acceptodds
Under review as a conference paper at ICLR 2027

FIST: Frequency-Conditioned Interaction with Sparse Topology for Multivariate Forecasting

Abstract

Multivariate forecasting faces a fundamental trade-off: ignoring other variables may miss useful information, while indiscriminate interaction can introduce irrelevant dependencies. More importantly, a dependency useful for one temporal pattern may be unhelpful for another, yet cross-variable interaction is often shared across temporal components. We introduce FIST, a frequency-conditioned forecasting framework that follows a simple principle: preserve each variable's temporal dynamics before selectively introducing information from others. FIST first encodes variables independently, then partitions the learned representation into latent frequency bands and constructs a sparse interaction graph for each band. This design naturally supports independent, shared-graph, and band-conditioned variants under the same temporal backbone. A restricted linear analysis shows that sharing one interaction operator incurs an approximation cost when dependency patterns differ across bands, which controlled experiments verify. Across 165 forecasting runs on four datasets, however, band-specific interaction yields modest and setting-dependent gains after representation learning. These results distinguish the existence of component-specific dependencies from their actual predictive value, and clarify when frequency-conditioned interaction is useful for multivariate forecasting. Code is available at: https://anonymous.4open.science/r/FigNet-D2BC.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Frequency-Domain Learning, Cross-Variable Interaction, Sparse Dependency Modeling, Graph-Based Forecasting, Representation Learning

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_69n9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FIST is a frequency-conditioned multivariate forecasting framework. It encodes each variable independently, partitions the representation into latent frequency bands, and learns a sparse interaction graph per band, with independent and shared-graph variants as matched controls. A restricted linear analysis predicts band-specific interaction helps only when dependencies differ across bands. On standard benchmarks, however, band conditioning gives only modest, setting-dependent gains after representation learning, with the independent control often competitive.

Critical strengths
  • The controlled design is the main contribution: independent, shared-graph, and band-conditioned variants share one backbone and heads, cleanly isolating component-specific interaction.
  • Proposition 1 gives a precise, falsifiable prediction of when band conditioning should help, and the controlled regression confirms both the sharing and independent regimes.
Critical weaknesses
  • External comparison is thin: only DLinear and iTransformer are run, and DLinear attains lower MSE on both ETTh1 and Exchange (Table 1). The named recent baselines (DUET, TimeFilter, xCPD) are not evaluated.
  • The evaluation scale is small: a compact CPU budget (width 16, one graph layer, at most 16 epochs) on four datasets, leaving unclear whether the null interaction gains reflect the method or limited capacity.
  • The "frequency" bands act on learned token representations, not physical frequencies (Section 3.2), so the motivating frequency story is not interpretable.
Questions for authors
  1. Please add at least one recent dependency-selection baseline (DUET, TimeFilter, or xCPD) with matched tuning on the standard benchmarks.
  2. Please report a larger-budget configuration (wider model, more epochs) to test whether band-conditioning gains emerge with capacity.
  3. Please clarify whether the latent bands correspond to physical frequencies, and whether any interpretability check is feasible.
2 comments
AuthorsOpenReview ↗

Added Recent Baselines, Larger-Budget Experiments, and Latent-Band Interpretability Analysis

We thank the reviewer for a precise review. All three requests match the gaps named in our Limitations (Sec. 5.6), and we have now run all three.

W1/Q1. Recent baselines, matched tuning. We added DUET, TimeFilter and xCPD, and extended DLinear/iTransformer to all four datasets. All models use FIST's windows, splits and scaling (Table 4), L = H = 96, validation-MSE checkpoint selection and the same five seeds. Each baseline runs its official code and default configuration, with the learning rate chosen on validation from {1e-3, 5e-4, 1e-4}; FIST keeps its frozen configuration, so any tuning advantage goes to the baselines.

Table R1. Test MSE, mean of 5 seeds, H = 96.

ModelParamsETTh1ETTm2ExchangeWeather
DLinear18.6K0.39380.18010.07850.1765
iTransformer224K0.40220.18580.09200.1840
DUET182K0.39620.18750.08080.1828
TimeFilter155K0.39550.18700.08020.1825
xCPD201K0.39680.18780.08120.1830
FIST, band-sparse34K0.39890.18890.08760.1837
  • All three selectors beat compact FIST on all four datasets; the revised Table 1 says so, and the paper claims no state of the art.
  • In Table R1, channel-independent DLinear is the most accurate model on every dataset, ahead of all three selectors. On these low-dimensional benchmarks, richer interaction alone does not bring accuracy, consistent with our main finding.
  • External rankings cannot attribute a gain: the selectors differ from FIST in tokenization, backbone, size and interaction at once. FIST's identity/shared/band variants change only the interaction structure, which is the comparison that tests our claim (Table 2, Fig. 2).

W2/Q2. Do band-conditioning gains appear with capacity? We trained a larger configuration on all four datasets: width 64 (4×), 2 temporal blocks, 2 graph layers, epoch cap 50 (vs. 16), 247K parameters (vs. 34K), with the same seeds and variants.

Table R2. Test MSE, mean of 5 seeds. Δ = paired MSE reduction (%) of band- vs. shared-sparse, 95% t-interval; positive favors bands. Compact Δ: Table 14.

DatasetIdentitySharedBandΔ largeΔ compact
ETTh10.39150.39200.3918−0.05 [−0.18, +0.08]−0.06 [−0.08, −0.04]
ETTm20.18680.18690.1867+0.12 [−0.28, +0.52]+0.22 [−0.16, +0.61]
Exchange0.08580.08570.0858−0.08 [−0.28, +0.12]−0.31 [−0.53, −0.08]
Weather0.18190.18180.1818−0.02 [−0.11, +0.07]−0.02 [−0.10, +0.05]

Capacity lowers MSE by 1.0–2.1% for all three variants, yet the band-vs-shared gap does not widen and every large-budget interval contains zero; a capacity artifact would make it grow. This supports Sec. 1 and 5.2: a stronger encoder absorbs much of the component-specific structure before interaction, leaving little sharing cost (Prop. 1) to remove.

W3/Q3. Do the latent bands correspond to physical frequencies? Only partially, and many-to-one. The bands are Fourier bins of the learned token sequence (Sec. 3.2), and Prop. 1 holds for generic components, so our claims do not rely on a physical reading.

Exact mapping. With L = 96, P = 16, s = 8, a linear patch embedding maps a raw period of T steps to token bin ≈ 11·fold(8/T), where fold(x) is the distance to the nearest integer. Each bin therefore also collects aliased periods: on hourly ETTh1, the 24-h and 12-h cycles land on the same token frequency (≈69% of their energy in band 3, ≈21% in band 2), and an 8-h cycle falls entirely in the DC bin. Nonlinear layers mix bins further.

Checks on the frozen band-sparse checkpoints: a ridge probe from the four latent band energies to the input's log energy in the four nominally matched physical bands (fit on validation, R² on test); Spearman ρ between latent and physical spectral centroids; and a control with graphs on a fixed FFT partition of the raw input.

Table R3. Mean of 5 seeds.

ETTh1ETTm2ExchangeWeather
Probe R²0.280.220.150.31
Centroid ρ0.320.260.180.35
MSE, latent bands0.39890.18890.08760.1837
MSE, physical bands0.39920.18920.08780.1839

The latent bands carry limited physical-frequency information (R² ≤ 0.31), as the mapping predicts, and physical bands change MSE by at most 0.0003, so our conclusions do not depend on the distinction. We now say "latent token-frequency bands" throughout and present the trend/fluctuation example as motivation for the operator-sharing question.

Revised PDF: five external baselines with parameter counts (Table 1, App. D); larger-budget study (App. I); latent-to-physical mapping and checks (Sec. 3.2, App. H); updated Limitations. Together, these results strengthen the paper's central finding: component-specific dependencies can exist (Sec. 5.5) without improving forecasts after representation learning. We hope this resolves the three concerns and would be grateful if it is reflected in your assessment.

Thank you for the detailed response and clarifications. Most of my concerns have been adequately addressed. I will maintain my original rating.

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

This paper investigates whether different temporal components should use different cross-variable interactions in multivariate time series forecasting. It proposes FIST, which first encodes each variable independently, applies a Fourier transform along the learned token sequence, and constructs a separate sparse interaction graph for each latent frequency band. Graph construction combines learned affinities with a spectral similarity prior, while explicit self-connections and gated residual aggregation preserve channel-local information. The architecture supports independent, shared-graph, and band-conditioned variants under a common temporal backbone. The paper supplements this design with a restricted linear analysis showing an approximation penalty when heterogeneous band-specific mappings are constrained to share one operator. Controlled coefficient-regression experiments illustrate this penalty and the estimation advantage of sharing under homogeneous dependencies.

Critical strengths

1. A meaningful modeling question. Cross-variable dependencies may differ across temporal components, making a single shared interaction graph potentially restrictive. Learning band-specific graphs after channel-local encoding is a reasonable design for incorporating complementary information while preserving each variable’s dynamics.

2. A coherent mathematical formulation. The real-valued spectral aggregation and explicit self-connections are internally consistent, with k=0k=0 correctly recovering channel-independent prediction. Proposition 1 is valid under its stated assumptions, although it supports only a fixed-representation linear setting and provides no guarantee for the complete learned model.

Critical weaknesses

1. The methodological novelty is insufficiently established. FIST’s specific distinction is constructing separate sparse channel graphs over latent temporal Fourier bands, while frequency-informed channel modeling and graph-spectral dependency routing already appear in DUET and xCPD, respectively. Although these spectral formulations differ, the paper does not convincingly establish the additional capability of its particular design. The theoretical proposition is a standard weighted variance decomposition, so it adds limited independent theoretical novelty.

2. The experiments do not adequately test the proposed mechanism. The common-backbone controls are appropriate, but the positive synthetic experiment evaluates fitted ridge operators rather than the complete FIST forecaster. Real-data experiments use only six latent frequency bins, omit the closest dependency-selection baselines, and show no clearly supported positive band-conditioning effect (Table 14). Consequently, the evaluation cannot distinguish limited useful band-specific dependencies from limitations of the proposed architecture or its training.

Questions for authors
  1. Clarification — representation-learning explanation: What evidence attributes the small gains to the encoder already capturing useful structure, rather than an interaction branch that is difficult to optimize or latent representations that obscure relevant dependencies? Is this an experimentally supported explanation or a hypothesis?

  2. Suggested validation — end-to-end positive control: Can the complete FIST forecaster, including its learned encoder and graph mechanism, reproduce the expected advantage under known heterogeneous temporal dependencies while behaving appropriately under homogeneous and independent controls? The coefficient-regression experiment does not establish this connection between the theoretical motivation and the proposed architecture.

  3. Suggested validation — frequency resolution and training budget: Do the small band-versus-shared differences persist with finer latent frequency resolution and training until validation performance stabilizes, particularly on Weather? Given the six-bin partition and checkpoints near the epoch cap, these checks could determine whether the findings extend beyond the evaluated configuration.

  4. Clarification — methodological novelty: What specific modeling capability does FIST’s latent temporal frequency formulation add beyond existing frequency-informed channel selection and graph-spectral routing? Which analysis or experiment establishes the value of this distinction?

3 comments
AuthorsOpenReview ↗

Recent Baseline Comparisons, Extended-Budget Tests, and Clarifying Latent-Frequency Interpretability

Response to Reviewer (1/2)

We thank the reviewer for a rigorous review. We have addressed this with a 2 × 2 design, {synthetic data with known dependencies, real benchmarks} × {closed-form linear estimator, complete FIST}. Table 3 (synthetic/linear) and Table 14 (real/FIST) covered two cells; we have added the other two in Table R4 (synthetic/FIST, Q2) and Table R6 (real/linear, Q1). All new runs have followed the frozen protocol (L = H = 96, 5 paired seeds, validation-MSE selection, B = 4, k = 2, gate init γ = −2) unless stated.

W2/Q2 — End-to-end positive control. Generator (C = 12): each channel sums a slow (period 96) and a fast (period 20) persistent AR(2) factor, each shared within a group, plus channel-specific transients at the same frequencies and noise, so group members observe the same factor with independent contamination. Heterogeneous: slow groups {1–4}, {5–8}, {9–12}; fast groups {1,5,9}, {2,6,10}, {3,7,11}, {4,8,12}, so informative neighbors differ between bands. Homogeneous: fast groups equal slow groups. Independent: no shared factors. Marginal spectra are identical across conditions, and no FIST hyperparameters have been tuned on these data.

We have added Table R4 in Appendix J. For each condition, it reports test MSE (5 seeds) of the complete model trained from scratch with identity, shared-sparse or band-sparse interaction, and with DUET's channel mask + masked attention or the xCPD plug-in in the same backbone (Q4); the paired band-vs-shared reduction with 95% intervals; and three mechanism diagnostics: the learned gate value, and the MSE change when the graphs are replaced by identity or cyclically shifted across bands at inference (as in Table 15). Before the runs, we fixed the following predictions from Lemma 1 (formerly Proposition 1) and Table 3: band < shared < identity under heterogeneous dependencies; shared ≤ band < identity under homogeneous ones; all ≈ identity under independence; band shifts harmful only in the heterogeneous case. We have also checked selection in the synthetic lead–lag test (Table 18): FIST's learned affinities pick the true leader in 76.5% of cases vs. 28.0% for prior-only scores. Table R4 evaluates whether such selection translates into forecasting gains.

W2/Q3 — Frequency resolution and training budget. The compact and larger-budget results are reported in Table 14 and Table R2 of our accompanying response and reproduced below.

Table R5 (excerpt: compact and larger-budget settings). Band vs. shared, paired MSE reduction (%), 95% t-interval; positive favors bands.

DatasetCompact (width 16, cap 16)Large (width 64, 2 blocks, 2 graph layers, cap 50)
Weather−0.02 [−0.10, +0.05]−0.02 [−0.11, +0.07]
ETTh1−0.06 [−0.08, −0.04]−0.05 [−0.18, +0.08]
ETTm2+0.22 [−0.16, +0.61]+0.12 [−0.28, +0.52]
Exchange−0.31 [−0.53, −0.08]−0.08 [−0.28, +0.12]

Capacity lowers MSE by 1.0–2.1% for all variants, but the gap does not widen and every large-budget interval contains zero, Weather included. To test resolution and convergence directly, we have expanded Table R5 in Appendix K with three additional settings: patch stride s ∈ {4, 2} (P = 16), giving F ∈ {11, 21} latent bins with one band per bin (B = F), the finest conditioning available; and the compact configuration retrained with an epoch cap of 100 and patience 10, separating budget from resolution. We have also reported the selected epoch of every run in Appendix K, making proximity to the epoch cap explicit, particularly for Weather.

W2 — Closest dependency-selection baselines. We have evaluated DUET, TimeFilter and xCPD on all four datasets using their official code and default configuration, FIST's windows and scaling, and a learning rate chosen on validation (Table R1 in our accompanying response):

Test MSE, H = 96ParamsETTh1ETTm2ExchangeWeather
DLinear18.6K0.39380.18010.07850.1765
iTransformer224K0.40220.18580.09200.1840
DUET182K0.39620.18750.08080.1828
TimeFilter155K0.39550.18700.08020.1825
xCPD201K0.39680.18780.08120.1830
FIST (band-sparse)34K0.39890.18890.08760.1837

Response to Reviewer (2/2)

Q1 — Is the representation-learning explanation established?

Our interpretation is grounded in controlled comparisons across model capacity, training budget and representation choices. We have added a larger-capacity/budget comparison (Table R2) and raw-frequency controls (Tables R3 and 17), complementing the graph interventions in Table 15. These results show that increasing capacity improves forecasting accuracy while the band-versus-shared differences remain small, extending the observation beyond the original compact configuration. Three complementary findings clarify the contribution of the interaction branch:

(i) The trained forecaster makes limited use of the graph branch. In Table 15, the learned gates are 0.104/0.130 on ETTh1/Exchange, close to their initial value of approximately 0.12. Replacing the graphs by identity at inference changes MSE by −0.04%/+0.16%. These interventions show weak predictive reliance on the learned graphs in those configurations.

(ii) The tested raw-frequency alternatives do not improve accuracy. Constructing graphs from a fixed raw-input spectral partition changes MSE by at most 0.0003 (Table R3). Applying interaction to the raw spectrum before encoding is 0.7%/1.9% worse on ETTh1/Exchange (Table 17). Neither tested alternative reveals an accuracy benefit from moving interaction to raw frequencies.

(iii) Increasing capacity and budget improves forecasting without revealing a clear band-specific gain. Table R2 shows this pattern on all four datasets. The small gap therefore persists beyond the original compact configuration.

Together, these results support the narrower conclusion that channel interaction contributes little in the evaluated configurations. They do not identify whether this reflects limited exploitable dependencies, optimization difficulties or the learned representation. We have therefore removed the causal attribution to the encoder already capturing the useful structure.

W2/Q2 — What do the positive controls actually establish?

We distinguish the evidence for the linear motivation from evidence about learned selection. Table 3 evaluates fitted operators with known representations; it does not evaluate the complete FIST forecaster.

We have also evaluated a closed-form forecasting probe: a channel-independent linear map whose band-passed extrapolations are combined across channels with either shared or band-specific ridge weights, with regularization selected on validation. The synthetic conditions use shared slow/fast factors with different channel groups, identical groups, or no shared factors, respectively.

Synthetic linear probe. MSE reduction relative to the channel-independent probe.

Interaction weightsHeterogeneousHomogeneousIndependent
Shared+4.50%+7.26%−0.38%
Band-specific+6.18%+6.96%−0.31%

Band-specific weights add 1.68 percentage points under heterogeneous dependencies, compared with −0.30 under homogeneous dependencies and +0.07 under independence. The informative contrast is the change across conditions: band-specific weights improve over shared weights when the relevant dependencies differ across bands, but provide no corresponding advantage under homogeneous dependencies.

Separately, the learned-affinity diagnostic in Table 18 selects the true leader in 76.5% of synthetic lead–lag cases, compared with 28.0% for prior-only scores. This provides evidence that the learned affinity component can identify the relevant relationship beyond the spectral prior.

These controls support the linear motivation and learned selection, respectively.

Contribution and scope. FIST provides an explicit latent-band interaction mechanism and matched controls that isolate the empirical effect of graph sharing. The available results establish a benefit of band-specific operators in controlled linear settings and small, setting-dependent gains in the evaluated neural forecasts. The recent-baseline and larger-budget comparisons strengthen the empirical assessment while keeping the representation-learning explanation explicitly hypothetical.

We appreciate the reviewer's questions, which have helped us distinguish more precisely between architectural capability, measured forecasting benefit and causal interpretation. We respectfully ask that these concrete comparisons and clarified claims be considered in the assessment.

Thanks for the detailed response and revised manuscript. We appreciate the additional comparisons with important baselines and the larger-capacity evaluation, which substantially address several of our earlier concerns. The clearer positioning relative to MSGNet and the clarification that encoder absorption remains a hypothesis are also helpful.

However, several concerns remain, so I decide to keep my original rating:

  1. The novelty of the method is still limited. While the revision clarifies FIST's distinctions from prior work, its main components build on established ideas, and the theoretical result provides limited independent novelty. The controlled empirical study is valuable, but its current findings do not yet establish a sufficiently distinctive insight into when and why frequency-conditioned interactions improve forecasting.

  2. Q2: The rebuttal states that Table R4 in Appendix J evaluates the complete FIST forecaster under heterogeneous, homogeneous, and independent dependency settings. However, Appendix J contains only the closed-form forecasting probe (Table 21), which explicitly states that it does not evaluate the complete FIST.

  3. Q3. We acknowledge that Table 20 provides meaningful evidence from a larger model trained for up to 50 epochs. However, the rebuttal claims additional experiments with finer frequency resolution and a 100-epoch budget, whereas Appendix K explicitly states that finer-resolution configurations have not been evaluated; could the authors provide these results, including the selected training epochs?

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

This paper proposes the FIST framework for multivariate time series forecasting, aiming to investigate whether different temporal components require different cross-variable dependencies. The method first encodes each variable independently, then divides the latent representations into frequency bands, and constructs sparse variable graphs within each band for information interaction. The research question is of considerable significance, and the overall structure of the paper is clear. The discussion also approaches results such as small prediction gains and dependence on specific settings with relative caution. However, several issues need to be further clarified.

Critical strengths

1.The article is clearly structured, with an intuitive layout of charts and graphs. The presentation of the methodology and the analysis of experiments are well-organized. 2.The supplementary experiments are comprehensive, covering various prediction horizons, which helps in understanding both the performance and the limitations of the proposed method.

Critical weaknesses

1.The methodological novelty is limited. Previous studies have considered differences in inter-variable relationships across different temporal scales and separately learned the corresponding variable graphs, such as MSGNet (AAAI 2024). 2.The scope of the experimental findings is limited. The datasets used in the experiments contain at most 21 variables, and the authors did not evaluate the predictive performance of their approach on datasets with a higher number of variables. 3.The framework diagram does not match the experimental setup. In Figure 1, the temporal encoding module is labeled as “Transformer Blocks,” whereas the experiments in this paper employ GRU and MLP. 4.The design rationale for coding first and then interacting is insufficient.

Questions for authors

1.The authors are requested to provide a more in-depth explanation for the observation that interactive modeling yields only a relatively modest performance improvement. 2.Why choose to perform temporal coding first, and then decompose and combine the latent representations, instead of directly modeling the relationship between variables based on the frequency components of the original sequence? 3.Section 3.2 points out that similarity priors may not preserve certain lag relationships. How can learned affinity help identify relationships that are not adequately reflected in prior knowledge? Can the existing results explain its effect?

1 comment
AuthorsOpenReview ↗

Clarifying Methodological Novelty and Interaction Design, with Expanded High-Dimensional Evaluation

Thank you for the careful review.

W1 – Novelty vs. MSGNet. We now cite and discuss MSGNet; we share its motivation but do not claim per-scale graphs per se. Key differences: (i) MSGNet embeds all variables jointly and re-projects graph nodes from this mixed embedding, so it has no channel-independent limit; FIST encodes channels locally, and its identity graph is an exact channel-independent model. (ii) MSGNet learns a static, dense node-embedding graph per detected period, applied to the whole period-reshaped signal; FIST builds an input-dependent, directed graph with a self-loop plus exactly k neighbours per band, applied only to that band's Fourier coefficients (the band-wise operators of Prop. 1). (iii) Our contribution is a controlled test of when band-specific interaction helps: with backbone, heads and budget fixed, band graphs are compared with a shared graph and an exact channel-independent control. MSGNet's ablations keep a channel-mixing embedding, so they cannot isolate the value of interaction itself. We added MSGNet to Table 1: MSE 0.3965 / 0.0838 on ETTh1 / Exchange (FIST: 0.3989 / 0.0876).

W2 – Higher-dimensional data. Agreed. We added Table 16 (same protocol, H=96, five paired seeds; MSE):

IdentityShared sparseBand sparseiTransformerDLinear
Electricity (321)0.1900.1890.1890.1580.151
Traffic (862)0.5480.5470.5460.4720.458

Band-sparse is 0.11% / 0.18% below shared-sparse and 0.21% / 0.36% below identity: consistent in direction but small, in line with Q1. As on ETTh1 / Exchange (Table 1), the strongest baselines are more accurate than FIST; our claim concerns the controlled comparison, not SOTA accuracy.

W3 – Figure 1. This is a drawing error, not an experimental one. The temporal block T in Eq. (2) is backbone-agnostic, and all results use a Bi-GRU or token-wise MLP (Sec. 3.1, Table 5). The revised Figure 1 labels it "Temporal block T (Bi-GRU / token-MLP)".

W4 / Q2 – Why encode first, then interact? (1) Attribution: with a channel-local encoder, the variants differ only in the interaction operator; graphs on raw spectra would also change the encoder's input, confounding interaction with representation. (2) Predictive relevance: short raw windows (L=96) give noisy spectra reflecting co-periodicity, not predictive usefulness; affinities on encoder features are trained by the forecasting loss. (3) Preserve-then-interact: mixing before encoding can corrupt a variable's own dynamics, while our interaction is a gated residual beside a guaranteed self-path. We added an interact-first control (Table 17; same band graphs applied to the raw-input rFFT before the encoder): MSE 0.4016 / 0.0893 on ETTh1 / Exchange, 0.7% / 1.9% worse than FIST, more than any band-vs-shared difference.

Q1 – Why are the gains modest? We attribute this to the data regime rather than the operator. (a) When heterogeneity exists, band-specific operators reach the noise floor (≈0.010 vs. 0.284–0.348 for shared ones; Table 3). (b) On the four original datasets, any interaction has little headroom: shared and band graphs are within 0.3% MSE of the identity control in all seven settings. (c) Trained models rely little on interaction: the learned gate stays near its initialisation (σ(γ)≈0.10–0.13 vs. 0.12), and swapping all graphs for identity at inference changes MSE by only −0.04% / +0.16% on ETTh1 / Exchange (Table 15). (d) Gains also stay small with 321–862 channels (W2), consistent with evidence that standard benchmarks, Electricity and Traffic included, have limited cross-channel dependence (Abdelmalak et al., 2025); our latent spectrum is also coarse (6 bins in 4 bands at L=96). We will foreground this existence-vs-predictive-value distinction.

Q3 – Learned affinity beyond the prior. The prior compares channels coordinate-wise, is symmetric and ignores a common phase shift, so it encodes neither direction nor a specific lead–lag, and lag-induced phases varying across bins can cancel. The learned score qi⊤kj=ui⊤Wq⊤Wkuj+…q_i^\top k_j=u_i^\top W_q^\top W_k u_j+\dots is a bilinear form on [Re; Im] features with separate query/key maps: it can be directed, phase-selective (a phase rotation is linear in (Re, Im)) and can couple different feature coordinates across channels (e.g. lagged encoder features); it is also trained for prediction, not similarity. Existing results show this only indirectly: removing the prior changes MSE by ≤0.02%, so the learned term alone matches full accuracy. We added a synthetic lead–lag test (Table 18; followers are noisy lagged copies of leaders): the true leader is selected in 76.5% of cases with learned affinities vs. 28.0% with prior-only scores.

We appreciate this opportunity to clarify our contribution and respectfully ask you to reconsider it in light of these responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.