters: not only how observations interact, but when shared context enters their representations. This is particularly relevant for clinical data, where co-recorded measurements can be jointly informative.
  • The proposed interface can be applied to different relational forecasters rather than being tied to a single backbone, increasing its potential relevance.
  • The experimental section follows a clear progression: it tests the interface across four architectures and three clinical datasets, compares MissHyper with existing forecasters, and examines the contributions of snapshot context, gating, and event-layout support through ablations. Overall, the experiments are reasonable and aligned with the paper’s main claims.
Critical weaknesses

Weakness 1

Proposition 1 provides limited support for the architectural diagnosis. Proposition 1 correctly shows that averaging and nonlinear transformation need not commute, but this is an elementary observation. More importantly, the proposition neither establishes that the proposed ordering is preferable nor that a relational network cannot recover the relevant context.

Weakness 2

Variable identity is not explicitly represented in the snapshot context. Equation 5 constructs each event embedding from its value, query flag, and support, but not from its variable identity. Therefore, measurements from different variables may be treated identically when these inputs coincide, even though their meanings differ. This design may remove variable-specific information before snapshot aggregation. The paper should justify this choice and discuss it as a limitation.

Weakness 3

Clarity and organization. Several central concepts and pieces of notation are introduced either too late or without sufficient definition, making the proposed method unnecessarily difficult to follow.

  • 3.1 The problem statement is not presented in the clearest order. The opening paragraph introduces the main architectural issue and defines pre-propagation synchronicity loss before clearly delimiting and referencing the relevant family of “event-centric” models supplied with relational forecasters. The necessary context appears only later in the Introduction, in the discussion of event/set representations and relational forecasters. Defining the affected model class and citing representative methods at its first mention would make the motivation easier to understand and prevent the initial claim from appearing broader than intended.

  • 3.2 Some basic concepts and notation in Section 4.1 are insufficiently defined. In particular, the paper does not explicitly define what constitutes an episode or an event. Moreover, in the event-layout matrix (Ωb∈0,1L×C)(\Omega_b\in{0,1}^{L\times C}), the dimensions (L) and (C) are not clearly introduced. The clipped local window (Ww(ℓ))(W_w(\ell)) should also be formally defined. The role of the window size (w) is explained only at the end of the section, whereas it should be introduced when (Ww(ℓ))(W_w(\ell)) first appears.

  • 3.3 Minor: The organization of the Related Work and hypergraph presentation could be improved. Related Work repeatedly restates the contribution, while some of its clearest scope statements would be more effective in the Introduction. It could instead focus on positioning the method and introducing the HyperIMTS backbone. Since MissHyper is the main realization, the methodological section should then clearly define its hypergraph construction—specifically, the event nodes, timestamp hyperedges, and variable hyperedges—before presenting the propagation operator in Section 4.4.

  • 3.4 Minor: Figure 1 and 2 are never referenced in the main text.

Altogether, these things strongly indicate that the text was heavily generated by an LLM and the text not checked/reviewed thoroghly enough by a human.

Weakness 4

The proposed “pre-propagation” module is itself a form of relational aggregation. The operator pools co-timestamp event states into a shared summary and then broadcasts that summary back to update each event. This is effectively a timestamp-local message-passing step placed before the existing backbone. Keeping the original propagation machinery unchanged therefore does not mean that the full model has the same amount of relational computation as the baseline. As a result, the gains may come from adding an extra local interaction stage, rather than from repairing a distinct architectural failure. The paper would be clearer if it framed the contribution as an early timestamp-context layer and compared it against a matched additional message-passing baseline.

Weakness 5

Novelty relative to prior set- and context-pooling designs is limited and under-discussed. Pooling a shared summary and returning it to individual elements is an established pattern in Deep Sets and PointNet. Related ideas also appear in irregular time-series work such as SeFT, which combines set-level summaries with individual observations. Hypergraph aggregation frameworks such as AllSet already cover the general form of node-to-group-to-node updates. The gated interpolation likewise resembles familiar transform/carry mechanisms. The potentially novel part is the specific combination of timestamp-local pooling, event-layout support, and placement before several irregular forecasting backbones. That contribution is narrower than identifying a previously overlooked general principle, and the related-work discussion should make this positioning more precise.

Weakness 6

The ablations do not fully isolate why the method improves. The design adds several ingredients at once: an observed/query flag, a local layout feature, a context projection, a gate, and additional nonlinear computation. Support also appears in multiple places. A comparison against the original model therefore tests a bundle of changes rather than synchronicity restoration alone. In addition, the variant without snapshot restoration still retains a substantial fraction of the improvement over HyperIMTS on all three datasets. This does not invalidate the full model, but it weakens the claim that synchronous context is the main driver. A useful control would keep the module active while replacing the shared summary with a self-only summary, since a singleton timestamp still receives an extra gated nonlinear transformation even when no co-timestamp neighbor exists.

Questions for authors
  • Q1 - W1 : Can you provide a controlled ablation that places the same snapshot-context module before versus after the first relational block, while keeping the remaining architecture fixed? This would more directly test whether the reported gains arise from operation ordering rather than from the additional context module or model capacity.

  • Q2 - W2 : What limitations does excluding variable identity impose on the proposed notion of snapshot context?

  • Q3 - W4 : Can you compare the proposed module against an ordinary additional timestamp message-passing layer with matched features, parameter count, and computational budget?

  • Q4 - W6 : Can you report a self-only control in which the snapshot summary is replaced by each event’s own transformed state, so that the extra nonlinear capacity is kept but co-timestamp sharing is removed?

Reviewer_1V9ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulate pre-propagation synchronicity modeling as a representation interface that separates context construction from relational forecasting. Experiments on three clinical benchmarks and four forecasting architectures show consistent improvements over existing models.

Critical strengths

The paper presents a clear motivation and the empirical results are consistent.

Critical weaknesses
  1. The core technical novelty is limited, as the method mainly adds gated fusion before relational propagation.
  2. The performance gains are limited in both the main experiments and the ablation studies.
  3. The evaluation is restricted to clinical datasets (Physionet and MIMIC), common non-clinical irregular benchmarks such as Human Activity and USHCN are missing (refer to benchmarks of tPatchGNN and HyperIMTS).
Questions for authors

Can the authors better clarify the technical novelty?

Can the authors include results on non-clinical irregular benchmarks such as Human Activity and USHCN?

Reviewer_o6PVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a snapshot-aware initialization method for irregular time-series forecasting. It aggregates observations sharing the same timestamp and uses a learned gate to combine this context with individual event representations before relational modeling. The method is tested with four forecasting models on three clinical datasets, showing improvements in MSE and MAE across all 12 model–dataset pairs.

Critical strengths

The method is simple and can be applied to different forecasting architectures. The experiments cover several models and datasets, with consistent improvements over the original models. The component ablations also help explain the contributions of snapshot aggregation and gating.

Critical weaknesses
  • The technical novelty is limited. The main method is essentially weighted aggregation of same-timestamp observations followed by gated fusion. The distinction from existing aggregation and message-passing operations needs to be clearer.

  • Proposition 1 only shows that nonlinear transformation and mean aggregation do not commute. This is a basic property and does not establish that existing relational models fail to learn synchronous information. The theoretical motivation is therefore weaker than the paper suggests.

  • Some improvements are small. For example, tPatchGNN's MSE on P12 decreases from 0.3168 to 0.3143 (Table 1). Since the method adds an aggregation and fusion module, a parameter-matched baseline or an additional propagation layer would help establish whether the gain comes from the proposed computation order.

  • The definition of synchronicity depends on timestamp resolution, which differs considerably across datasets. The paper does not test how changing the resolution affects the results.

  • The implementation of the proposed interface in the non-hypergraph baselines is not described in enough detail, making those comparisons difficult to assess.

Questions for authors
  • Does the method still outperform the baselines when model capacity is matched, for example by adding an extra feed-forward or relational layer?

  • How sensitive are the results to timestamp discretization? What proportion of timestamps contain multiple observed variables in each dataset?

  • How exactly is the proposed representation module integrated into tPatchGNN and Warpformer?

  • Equation (4) includes predefined query positions in the event-layout support. How much of the improvement comes from these positions rather than observed events alone?

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

Related papers

Discussion (0)

Sign in to comment.