acceptodds
Under review as a conference paper at ICLR 2027

Not All Events Are Equal: Capturing Heterogeneous Impact of Aperiodic Events in Long-Term Spatio-Temporal Forecasting

Abstract

Aperiodic events, such as traffic incidents and extreme weather, induce irregular deviations from recurring spatio-temporal patterns. Accurately forecasting these deviations is particularly challenging over long horizons. Existing methods typically approach this challenge by decomposing signals into regular and aperiodic components, but usually process all aperiodic components with an event-invariant parameterization. This overlooks the fact that the impact of an aperiodic event is heterogeneous, depending on when and where the event occurs and what type it is. A event-invariant model, agnostic to this heterogeneity, cannot adequately capture it. To address this challenge, we propose HEAM (Heterogeneous Event-Adaptive Modulation), a plug-and-play module that can be integrated into existing forecasting backbones to achieve event-adaptive modeling of the aperiodic component instead of event-invariant modeling.Specifically, we propose a Decompose-Situate-Identify pipeline, that generates modulation vectors from a library of modulator pools, conditioned on the spatio-temporal context and the latent event type of the aperiodic component. These vectors then modulate the aperiodic representation, so that the heterogeneous impact is explicitly captured.Experiments on datasets spanning multiple domains show that HEAM improves strong backbones of different architectural families, on average across all datasets and horizons. The gains are largest at the longest forecast horizon and in event-dense periods, particularly where the modeling of the heterogeneous impact of aperiodic events matters most.

Primary area:
learning on time series and dynamical systems
Keywords:
Data Mining, Spatio-temporal Data

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

The paper argues that existing decomposition-based spatio-temporal forecasters model the aperiodic (event-induced) component with one shared parameterization, although the effect of an event depends on when, where and what kind it is. It proposes HEAM, a module inserted before every layer of a backbone. At each layer the hidden representation Z is split into a "context" part Zc, produced by an MLP from time-of-day, day-of-week and node embeddings and trained to reconstruct stop-gradient(Z), and an "aperiodic" residual Za = Z − Zc. Zc is soft-assigned by cosine similarity to G pool prototypes ("Situate"); a projection of Za is soft-assigned to K prototypes within each pool ("Identify"); prototypes are updated by EMA. The two assignments mix a G×K table of learnable vectors into a modulation vector m, and Za is rescaled as (1 + γ·m) ⊙ Za, concatenated with Z and passed to the (widened) next layer. HEAM is added to GraphWaveNet and to a patch-Transformer variant of VisiFold, and evaluated on PEMS08, METR-LA and Solar-AL at horizons of 24/36/48 steps with five seeds, plus a peak/off-peak breakdown, an ablation at horizon 48 and visualizations of the routing.

Critical strengths
  1. Backbone-controlled evaluation with multiple seeds. HEAM is compared with its own host backbone under the same optimizer, schedule and epoch budget, over five seeds with standard deviations (Table 4), on two architecturally different hosts. For MAE at the 36- and 48-step horizons the improvements (e.g. PEMS08 16.36 → 16.05, METR-LA 5.01 → 4.72) are several standard deviations wide. This is more careful than most plug-in papers and is the main reason I do not rate the paper lower.

  2. A sensible problem statement and a reasonably light module. Conditioning the treatment of the non-regular residual on spatio-temporal context is a plausible refinement of decomposition-based forecasters, and the module adds no cross-node or cross-token computation.

Critical weaknesses
  1. The paper never shows that HEAM improves forecasting of aperiodic events; what is evaluated is rush hour.

    • There is no definition, detection or annotation of events anywhere. The "event-dense periods" used in RQ2 are fixed clock intervals (06:30–08:00 and 16:00–17:30 on PEMS08, 07:30–09:00 and 16:00–17:30 on METR-LA, 12:00–15:00 on Solar-AL). Morning and evening peaks are the most periodic structure in traffic data, so a larger gain there is evidence about recurring high-variance periods, not about incidents or extreme weather. At horizon 24 (Figure 8) the peak/off-peak difference is negligible on Solar (−2.2% vs −2.0%) and reversed on METR-LA (−2.7% vs −2.9%), where HEAM-V is 0.6% worse than its backbone in peak hours.
    • The "aperiodic representation" is whatever part of a hidden state is not predictable from (time-of-day, day-of-week, node id). That residual contains all input-dependent information, including ordinary level and trend, so calling it event-induced is an interpretation that is not verified. The evidence for "latent event types" is one node on one day (Figure 5b).
    • A direct test is available and missing: stratify test error by the magnitude of the deviation from the historical time-of-week profile, or use incident logs, and show the gain concentrates in high-deviation windows.
  2. The gains are small, confounded with added capacity, and the baseline table is not reliable enough to support the state-of-the-art claim.

    • HEAM concatenates the modulated residual with Z and doubles the input width of every consuming layer, in addition to per-layer embedding tables, prototypes and modulators. No parameter count, FLOPs or run-time is reported, and no width-matched backbone is included, so capacity is a live alternative explanation. The ablation variants "w/ Decompose" and "Concat" are the relevant controls but are shown only as bars spanning about 0.2–0.3 MAE at a single horizon and backbone, without numbers or deviations; the hyperparameter study shows the same narrow range.
    • At 24 steps HEAM-V is worse than its backbone on PEMS08 RMSE (25.61 → 25.66) and on METR-LA MAE and MAPE (4.42 → 4.43, 11.69 → 11.79), and on PEMS08 it is behind STAEformer in MAE at 24 steps (14.60 vs 14.91) and behind VisiFold in RMSE at 48 steps (27.76 vs 27.79). The text reports only the favourable 48-step MAE comparison.
    • Several baseline rows are implausible: on METR-LA, DCRNN has MAE 19.34 / MAPE 52% at 24 steps but 8.16 at 36 steps; on PEMS08, DualCast jumps from MAE 19.66 (36 steps) to 26.09 (48 steps). Baselines use "configurations adapted from BasicTS" with input length equal to horizon and a single run; they appear untuned for this setting.
    • Scope is narrow for a general plug-in claim: three datasets, two hosts, one of which had to be re-architected to accept the module. The mechanism (prototype routing into a pool of channel-scaling vectors) is close to memory- and prototype-based heterogeneity models (e.g. MegaCRN, which targets incidents and is neither cited nor compared) and to DPRNet, from which the orthogonality regularizer is taken. The learned gate γ, which determines how much modulation is applied at all, is never reported.
Questions for authors

Clarifications that could change my assessment:

  1. Please stratify test error by a data-driven event measure (e.g. deciles of |x − historical time-of-week mean| in the input or target window, or incident records for METR-LA / PEMS). Do HEAM's gains concentrate in the high-deviation strata?
  2. Report parameters, FLOPs and training/inference time for each backbone with and without HEAM, and add a backbone whose layers are widened to the same parameter count.
  3. Give the ablation of Figure 4b as a table with mean ± std, for both backbones and all three horizons.
  4. What values does γ take after training, per layer and dataset? What is the distribution of the resulting scale (1 + γ·m)?
  5. Please check the DCRNN (METR-LA) and DualCast (PEMS08) rows and state how baseline hyperparameters were chosen for 24–48-step inputs.
  6. Does HEAM help under the standard 12-step protocol and on larger benchmarks (e.g. LargeST), and when plugged into a decomposition-based host such as STDN or HyperD, which is the family the paper criticises?

Suggestion for future work (does not affect my rating): comparison with memory-based heterogeneity models such as MegaCRN.

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

The paper points out that most existing models ignore the difference among aperiodic events, and treating them as the same through a fixed parameters. Motivted by this oberavtions, this paper proposes HEAM, a plug-and-play module for adapting forecasting representations to heterogeneous deviations from regular spatial-temporal patterns.

Critical strengths

S1. The paper explicitly highlights the heterogeneity of aperiodic events and proposes a plug-and-play framework to address it, demonstrating its effectiveness on two forecasting backbones.

S2. The evaluation covers multiple types of spatial-temporal forecasting tasks and includes recent baselines that explicitly decouple regular and irregular components.

Critical weaknesses

W1. The presentation is not easy to follow. For example, Sections 4.3–4.4 introduce context and event prototypes, but the event prototype ee appears abruptly and is used for matching before its construction and update mechanism are explained. Deferring the detailed updates to Appendix A further disrupts the explanation. This obscures the core mechanism and makes it difficult to identify the methodological novelty beyond existing prototype learning and regular–aperiodic disentanglement approaches. Besides, the notation is potentially confusing: mm denotes both the superscript label in πm\pi^m and the modulation vector, alongside the modulator bank MM. Please distinguish these roles more clearly.

W2. The connection between the two-level prototype structure and the heterogeneous future impacts motivating the introduction is insufficiently established. Its methodological novelty over existing prototype learning and environment-codebook approaches also remains unclear.

W3. The peak-hour advantage in Figures 4(a), 7, and 8 is inconsistent: HEAM-V achieves identical peak/off-peak improvements on PEMS08 at horizon 36 and worsens peak-hour performance on METR-LA at horizon 24.

Questions for authors

Q1. CaST’s [1] environment codebook can be viewed as learning heterogeneous environmental representations, while DualCast’s environment loss explicitly encourages diverse aperiodic patterns. How does HEAM’s hierarchical prototype learning differ from these existing treatments of heterogeneity? Given this overlap, is the characterization of prior methods as “event-invariant” too broad? Please clarify the specific limitation HEAM addresses and the evidence supporting its claimed advance.

[1] Deciphering Spatio-Temporal Graph Forecasting: A Causal Lens and Treatment.

Q2. Why use the same number KK of event prototypes across contexts with potentially different event diversity? Does this cause prototype redundancy or merge distinct event patterns?

Q3. The paper does not justify increasing KK from 8 to 32 specifically for HEAM-G on SOLAR-AL. It also leaves the construction of the gray regular component in Figure 5(b) unclear, limiting the interpretation and reproducibility of this visualization.

Q4. How are aperiodic events operationally identified, particularly on SOLAR-AL? What evidence supports using 12:00–15:00 as a proxy for periods where their impacts are most consequential?

Reviewer_WCSuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the problem of long-term spatio-temporal forecasting. The authors propose HEAM (Heterogeneous Event-Adaptive Modulation), a plug-and-play module that modulates the representation of the aperiodic component in an event-adaptive manner. They propose a Decompose-Situate-Identify pipeline that generates modulation vectors tailored to each event’s context and latent type, such that the heterogeneous impact of events can be captured. The authors conduct experiments on three public benchmarks and demonstrate the effectiveness of the proposed solution.

The motivation and problem definition are clear. The focus on the impact of aperiodic events looks interesting. The related work section seems generally well explained. Also, the proposed framework seems overall reasonable. However, I am concerned that the proposed solution's accuracy improvement is modest and its effectiveness may be limited.

Critical strengths

(S) The motivation and problem definition are clear

(S) The focus on the impact of aperiodic events looks interesting

Critical weaknesses

(W) The proposed solution's accuracy improvement seems modest, and its effectiveness may be limited

Questions for authors

(Q1) In Figure 1, there are several sensors (sensors 1, 67, 161, and 102). I was wondering if some of these numbers are incorrect.

(Q2) Could you please explain the theoretical and mathematical definition of “events”?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.