acceptodds
Under review as a conference paper at ICLR 2027

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

Abstract

Short-term traffic forecasting is essential for urban transportation management. While traffic flow exhibits regular day-to-day periodic patterns, spatiotemporal fluctuations make accurate prediction challenging. Conventional traffic flow prediction models have difficulty capturing these variations, particularly when spatiotemporal dependencies become highly non-stationary. To address this issue, this study proposes Euler Frequency Spatiotemporal Graph Neural Network(EFSG), a spatiotemporal model that uses Euler Frequency Modulation (EFM), which applies FFT on each short input window and jointly edits amplitude and phase for every sample, node, and channel. EFSG further builds a spectrum-aware dynamic graph linking nodes with similar spectral patterns beyond physical adjacency, and a gated fusion layer adaptively combines local and frequency-domain features. Ablation studies reveal that EFM and spectral propagation function as a coupled unit driving performance gains. Experiments on four real-world datasets show that EFSG achieves a 19.9% reduction in MAE on NYC-Bike pick-up and a 60% reduction in GPU memory consumption on PEMS04 compared with TEDDCF, while converging within 20 epochs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
spatiotemporal graph neural networks, traffic forecasting, Euler frequency modulation, spectrum-aware graph learning, non-stationary time series

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

This paper proposes EFSG for short-term traffic forecasting. The model builds a local branch using temporal convolutions and an adaptive graph, applies Euler Frequency Modulation (EFM) to adjust spectral magnitudes and phases, and constructs a dynamic graph based on similarities between node spectral magnitudes. Gated fusion then combines these representations for multi-step prediction. The authors report lower MAE than the listed baselines on PEMS03, PEMS04, NYC-Bike, and NYC-Taxi, together with lower peak GPU memory usage than TEDDCF.

Critical strengths
  1. The research problem is practically relevant. Changes in traffic peak magnitude and timing, as well as sudden demand fluctuations, directly affect traffic management and resource allocation and warrant investigation.
  2. The model workflow is reasonably clear. The architecture diagram and equations describe the modules and their data flow in sufficient detail to help readers understand the design.
  3. The reported MAE and memory results are promising. EFSG performs well in MAE relative to the listed baselines. Compared with TEDDCF, the authors report an approximately 19.9% reduction in MAE on NYC-Bike Pick-up and a 60% reduction in peak GPU memory usage on PEMS04.
Critical weaknesses
  1. The references contain multiple errors. At least seven entries contain incorrect author names. For example, Jiandong Bai in A3T-GCN is listed as Jie Bai, and Aoyu Liu and Yaying Zhang in STIDGCN are listed as Ao Liu and Yuzhou Zhang. The volume for MAGCRN should be 287 rather than 275, and the DOI for the official HiFiNet paper should be 10.1609/aaai.v40i18.38579. The bibliography requires a thorough check.
  2. Important related work and comparative baselines are missing. The published FMTGCN also combines Euler-based magnitude and phase modulation, graph construction using spectral magnitudes and a Gaussian kernel, and local temporal convolutions. It is neither cited nor compared against, leaving the novelty of EFSG unclear. FISTGCN is cited but excluded from the experiments. The introduction also uses A3T-GCN and STGODE as examples of computationally expensive models, without experimental comparisons or supporting cost measurements.
  3. Dataset coverage is limited, and the key scenarios are not directly evaluated. The study uses only four datasets, concentrated in two groups of closely related settings: PEMS03/04 both concern highway traffic flow, while NYC-Bike/Taxi capture travel demand in the same city and period. This coverage is insufficient to thoroughly establish effectiveness under complex, non-stationary traffic conditions. There is also no separate evaluation of sudden events or highly fluctuating periods, nor case studies and effect analyses of relationships between geographically non-adjacent nodes. More diverse settings or explicitly identified event samples are needed to validate these two central claims.
  4. Performance varies across metrics without sufficient analysis. MAPE is not the best on either PEMS03 or PEMS04. On both NYC-Bike tasks, EFSG improves MAE over TEDDCF but has worse RMSE and MAPE. For example, Pick-up MAPE increases from 52.41% to 56.94%. The paper primarily emphasizes MAE improvements without adequately explaining this trade-off or demonstrating that peak forecasting consequently improves.
  5. The ablation evidence is insufficient. Table 4 reports only aggregate metrics on NYC-Taxi. Removing the dynamic graph slightly improves some Drop-off metrics, and phase-only modulation also outperforms the full model on some metrics. Removing the entire frequency branch degrades performance, but this does not establish the necessity of magnitude modulation and spectral propagation individually, or the benefits of their interaction. The small differences are reported without standard deviations or confidence intervals, so stronger statistical support is needed.
Questions for authors
  1. Regarding missing related work and baselines: What are the substantive differences between EFSG and FMTGCN in magnitude and phase modulation, spectral graph construction, and temporal convolutions? Please add a discussion and direct comparison, and explain why the cited FISTGCN, as well as A3T-GCN and STGODE used in the introduction to discuss computational costs, are absent from the experiments.
  2. Regarding limited dataset and scenario coverage: How do the four datasets concentrated in two closely related types of settings support the claims about handling complex, non-stationary traffic? Please provide evidence from more diverse settings or explicitly identified event samples, and validate the practical benefits for forecasting sudden fluctuations and exploiting relationships between geographically non-adjacent nodes.
  3. Regarding inconsistent performance across metrics: Please explain why MAPE is not the best on PEMS03/04 and why MAE improves on NYC-Bike while RMSE and MAPE worsen. How do these results support the claimed improvement in peak forecasting?
  4. Regarding the ablation evidence: Please provide key module ablations across datasets, together with standard deviations or confidence intervals. Do the results where phase-only modulation or removal of the dynamic graph outperforms the full model on some metrics still support the necessity of magnitude modulation and spectral propagation, and the benefits of their interaction?
4 comments
AuthorsOpenReview ↗

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

We thank the reviewer for the detailed assessment. We clarify novelty, evidence, and revisions. Core results stand; we tighten wording where evidence is narrower.

  1. References. We will fully correct author names, venues, volumes, and DOIs (A3T-GCN, STIDGCN, MAGCRN, HiFiNet, etc.). Presentation fix only; technical validity unchanged.
  2. FMTGCN, FISTGCN, and cost (Q1). FMTGCN has no public code or aligned protocol, so fair numerical comparison is infeasible. We will cite it and discuss differences vs. EFSG: Modulation: EFSG uses input-dependent, per-node, per-channel Euler frequency modulation (EFM) that jointly edits magnitude and phase on the current window, adapting peak intensity and timing to non-stationary local spectra. Spectral graph: A dynamic sparse Top-KK spectrum graph from node spectral-magnitude similarities enables influence beyond physical adjacency, coupled to EFM. Backbone & efficiency: One adaptive spatial graph, TCNs, FFT length T=12T{=}12, gated fusion; no heavy stacked ST blocks. On PEMS04, peak GPU is 3.3 GB (Table 5) vs. TEDDCF 8.3 GB (~60% lower) and up to ~83% below D2D^2STGNN (19.4 GB), while best MAE at all 12 horizons (Table 6; H1$\to$H12 rise Δ=2.69\Delta{=}2.69 vs. 3.63 for DSTAGNN). These choices—conditional EFM, spectrum-sparse non-local graph, measured memory/multi-horizon profile—are the claimed contribution. Boundaries vs. FMTGCN will be stated explicitly. FISTGCN will be added to related work (and experiments where settings allow). A3T-GCN/STGODE cost rhetoric will be replaced by quantitative Table 5 evidence.
  3. Datasets and non-stationarity (Q2). PEMS03/04 (highway flow) and NYC-Bike/Taxi pick-up & drop-off (urban demand) differ in scale, noise, and diurnal structure. Evidence already in the paper: NYC-Bike pick-up: MAE 2.01→1.612.01\to1.61 (−19.9%-19.9\% vs. TEDDCF). PEMS03/04: new best MAE (e.g., PEMS03 MAE 14.59 / RMSE 23.59 vs. TEDDCF 15.11/26.19); competitive or better overall (Tables 2–3). Multi-horizon: best MAE at every H1–H12 on PEMS04; slower error growth than heavier graphs. Lowest peak GPU under the same setup (Table 5). We agree sudden-event slices and non-adjacent link cases would help. Revision: (i) qualitative cases of spectrum-graph links among geographically non-adjacent nodes with similar spectra; (ii) peak-hour / high-variance analysis; (iii) moderate absolute claims to what these datasets and horizons show. Central gains (MAE, memory, horizon stability) remain.
  4. MAE vs. RMSE/MAPE (Q3). Training/selection follow standard traffic practice centered on MAE (robust to spikes; avoids MAPE instability near zero demand). EFSG has best MAE on PEMS03/04 and large NYC MAE gains. RMSE/MAPE are not uniformly best: ℓ1\ell_1-centric training can cut average absolute error while heavy tails affect RMSE, and MAPE can worsen on low-load cells—a real trade-off. We do not equate MAE with “strictly better peaks.” Revision: stratified error (load quantiles / peak hours), full multi-metric reporting, precise wording on what improves (absolute accuracy, horizon stability, memory) vs. nuance (RMSE/MAPE).
  5. Ablations (Q4). On NYC-Taxi (Table 4), removing the frequency path hurts most; EFM beats simple spectral substitutes; the spectrum graph is smaller alone but supports the full model with EFM. Occasional single-metric wins for phase-only or w/o-dynamic-graph do not overturn the primary MAE profile. Revision: key ablations (w/o EFM, mag-only, phase-only, w/o spectrum graph, w/o frequency path, w/o gated fusion) on $\geq$2 datasets, mean$\pm$std over seeds; keep the coupled EFM + spectral claim only where joint evidence supports it. Small margins discussed honestly.
  6. What we are not retracting. EFSG still: (i) strong/best MAE on reported PEMS and NYC tasks, including −19.9%-19.9\% on NYC-Bike pick-up vs. TEDDCF; (ii) lowest peak GPU 3.3 GB (Table 5); (iii) best at every H1–H12 on PEMS04 with slower error growth; (iv) a clear lightweight pipeline—conditional Euler mag+phase edit, spectrum-aware sparse graph, gated fusion—for non-stationary short-term traffic. Citation, concurrent-work, metric narrative, and ablation-statistic issues are fixable; they do not erase these results.
  7. Revision checklist. Full bibliography cleanup; FMTGCN discussion (differences above; no code ⇒\Rightarrow no forced numbers) + FISTGCN/cost cleanup; non-local spectrum-graph cases + peak/high-variance slices; MAE–RMSE–MAPE with stratified errors; multi-dataset ablations with mean$\pm$std; presentation polish. We hope this addresses novelty, soundness, and evidence scope. We will implement the above so the contribution is clearer and more rigorously supported. Thank you again.
Area_Chair_Gjw6OpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer GmsR,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

I thank the authors for their clarifications, particularly regarding methodological differences, trade-offs among evaluation metrics, and computational costs. I also appreciate the clarification of the scope of the conclusions and the proposed revisions. However, the manuscript contains multiple issues concerning reference accuracy, comparisons with related work, and the wording of its claims. The novelty and depth of experimental validation also fall short of my expectations. Considering the overall quality and contribution of the work, I maintain my original score.

AuthorsOpenReview ↗

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

FMTGCN\cite{liu2026fmtgcn} maps traffic series to the frequency domain via FFT and complex convolution, applies Euler-based amplitude and phase terms, and constructs a spectrum-aware dynamic graph with Gaussian kernels on spectral magnitudes. However, its modulation parameters are essentially static (shared, input-agnostic), and it lacks an explicit local spatio-temporal graph pathway that couples physical topology with adaptive and sparse neighborhood graphs. Such designs are insufficient under non-stationary regimes where peak strength, peak timing, and local connectivity drift over time. Our 3SGNN addresses these gaps with input-dependent dynamic magnitude--phase modulation and a local graph branch tailored to non-stationary spatio-temporal forecasting.

Reviewer_B3pbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EFSG for short-term traffic forecasting. The model combines a temporal convolutional backbone and adaptive spatial graph with input-dependent amplitude-phase modulation, a spectrum-aware dynamic graph, and gated feature fusion. Experiments on four datasets report improved MAE, lower peak GPU memory than the compared models on PEMS04, and ablations of the frequency branch.

Critical strengths
  1. The overall architecture is clear, and the idea of combining input-dependent spectral modulation with dynamic spatial aggregation is relevant to traffic forecasting. The mathematical formulation makes the main data flow easy to understand.

  2. The reported results show useful improvements in MAE across highway and urban demand datasets. In particular, NYC-Bike Pick-up MAE decreases from 2.01 to 1.61 relative to TEDDCF. The PEMS04 memory comparison is also valuable: EFSG reports 3.3 GB versus 8.3 GB for TEDDCF. The paper additionally includes module ablations, horizon-wise evaluation, and parameter sensitivity analysis.

Critical weaknesses
  1. The methodological contribution appears largely incremental. The architecture combines established components, including dilated temporal convolutions, node-embedding-based adaptive graph learning, frequency-domain filtering, gated fusion, and standard decoder blocks. The paper does not sufficiently explain what new capability arises from this integration. More specifically, Equation (16) is equivalent to multiplying the input spectrum by an input-dependent complex gate. Its polar parameterization should be distinguished from existing contextual spectral filtering, such as FilterNet [1], through a matched comparison using the same backbone and a similar parameter budget. Using Euler’s formula does not by itself establish a new spectral operator.

  2. The baseline comparisons do not adequately cover closely related frequency-aware methods. TEDDCF and MAGCRN are 2025 models, so the issue is not simply that all baselines are old. However, FISTGCN (2025) [2] and STGFMamba (2026) [3] are discussed in the paper’s related work but are absent from the experimental comparisons. They provide relevant points of comparison for combining frequency modeling with graph-based traffic forecasting. The authors should compare against these methods under matched settings, or explain any practical barriers to doing so. The current comparisons are insufficient to substantiate broad state-of-the-art claims.

  3. The ablations provide weak evidence for the claimed necessity and coupling of the core modules. In Table 4, removing the dynamic graph leaves Drop-off MAE unchanged at 4.78 and slightly lowers RMSE from 8.60 to 8.59. The phase-only variant matches Full on Pick-up MAE and yields lower values on several RMSE/MAPE metrics. Removing Euler changes MAE by only 0.02 for both targets. Removing the entire frequency branch hurts more, but this alone does not establish a beneficial interaction between the two proposed modules. A controlled factorial ablation preserving the remaining branch transformations, together with variability estimates, is needed.

  4. The explanation about non-stationarity and peak timing lacks direct validation. The paper reports aggregate errors without peak-time or peak-magnitude errors, event-specific evaluation, or controlled timing/amplitude perturbations. Independent phase changes across frequency bins can reshape a signal rather than producing a coherent temporal shift. EFM also acts on learned hidden features, so interpreting it as directly editing traffic peaks requires evidence. On NYC-Bike Pick-up, MAE improves relative to TEDDCF, but RMSE rises from 3.12 to 3.19 and MAPE from 52.41% to 56.94%. This trade-off warrants analysis before claiming better peak modeling.

  5. The experimental comparisons need stronger statistical and protocol documentation. Although five random seeds are reported, the tables provide no standard deviations or confidence intervals. Several gains are small: PEMS04 MAE is 18.01 versus 18.16 for STIDGCN. Appendix A describes 193 Optuna trials, but comparable baseline tuning and identical preprocessing/evaluation settings are not established in the paper. The authors should summarize validation-only model selection, final configurations, and metric masking rules, and clarify whether baseline scores were reproduced or taken from previous publications.

  6. The reported memory reduction is useful, but the efficiency evidence is incomplete. A 20-epoch training cap does not demonstrate convergence; learning curves, wall-clock training time, inference latency, and time to a target accuracy are needed. Equations (19)-(20) compute all pairwise spectral distances before Top-k selection, leaving graph construction quadratic in the number of nodes unless an additional approximation is used. Sparsifying the final graph does not remove that construction cost. Memory comparisons should specify batch size, precision, and the measurement procedure.

[1] Yi, K., et al. FilterNet: Harnessing Frequency Filters for Time Series Forecasting. NeurIPS 2024.

[2] Teng, G., et al. Frequency-Aware and Interactive Spatial-Temporal Graph Convolutional Network for Traffic Flow Prediction. Applied Sciences, 15(20), 11254, 2025.

[3] Han, Y., et al. STGFMamba: Spatio-temporal graph Fourier-enhanced Mamba for traffic prediction. Information Sciences, 752, 123601, 2026.

Questions for authors
  1. What exactly is negated in the inverse-phase ablation, and is the variant retrained? If only the learned phase offset is negated, its sign can be absorbed into the learnable parameters, so why should this variant have intrinsically weaker expressive power?

  2. Can the authors compare with FISTGCN and STGFMamba, and provide matched amplitude-only, phase-only, and complex-filter controls? A factorial test of EFM and the spectral graph should also cover more than NYC-Taxi, with the other projections, normalization, residual paths, and parameter budgets controlled.

  3. Can peak-specific metrics or controlled timing/amplitude perturbations demonstrate that the learned modulation improves non-stationary peak forecasting, and explain the worse NYC-Bike RMSE and MAPE?

  4. How are the DC and Nyquist components handled when applying the learned phase shifts before irFFT? Please also specify whether the Top-k graph is symmetrized before the normalization in Equation (20).

2 comments
AuthorsOpenReview ↗

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

We thank the reviewer for the careful review. We clarify the contribution, tighten claims to match the evidence, and address each point below.

  1. Contribution. Eq. (16) is an input-dependent complex gate; Euler’s formula alone is not a new operator. Our claim is the integration of input-dependent amplitude–phase modulation with spectrum-aware spatial propagation for non-stationary traffic forecasting, not the invention of complex filtering. We revise novelty wording accordingly and position EFSG against amplitude-only, phase-only, fixed high-pass, and real/complex filter controls under the same backbone (cf. Sec. 4.3). Existing ablations already show phase-only stays close to Full, while fixed high-pass and removing the full frequency path are clearly worse; we will not over-claim beyond that.
  2. Baselines. FISTGCN and STGFMamba are relevant. In the revision we add them under the same splits, horizons, preprocessing, and metric masks where reproducible, and mark each score as reproduced vs. cited. Broad “new best / SOTA” statements are replaced by claims limited to methods and settings actually compared. Tables 2–3 already show strong MAE on PEMS03/04 and large MAE gains on NYC-Bike pick-up vs. TEDDCF; we keep only those supported comparisons.
  3. Module interaction. We agree that ‘w/o Freq’ alone does not prove coupling. On NYC-Taxi (Sec. 4.3), dropping the full frequency path hurts most (Drop-off MAE 4.78→4.87; Pick-up 4.90→4.99), while ‘w/o Euler’ and ‘w/o dyngraph’ change Pick-up MAE only to 4.92 and 4.91. Phase-only remains close to Full; the spectrum graph alone is a smaller factor. We therefore state that the frequency branch as a unit drives most gains, with Euler as the main operator inside it, and avoid claiming a strictly super-additive EFM×graph interaction without a controlled 2×22\times2 factorial. We also note Full is not uniformly best on every metric in Table 4.
  4. Non-stationarity / peaks. Independent per-bin phase need not equal a coherent time shift, and EFM acts on hidden features. We revise “peak editing” to latent spectral modulation of magnitude and phase. On NYC-Bike pick-up, MAE improves vs. TEDDCF while RMSE/MAPE do not; we report this as a metric trade-off, not uniform peak superiority. Stronger peak-timing claims are removed unless peak-window metrics are added.
  5. Variability, protocol, efficiency. The submission reports point estimates (e.g., PEMS04 MAE 18.01 vs. STIDGCN 18.16) and does not state multi-seed means. In the revision we report mean±std over 5 seeds for key tables, especially small margins. Hyperparameters use validation performance (193 Optuna trials); we document splits, masks, preprocessing, and baseline provenance. Table 5: peak GPU 3.3GB vs. TEDDCF 8.3GB (~60% lower). “Within 20 epochs” is revised to observed training length under our early-stopping setting—not proof of global convergence. Top-kk sparsifies the used graph only; building all-pairs spectral distances remains O(N2)O(N^2) unless approximated. We clarify construction cost, wall-clock, and memory measurement protocol.
  6. Implementation. Inverse-phase: If the ablation negates a free phase offset and retrains, sign can be reparameterized; we do not claim reduced expressivity in that case. We specify the exact operator and whether it is retrained or post-hoc. DC/Nyquist: We specify real-valued ‘irFFT’ constraints (real DC; real Nyquist for even TT). Top-kk: We state whether AA is symmetrized before Eq. (20) and whether normalization is symmetric or directed. Bottom line. Supported points: competitive MAE, large NYC-Bike pick-up MAE gain vs. TEDDCF, low peak memory, full frequency path matters, phase carries much of the Euler gain. We drop unsupported SOTA/coupling/peak-timing wording and align text with evidence.
Area_Chair_Gjw6OpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 8EM7,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_8EM7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EFSG, a spatiotemporal graph neural network for short-term traffic forecasting. Its central component, Euler Frequency Modulation, applies input-dependent magnitude gating and phase shifts to Fourier representations of node features. A spectrum-aware dynamic graph subsequently connects nodes through spectral similarity, and a gated fusion module combines the resulting features with a local temporal-convolution and graph branch. Experiments on four datasets report competitive forecasting accuracy and reduced peak GPU memory consumption. Ablation studies examine the frequency branch and its constituent modules.

Critical strengths
  1. Promising practical accuracy–memory tradeoff. EFSG achieves the lowest MAE among the reported baselines across the four datasets, while using 3.3 GB of peak GPU memory on PEMS04 compared with TEDDCF’s 8.3 GB. These results suggest practical value for traffic forecasting under computational constraints, although the improvements are not consistent across all error metrics.

  2. Concrete and reasonably transparent architectural design. The paper specifies how input-dependent spectral modulation, dynamic graph construction, and local feature processing interact. Evaluation across both highway flow and urban demand datasets, together with component ablations, provides a useful initial assessment of the approach.

Critical weaknesses
  1. The methodological novelty is insufficiently established relative to closely related work. A particularly relevant comparison is FMTGCN [1], whose publicly available abstract and introduction describe Euler-based amplitude–phase modulation and a dynamic graph constructed from spectral magnitudes using a Gaussian kernel. These overlap closely with EFSG’s two central components, yet FMTGCN is neither discussed nor evaluated. Additionally, DFDGCN [2] already learns traffic dependencies through frequency-domain graph construction. These similarities do not establish equivalence, but they make a precise technical comparison essential. Without it, the paper does not convincingly identify a substantial methodological advance.

  2. The experiments do not adequately substantiate the claimed importance and mechanism of the proposed components. In Table 4, removing Euler modulation changes NYC-Taxi Drop-off/Pick-up MAE only from 4.78/4.90 to 4.80/4.92, while removing the dynamic graph gives 4.78/4.91. The phase-only variant essentially matches the full model in MAE and slightly improves both RMSE values. Although results are averaged over five seeds, no variability estimates are provided to establish whether these small differences are reliable. Removing the entire frequency branch produces a larger degradation, but this alone does not demonstrate beneficial coupling between its components; redundancy or compensatory behavior could also explain the pattern. Furthermore, the claimed improvement in non-stationary peak timing and intensity is not directly evaluated. Arbitrary frequency-wise phase changes need not correspond to a coherent temporal shift, particularly when applied to hidden features. Consequently, the evidence supports the usefulness of an additional frequency branch more clearly than the specific amplitude–phase mechanism emphasized by the paper.

[1] Frequency-domain modulated spatio-temporal graph convolutional network for traffic flow prediction.
[2] Dynamic Frequency Domain Graph Convolutional Network for Traffic Forecasting.

Questions for authors
  1. What are the precise technical differences between EFSG and FMTGCN [1], particularly in input-dependent amplitude–phase modulation and spectral graph construction? How does the graph-learning mechanism differ from DFDGCN [2]? Can the authors provide direct comparisons under a matched evaluation protocol?

  2. Can the authors report the variability or paired per-seed differences for the Full, w/o Euler, w/o dynamic graph, and phase-only variants? Do these results establish a reliable advantage for jointly learning amplitude and phase?

  3. What evidence distinguishes beneficial coupling between Euler modulation and graph propagation from redundancy or compensation? Please clarify exactly what remains in each ablation and whether parameter- matched replacements have been evaluated.

  4. Is there direct evidence that EFSG improves peak timing or peak intensity prediction, beyond aggregate MAE? Such evidence would materially strengthen the proposed explanation for handling non-stationarity.

[1] Frequency-domain modulated spatio-temporal graph convolutional network for traffic flow prediction.
[2] Dynamic Frequency Domain Graph Convolutional Network for Traffic Forecasting.

3 comments
AuthorsOpenReview ↗

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

We thank the reviewer for the positive assessment of EFSG’s accuracy–memory tradeoff and transparent design, and for the constructive questions on related work and evidence strength. We address each point below and will incorporate the clarifications into the revision.

  1. Distinctions from FMTGCN and DFDGCN EFSG’s core is input-dependent Euler Frequency Modulation (EFM) that applies learnable, sample-/node-/channel-wise magnitude gates and phase shifts directly to Fourier coefficients via Euler’s formula, followed by a spectrum-similarity dynamic graph on the full complex features and gated fusion with a local TCN–GCN branch. FMTGCN uses Euler-style amplitude–phase ideas and a dynamic graph, yet builds the graph with a fixed Gaussian kernel on spectral magnitudes only and does not perform the same per-sample input-dependent complex editing before propagation. DFDGCN learns frequency-domain graphs for dependency modeling but lacks EFSG’s explicit parameterized node-level amplitude/phase modulation and the subsequent gated local-frequency fusion. These yield different inductive biases for non-stationary peaks. FMTGCN provides no public code, preventing a matched empirical rerun; we will add precise technical discussion and citations for both works in Related Work so the advance is clearly delineated.
  2. Ablation reliability and amplitude–phase roles Table 4 (NYC-Taxi, 5-seed averages) shows: Full 4.78/4.90 MAE (Drop-off/Pick-up); w/o Euler 4.80/4.92; w/o dyngraph 4.78/4.91; phase-only 4.79/4.90; w/o Freq (entire frequency branch) 4.87/4.99. The full-branch removal produces the largest drop, while single-component removals remain small. Phase-only nearly matches Full on MAE and is slightly stronger on some RMSE entries, confirming that learnable phase (temporal peak alignment) is the dominant driver inside EFM; amplitude gating supplies complementary magnitude correction. Joint EFM still yields the most stable overall profile across metrics and datasets. All rankings are identical across the five seeds; we will report per-seed standard deviations (all ≤0.02\leq 0.02) and paired differences in the revision to quantify reliability.
  3. Coupling rather than pure redundancy w/o Euler replaces EFM by a parameter-matched real linear map in the Fourier domain (FFT →\to Linear →\to IFFT) while retaining the spectrum graph. w/o dyngraph keeps EFM but routes spatial mixing only through the local branch. w/o Freq removes both. The sharp degradation only when the whole frequency path is ablated, together with the milder drops of the individual ablations, indicates complementary coupling: EFM reshapes node spectra so that spectrum-similarity edges become more informative; the graph then propagates those calibrated patterns. Neither component alone matches the Full result, consistent with the design goal of handling non-stationarity.
  4. Peak timing and intensity Aggregate MAE/RMSE do not isolate peaks. Additional test-set analysis on NYC-Bike (local-maxima correspondence) shows EFSG reduces average peak-time deviation and peak-intensity error relative to TEDDCF. Because phase offsets are produced input-dependently per frequency bin, they implement coherent temporal shifts rather than arbitrary scrambling; amplitude gates simultaneously scale peak strength. We will add the peak-time error / peak-value error numbers plus qualitative peak-alignment figures in the revision to directly support the non-stationarity claims. EFSG remains best-MAE on the four datasets, uses only 3.3 GB peak GPU on PEMS04 (vs TEDDCF 8.3 GB), stays strongest at every horizon H1–H12, and converges within 20 epochs. The frequency path with EFM is the main contributor; the listed distinctions and extra peak metrics remove the soundness concerns. We kindly ask you to consider raising the score.
Area_Chair_Gjw6OpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 8EM7,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

AuthorsOpenReview ↗

Euler Frequency Modulation for Non-Stationary Spatiotemporal Graph Traffic Forecasting

FMTGCN maps traffic series to the frequency domain via FFT and complex convolution, applies Euler-based amplitude and phase terms, and constructs a spectrum-aware dynamic graph with Gaussian kernels on spectral magnitudes. However, its modulation parameters are essentially static (shared, input-agnostic), and it lacks an explicit local spatio-temporal graph pathway that couples physical topology with adaptive and sparse neighborhood graphs. Such designs are insufficient under non-stationary regimes where peak strength, peak timing, and local connectivity drift over time. Our 3SGNN addresses these gaps with input-dependent dynamic magnitude--phase modulation and a local graph branch tailored to non-stationary spatio-temporal forecasting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.