acceptodds
Under review as a conference paper at ICLR 2027

Not All Frequencies Are Equal: Asymmetric Spectral Modeling for Cloud Workload Forecasting

Abstract

Workload forecasting is essential for cloud resource management, yet real-world workloads often exhibit complicated periodic patterns across multiple temporal scales. Although frequency-domain forecasting provides an effective representation for such patterns, existing methods commonly apply homogeneous transformations across the spectrum, overlooking the distinct structures of low- and high-frequency components. To address this issue, we propose DuoFreq, a lightweight frequency-domain forecasting framework based on Dual-Band Spectral Modulation (DBSM). DuoFreq separates the spectrum into low- and high-frequency bands and models them with specialized mechanisms. Smooth Low-Frequency Modulation (SLFM) preserves and adapts low-frequency structures through smooth residual modulation, while Local-Global High-Frequency Modulation (LG-HFM) captures neighboring and band-wide dependencies, together with Complex Spectral Transformation (CST) for complex-valued high-frequency modeling. Extensive experiments on five real-world cloud workload datasets show that DuoFreq achieves the best MSE in 17 of 20 forecasting settings and the best MAE in 18, with average reductions of 9.48% and 7.09%, respectively. DuoFreq also generalizes well to unseen workload instances while maintaining a favorable accuracy–efficiency trade-off with approximately 0.61M parameters.

Primary area:
learning on time series and dynamical systems
Keywords:
Cloud workload forecasting, Time series forecasting, Frequency-domain learning, Spectral modeling, Asymmetric frequency modeling

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

The paper starts from the claim in its title: the low- and high-frequency parts of a workload spectrum have different structure and should not be handled by the same operator. DuoFreq implements this by linearly extrapolating the lookback into the horizon, taking an rFFT over the concatenation of history and extrapolation, and then applying a smooth identity-centred real gain to the low band (SLFM) and local/global gains plus complex linear blocks to the high band (LG-HFM, CST). Experiments use five cloud traces, eleven baselines and four horizons, plus ablations and an efficiency study; the model has about 0.61M parameters and reports the best MSE in 17 of 20 settings.

Critical strengths
  1. Well written. The motivation, the two-band design and the figures are easy to follow, and the paper reads cleanly from start to finish.
  2. Targets a real problem and stays lightweight. Cloud workload forecasting underpins auto-scaling and resource provisioning, and at 0.6M parameters the model is small enough to be deployed per service.
Critical weaknesses
  1. The transfer results are hard to accept as they stand. RQ3 evaluates on variables that were not used in training, so it should not come out easier than RQ1, which trains and tests on the same variables. Yet the unseen-half error is lower on four of five datasets (Huawei Memory 0.7068 vs 1.1057; Huawei CPU 0.4964 vs 0.6269; FaaS 0.1968 vs 0.2120; PaaS 0.0458 vs 0.0589). That direction is not explained.
  2. The role of SLFM needs stronger support. Removing it costs only 0.2–0.7% MAE, the smallest drop among the three proposed modules on four of five datasets. So a dedicated low-frequency module does not clearly help.
Questions for authors
  1. Unify the setting for the seen and unseen variables, especially the normalisation. Then explain why the seen numbers come out worse than the unseen ones.
  2. Give stronger support for the role and significance of SLFM.
1 comment

Thank you for pointing out the difference between the reported RQ1 and RQ3 errors and for asking us to clarify the role of SLFM. We have expanded the transfer evaluation and added two structural comparisons and a zero-cutoff experiment.

1. Why the reported error on unseen variables can be lower

RQ1 reports the average error over all variables in each dataset. The original RQ3 experiment reports the average only over the held-out second half. Workload variables differ in their scales and patterns of variation. Both evaluations use the same per-variable standardization rule, so differences in raw scale are accounted for. However, standardized forecast errors can still vary considerably across variables. A lower error averaged over the held-out subset can therefore arise from which variables enter the average and does not imply that forecasting unseen variables is easier.

For each variable, we compute the standardization statistics from the first 70% of its timeline. In RQ3, the historical observations of the target variables are used only for standardization. Model training, validation, and checkpoint selection use source variables only.

The revised Table 4 averages results across the original partition and two additional random partitions, shared by all nine models. On FaaS, the transfer MSE changes from 0.1968 under the original partition to 0.2127 after averaging across the three partitions, close to the RQ1 value of 0.2120. On PaaS, it changes from 0.0458 to 0.0590, close to the RQ1 value of 0.0589. These results show that the gap between the two reported errors depends on the choice of target variables.

2. What SLFM contributes

The SLFM ablation bypasses the gain predictor and removes the smoothness penalty. The low-frequency coefficients enter reconstruction unchanged, so the comparison isolates the benefit of learning to refine them. The improvement is clearest on Huawei Memory, where removing SLFM raises the horizon-averaged MSE from 1.1057 to 1.1283 and MAE from 0.5481 to 0.5521. On Huawei CPU, the average MSE rises from 0.6269 to 0.6336.

The two new structural comparisons test how the bands should be processed. Symmetric-HFM-CST retains the band split and applies the high-frequency pipeline separately to both bands. FullSpectrum-HFM-CST applies the same pipeline to the entire spectrum. On Huawei CPU, their horizon-averaged MSEs are 0.6650 and 0.6497, respectively, compared with 0.6269 for DuoFreq. The symmetric comparison is particularly relevant here because both designs retain the two bands and differ in how they process the low band.

We also added a zero-cutoff setting that keeps DC protection and removes the separate low-frequency band. On Huawei CPU at a forecast horizon of 288 steps, all tested settings that assign a low-frequency band to SLFM achieve lower validation and test MSE than the zero-cutoff setting. Together, these results support refining the retained low-frequency coefficients with smooth residual gains and giving the two bands different transformations.

Reviewer_LyK9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DuoFreq, a lightweight frequency-domain framework for cloud workload forecasting. It first generates an initial future estimate through Temporal Extension (TE), then transforms the extended sequence into the frequency domain and separates low- and high-frequency components. Smooth Low-Frequency Modulation (SLFM) applies smooth residual modulation to low frequencies, while Local-Global High-Frequency Modulation (LG-HFM) and Complex Spectral Transformation (CST) model high-frequency dependencies. The processed spectrum is transformed back into the time domain to produce forecasts. Experiments on five datasets and four forecast horizons show competitive performance against eleven baselines, supported by ablation, sensitivity, transfer, and efficiency analyses.

Critical strengths
  1. Clear motivation and presentation. The rationale for treating low- and high-frequency components differently is intuitive, and the paper is well organized and easy to follow.
  2. Broad empirical evaluation. The method achieves strong results across five cloud workload datasets and multiple forecasting horizons. The paper also includes useful ablation, sensitivity, generalization, and efficiency studies.
Critical weaknesses
  1. Insufficient differentiation from prior work. The novelty of DuoFreq relative to prior frequency-domain forecasting methods is not sufficiently clarified. In particular, Fremer (Ye et al., PVLDB 2025) introduces Learnable Linear Padding (LLP) and Frequency Reversible Instance Normalization (F-RIN), which are closely related to DuoFreq's Temporal Extension (TE) and spectral normalization, respectively. Both methods also employ frequency-specific processing and complex-valued spectral modeling. Although their architectures differ, the incremental contribution remains unclear, and a direct comparison with Fremer is missing.

    Reference: Hengyu Ye et al., Fremer: Lightweight and Effective Frequency Transformer for Workload Forecasting in Cloud Services. Proceedings of the VLDB Endowment, 18(11):3812–3825, 2025. DOI: 10.14778/3749646.3749656.

  2. Lack of direct evidence for asymmetric modeling. Existing ablations remove individual components but do not compare the proposed design with symmetric alternatives under comparable model capacity. Thus, the necessity of asymmetric spectral modeling remains insufficiently validated.

  3. Limited evaluation of generalization. Although the paper demonstrates promising performance on unseen workload instances, the transfer experiments are restricted to a single fixed split of variables within each dataset, with equal numbers of source and target variables. It remains unclear whether the reported generalization performance holds under different variable partitions or across heterogeneous workload datasets.

Questions for authors

Please see weakness.

3 comments

Thank you for pointing out the overlap with Fremer and for suggesting a direct comparison with symmetric models. We have uploaded a revised manuscript that addresses these points and expands the transfer evaluation.

1. Comparison with Fremer

Fremer’s Learnable Linear Padding and F-RIN are closely related to our temporal extension and spectral normalization. The main difference comes after these steps. Fremer preserves the low-frequency coefficients and uses attention over frequency combinations. DuoFreq learns small, smooth adjustments to the low-frequency coefficients. In the high band, it combines information from neighboring frequencies with a summary of the whole band before applying a complex MLP.

Table 2 now compares the two methods across five datasets and four horizons. DuoFreq has lower horizon-averaged MSE and MAE than Fremer on all five datasets. On FaaS and PaaS, where Fremer is the strongest MSE baseline, DuoFreq reduces MSE by 3.42% and 5.00%, respectively.

2. Asymmetric modeling at comparable capacity

We added two comparisons to test whether the low and high bands benefit from different processing. The symmetric model applies the high-frequency pipeline separately to each band. The full-spectrum model applies the same pipeline once to the entire spectrum. Both use the same hidden widths as DuoFreq.

On Huawei CPU at P=288P=288, the full-spectrum model is close in size to DuoFreq (0.617M vs. 0.608M parameters), but its MSE is higher (0.8681 vs. 0.8502). The symmetric model has more parameters (0.749M) and gives an MSE of 0.8873. Across all five datasets and four horizons, DuoFreq has lower MSE than both alternatives in 18 of the 20 settings (Table 7).

3. Transfer across variable partitions

Table 4 now covers the original partition and two additional random partitions. All nine models use the same splits and four forecasting horizons. Training and checkpoint selection use only source variables. Each target variable’s normalization statistics come from its own historical training period.

Averaged over these partitions and horizons, DuoFreq has the lowest MSE on four of five datasets and the lowest MAE on all five. We kept the source and target groups equal in size because several baselines depend on the number of input variables. These results support transfer to unseen instances within a dataset. Transfer between different cloud datasets remains an open question.

Thank you also for flagging the missing AI Use Statement. It now appears on page 10, before the references.

We hope these comparisons address your concerns and that you’ll consider updating your score.

I have read through the response, and increased the score since my concerns are addressed.

Thank you for taking the time to review our response and the revised manuscript. We appreciate your constructive feedback, which helped us clarify the contribution and strengthen the evaluation.

Reviewer_TrG6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

DuoFreq first makes a linear forecast of the horizon from the centered lookback (a linear map from L to T, "Temporal Extension"), appends it to the history, and takes the rFFT. After normalizing the spectrum, the first KlK_l bins go through a near-identity gain (SLFM, with a smoothness penalty), and the remaining bins go through a gain from a depthwise convolution and a pooled MLP (LG-HFM) followed by a complex-valued MLP (CST). The inverse FFT gives the forecast. All weights are shared across channels. On five cloud workload datasets and four horizons it has the best average MSE on four of five datasets against eleven baselines, and it is also tested on unseen channels and for speed.

Critical strengths
  1. The model is small (0.61M parameters), and the timing protocol is careful: warm-up, CUDA events, and five repeated trials (Appendix A.5).
  2. The comparison is broad: five cloud datasets, eleven recent baselines and four horizons, with full per-horizon tables in the appendix.
  3. The method is easy to follow, with one equation per module, and the data protocol is clean: chronological splits, standardization from the training split only, and a fully specified unseen-instance protocol (Appendix A.4).
Critical weaknesses
  1. The model is built from known parts, and its two closest relatives are missing. The first step is a linear L-to-T forecast as in DLinear and RLinear. The rest is a band split, a gated identity on a few low bins, and a complex-valued frequency MLP already proposed by Yi et al. (2023), cited here. FP-MLP (Lu et al., 2025) already models spectral bands separately for cloud workloads, and Fremer (Ye et al., 2025) is the frequency model built for this task. Both are cited and neither is a baseline; FITS is missing too.

  2. The ablation does not test the title. The direct test, one module on the whole spectrum, is never run. Removing SLFM, the low-band half of the design, changes MSE by 0.4% to 2% (FaaS 0.2120 to 0.2130, IaaS 0.6925 to 0.6950; Table 3), and removing LG-HFM improves PaaS MAE (0.1342 vs. 0.1348). Moving the cutoff from 3 to 30 of 505 bins is "relatively stable", so where the split sits barely matters. The default cutoff, ρ\rho, λ\lambda and the CST depth are not given, and no variant removes Temporal Extension.

  3. Small single-run gains sit behind an inflated headline. The abstract's 9.48% and 7.09% are averages over all eleven baselines, weak ones included. Against the best baseline per dataset, the MSE gain is 2.6% to 5.3%, and on Huawei Memory DuoFreq loses to FreEformer (1.1057 vs. 1.0942; Table 2). Some wins are by 0.0001 (0.7067 vs. 0.7068; Table 6). There are no seeds or standard deviations, no word on how baselines were tuned, and no source or license for the Huawei data.

  4. The extra tables drop the strongest baseline, and the transfer test favors DuoFreq by design. Tables 4 and 5 omit TFDNet, the best baseline on FaaS (0.2203; Table 2), and the text then calls DUET "the strongest baseline on FaaS". The transfer test trains on half the channels and tests on the other half; DuoFreq shares weights across channels, while several baselines have parameters tied to channel positions (Appendix A.4). The channel-independent DLinear comes close (Huawei CPU 0.5092 vs. 0.4964). FilterNet and DLinear are smaller and 5 to 9 times faster at inference (Table 5), and no standard benchmark such as ETT or Weather is reported.

Questions for authors

See weaknesses.

Comments on the paper

2 comments

Thank you for the response. My central concern is the lack of substantive novelty: I do not see a meaningful new methodological contribution beyond incremental refinements of ideas already present in FP-MLP and Fremer. The additional experiments address some empirical concerns but do not resolve this limitation. Furthermore, selecting hyperparameters using test MSE compromises the affected evaluation. Given the unresolved novelty concern and this additional soundness issue, I therefore maintain my reject recommendation.

Thank you for clarifying your remaining concerns. For the main forecasting comparison in Table 2, we used grid search for DuoFreq and all baselines and selected each model’s configuration based on validation performance. The test set was used to evaluate the selected configurations. Our earlier wording did not accurately describe the selection procedure for this table.

What DuoFreq changes. The methodological contribution lies in what the model can learn in each frequency band. FP-MLP uses frequency patches to represent different regions of the spectrum. Fremer routes low-frequency coefficients around its prediction backbone and applies complex attention to frequency combinations. DuoFreq gives the two bands different forms of adaptation. In the low band, each coefficient is refined by a bounded residual gain that starts at one, with a fixed DC gain and a smoothness penalty across neighboring gains. In the high band, the model uses neighboring coefficients at the original spectral resolution and pooled information from across the band to compute a gain for each frequency. A complex residual MLP then mixes the modulated coefficients. The two stages have distinct roles: the gain adjusts how strongly each frequency enters the transformation for the current input, and the MLP learns how to combine information across frequencies.

Why the distinction matters. The structural comparisons apply the same high-frequency pipeline separately to both bands or once to the entire spectrum. On FaaS, the resulting average MSEs are 0.2321 and 0.2303, compared with 0.2120 for DuoFreq. On Huawei CPU at a forecast horizon of 288 steps, DuoFreq achieves an MSE of 0.8502 with 0.608M parameters, compared with 0.8681 and 0.617M parameters for the full-spectrum control. The SLFM ablation examines the contribution of low-frequency refinement more directly. On Huawei Memory, removing the learned gains while retaining the low-frequency coefficients raises average MSE from 1.1057 to 1.1283.

Accuracy and computational cost. In Table 2, DuoFreq achieves lower horizon-averaged MSE and MAE than Fremer on all five datasets. At a forecast horizon of 288 steps on FaaS, DuoFreq achieves an MSE of 0.2610 versus 0.2689 for Fremer, with 14.8% fewer parameters and 27.4% lower per-sample inference latency at a batch size of 32. The result shows the practical value of the design in this setting: constrained refinement in the low band and contextual mixing in the high band improve forecasting accuracy while reducing model size and inference time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.