acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Channel-entity Bipartite Graph for Irregular Multi-channel Time Series

Abstract

Irregularly sampled multi-channel time series exhibit asynchronous observations, sparsity, and uneven sampling intervals, while their sampling patterns may themselves convey task-relevant information. Existing graph-based methods organize interactions by learning channel dependencies over fully connected candidate graphs, constructing channel relations from semantic priors, or building channel-time graphs within predefined temporal patches. However, fixed patching may leave some patches with insufficient information and split key temporal dynamics across patch boundaries, while dense or prior-based graphs may propagate irrelevant information when relations are insufficiently filtered or priors are mismatched. To address these challenges, we propose the Reference-Entity-Augmented Dynamic Graph model (REA-DG), which organizes observation-driven local interactions through a sample entity. REA-DG constructs an entity-channel bipartite graph in which the sample entity connects only to currently observed channels and mediates their interactions. To complement sparse local information, shared learnable reference entities connect to all channels through channel-specific learnable edges, injecting context learned from the training data into message passing. Experiments demonstrate that REA-DG achieves state-of-the-art performance across financial forecasting, clinical time-series classification, and network-traffic forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
irregular time series, multi-channel time series, missing observations, temporal modeling, dynamic graph learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GN4XOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces REA-DG for irregularly sampled multivariate time-series classification and forecasting. At each observation step, a sample entity connects to currently observed channels and mediates their interactions. Learnable reference entities provide shared context through channel-specific edges. Fast and slow channel memories retain historical information, which is retrieved to initialize subsequent entity states.

The paper evaluates the method on four clinical classification settings and two forecasting datasets, FNSPID and CESNET. REA-DG achieves the highest reported mean performance among the listed baselines. Ablations on P19 examine the reference entities, graph message passing, and memory components.

Critical strengths
  1. The architecture addresses relevant aspects of irregular observations. Observation-dependent connectivity, channel-specific memories, and historical retrieval provide a coherent approach to combining asynchronous measurements. Figure 2 makes the overall information flow reasonably clear.

  2. The clinical results are promising, particularly on P19. In Table 1, REA-DG improves P19 AUPRC from the strongest listed baseline's 62.5 to 71.3, an increase of 8.8 percentage points. The improvement on MIMIC-III is also substantial, from 48.4 to 52.2. These results warrant further investigation.

  3. The ablations establish that several components contribute to performance. In Table 3, removing graph message passing reduces P19 AUPRC from 71.37 to 60.26, while removing reference entities reduces it to 70.03. The reference-count experiment also shows that increasing the number of references does not consistently improve performance.

  4. Several evaluation choices are explicitly documented. The appendix describes patient-disjoint clinical splits, training-only normalization statistics for clinical data, validation-based checkpoint selection, and repeated runs. It also discloses checkpoint averaging for forecasting.

Critical weaknesses

W1. The reference-entity mechanism is not specified precisely enough to reproduce

Equation (4) updates the reference state to ut,k+u^+_{t,k}, whereas Equation (5) constructs reference messages from ut,ku_{t,k}. In contrast, the real-entity message explicitly uses the updated state Et+E_t^+. It is unclear whether this asymmetry is intentional or a notation error.

The computation of the attention weights αk,j\alpha_{k,j} is also missing. Nonnegativity and normalization do not establish whether these weights are learned static parameters or depend on the current sample, time step, channel state, or reference messages. These alternatives imply different mechanisms for providing shared context.

Furthermore, the paper specifies how the real entity is initialized at subsequent steps but does not clearly define whether channel-node and reference-entity states are reset from learned embeddings or carried forward across time. Shared trainable parameters should be distinguished explicitly from sample-specific intermediate states.

A complete algorithm describing initialization, within-layer updates, and state transfer across observation steps would resolve these ambiguities. Important configuration values are also missing, including the backbone dimension DD, graph depth LL, clinical learning rates, and the regularization parameters ρ\rho and λcos⁡\lambda_{\cos}.

W2. The provenance and comparability of the clinical baseline results require clarification

Several baseline means and standard deviations in Table 1 exactly match the corresponding entries in the original KEDGN paper. The row labeled "KEDGN" corresponds specifically to KEDGN-Name. Other variants in that paper perform better on some metrics. For example, KEDGN-ChatGPT reports P12 AUROC/AUPRC of 87.8/54.5, compared with 87.1/54.1 in the submitted table. Relative to that published variant, REA-DG's P12 improvements are therefore 0.1/0.9 percentage points [1].

This does not reverse the direction of the reported improvements, but it changes their magnitude on some tasks.

More importantly, the manuscript does not identify which results were reproduced and which were taken from previous publications. Appendix A.1 specifies particular fixed splits, but the correspondence between these splits and the reported baseline results is not established. Matching published numbers does not itself demonstrate an unfair comparison; however, the current description is insufficient to verify comparability.

The authors should identify result sources and model variants, and confirm consistency in patient splits, preprocessing, input variables, and metric implementations. A matched comparison with the strongest relevant clinical baseline would be particularly useful.

W3. The CESNET results do not isolate the contribution of the proposed backbone

Equation (16) presents a general forecasting decoder. Appendix A.4, however, states that the CESNET results use a dedicated time-basis query head with fixed multiscale frequencies, and that the original-head control is not included in the main results. Forecasting results also use parameter averaging across the three checkpoints with the lowest validation MSE.

These are legitimate modeling choices, but the resulting gains combine the effects of the backbone, decoder, and checkpoint-selection procedure. Section 4.2 largely attributes the improvements to the integration of current observations, historical information, and reference context.

On CESNET, MSE decreases from 0.8944 for TimeCHEAT to 0.8677 for REA-DG, approximately a 3% relative improvement. The original-head result is therefore important for interpreting the source of this gain.

A comparison between the original and time-basis heads on the same REA-DG backbone would directly address this concern. A simple time-feature-only control would also help establish how much predictive information the historical encoder contributes on this small dataset.

W4. The ablations support component usefulness more clearly than the proposed mechanistic explanations

The real-entity branch in Equations (4)–(6) aggregates messages from observed channels into a shared state and broadcasts contextual information back to channels. This is a reasonable inductive bias, but its advantages over comparable masked aggregation or attention-based feedback mechanisms are not established.

TimeCHEAT already uses channel–time bipartite graphs to represent observations. The distinctive contribution of REA-DG therefore needs to be established through its persistent entity state, historical retrieval, and reference-entity mechanism [2].

The w/o Graph experiment removes the complete message-passing module. Its performance drop demonstrates that this module is useful, but does not isolate the benefit of the proposed topology. The exact replacement computation and retained parameters are also unspecified.

Other interpretations similarly exceed the available controls:

  • The claim that reference entities compensate for extreme sparsity is supported only by aggregate P19 results, without analysis across observation-density levels.
  • Removing fast or slow memory may alter model capacity; it is unclear whether total state dimensionality is preserved.
  • The conclusion attributes benefits to interval-dependent decay, but Table 3 does not include a control that retains fast memory while disabling decay.
  • Figure 3 visualizes attention weights, but does not establish the effectiveness of the reference-message diversity regularizer.

A decay-disabled control and an observation-density analysis on the existing test set would provide focused evidence for central claims without requiring a substantially different experimental study.

W5. The CESNET evaluation provides limited evidence for broad forecasting generalization

Appendix A.3 reports only 11 entities, 23 training windows, 11 validation windows, and 12 test windows for CESNET. This can support a valid comparison, but the scope of the conclusion should reflect the small evaluation set.

The very small standard deviation across random seeds indicates stable optimization on this particular split. It does not quantify uncertainty across entities, time periods, or test-window composition. Numerous temporally correlated scalar targets also do not constitute an equivalent number of independent evaluation samples.

Table 4 evaluates overlapping cumulative portions of the same 168-hour forecasts. These results characterize performance within the existing forecast window, but do not provide independent replications.

Per-entity or per-window error differences, together with uncertainty estimates that account for within-entity dependence, would strengthen this evidence.

W6. The interval definition needs a clearer interpretation and forecasting boundary

Equation (8) uses

Δt,j=t+−t−2\Delta_{t,j}=\frac{t^+-t^-}{2}

when both neighboring observations exist, where t+t^+ is the next observation time. This quantity describes local sampling spacing rather than simply the elapsed time since the previous observation.

Using such information is compatible with encoding a complete historical window, provided that both neighboring times lie within the available input history. However, if the intervals or Tmax⁡T_{\max} are computed from the full sequence before separating history and forecast targets, the encoder could incorporate observation-pattern information from beyond the forecast origin.

Appendix A.3 states that future target values are excluded from model inputs, but does not establish the scope used to compute these temporal statistics. The authors should clarify this preprocessing order and explain why a two-sided interval is appropriate for decaying historical memory.

I do not regard the equation alone as evidence of leakage. The concern is the unspecified boundary between available history and future observation information.

Questions for authors
  1. Does Equation (5) use ut,ku_{t,k} or ut,k+u^+_{t,k}? How are αk,j\alpha_{k,j} computed, and do they depend on the sample or time step?
  2. How are channel-node and reference-entity states initialized and propagated across observation steps?
  3. Which entries in Table 1 are reproduced results, and which are quoted? Are their splits and preprocessing identical to those used for REA-DG?
  4. What is REA-DG's CESNET performance with the original prediction head? How much improvement comes from the time-basis head and checkpoint averaging?
  5. What computations replace the removed components in the ablations? Is memory dimensionality preserved, and what happens when decay alone is disabled?
  6. Are t+t^+ and Tmax⁡T_{\max} computed strictly within the input history? How are observations at the forecasting boundary handled?
Reviewer_WZeZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

REA-DG models irregularly sampled multivariate time series with a sample-level "real entity" that connects only to the channels observed at each step, plus K shared learnable "reference entities" and fast/slow channel memories. It reports the best results on four clinical classification datasets and two forecasting datasets.

Critical strengths
  • The design is simple and avoids learning a dense channel graph from sparse data.
  • The reported performance is strong: REA-DG has the best AUROC and AUPRC on all four clinical datasets (e.g., +8.8 AUPRC over the strongest baseline on P19), and the lowest MSE and MAE on both forecasting datasets, with results over 5 seeds.
Critical weaknesses
  • It is not clear what is new in the proposed components. Using an intermediary node to pass information between nodes, and using several learnable nodes shared across samples, are both common techniques in GNNs and related areas, but the paper does not mention or discuss this connection. Within time series, the related work lists SeFT, STraTS and DuETT (set/attention aggregation across channels) and GRU-D (incorporating elapsed time into state updates through decay), and the paper does not discuss how the real entity and the interval-dependent decay differ from them. I would suggest a short discussion of these differences, and a controlled test on P19 that replaces the real entity with attention pooling over the observed channels, and the decay with a GRU-D-style decay. This would show which part of the design matters.

  • The ablation does not isolate the proposed components. On P19, the gap between REA-DG (71.37 AUPRC) and KEDGN (62.5) is 8.8 points, but removing the reference entities, slow memory or fast memory lowers AUPRC by only 1.34, 0.68 and 3.12 points, and each of these variants still scores above KEDGN. The "w/o Graph" variant (60.26) is not defined in the paper, so it is unclear what remains after removing it and whether it isolates the observation-driven structure. The conclusion credits "interval-dependent memory decay", but there is no decay vs. no-decay comparison. I would suggest describing the w/o Graph variant, and adding a decay vs. no-decay comparison and the attention-pooling control above.

Questions for authors

What are the parameter count and training/inference time of REA-DG compared with the baselines?

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

This paper addresses irregularly sampled multivariate time series with asynchronous and sparse observations. To overcome the limitations of fixed temporal patching and predefined channel dependencies, the authors propose the Reference-Entity-Augmented Dynamic Graph (REA-DG). It constructs an observation-driven channel-entity bipartite graph to facilitate cross-channel interactions, while introducing shared reference entities to provide complementary contextual information. The constructed experiment has testified to the effectiveness of the proposed method.

Critical strengths
  1. Practically relevant problem. This paper addresses important challenges in existing graph-based methods for irregularly sampled time series. In particular, fixed temporal patching may result in insufficient information within certain patches and disrupt temporal dependencies due to inappropriate patch boundaries.
  2. Intuitive method design. REA-DG employs an observation-driven channel-entity bipartite graph to facilitate cross-channel interactions, while incorporating reference entities and fast-slow memories to provide shared contextual information and preserve historical observations, respectively.
  3. Competitive experimental results. REA-DG achieves strong performance across multiple benchmarks.
Critical weaknesses
  1. Unclear advantages of the proposed Channel-Entity Graph. REA-DG aggregates information from observed channels through a real entity and subsequently distributes it back to individual channels. From an information propagation perspective, this remains a form of indirect channel-channel interaction and may therefore retain similar limitations, such as the propagation of irrelevant information across channels. Although the bipartite graph differs in connectivity and computation, the paper does not sufficiently explain how this design addresses these limitations or what unique modeling advantages it offers over existing cross-channel aggregation methods.
  2. Insufficient motivation for introducing Reference Entities. The paper argues that shared reference entities capture cross-sample knowledge to compensate for sparse observations within individual samples (Lines 78-81). However, learning shared patterns across samples is already a fundamental capability of neural networks through parameter sharing. Therefore, capturing shared knowledge alone does not adequately justify the necessity of introducing additional reference entities.
  3. Insufficient validation of robustness to sparse observations. One of the main motivations is to address insufficient information within fixed temporal patches. However, REA-DG constructs graphs at individual observation steps, where the available information may be even more limited, particularly under nonstationary sampling patterns. Although historical memories may partially alleviate this issue, it remains unclear whether the proposed approach effectively handles severe sparsity. Further experiments, such as those under varying observation rates and sampling interval distributions, are needed to substantiate this claim.
Questions for authors
  1. Reference entities are primarily constructed from learnable embeddings and channel-specific learnable edge representations. How do they differ from conventional learnable global representations or latent tokens?
  2. How does the proposed method perform in more realistic forecasting or imputation scenarios, such as on the MIMIC dataset?
  3. Given the substantial differences in sampling patterns across tasks, can the authors provide ablation studies on additional datasets to verify the consistent contributions of the proposed components?
  4. What is the computational efficiency of REA-DG compared with existing methods, particularly in terms of runtime and memory overhead?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.