acceptodds
Under review as a conference paper at ICLR 2027

SCPEC: State-Conditioned Pattern-Expert Composition for Workload Forecasting

Abstract

Workload forecasting in dynamic multi-tenant edge-cloud platforms is crucial for resource provisioning, application deployment, and service assurance. However, heterogeneity across nodes and workload dynamics within each node make accurate forecasting challenging. Recent studies have adopted unified forecasting models to handle both, but representative ones still rely on a prototype pool of stored historical patterns, which limits their generalization to evolving workloads. To address this limitation, we reformulate multi-node workload forecasting as a multi-view regression problem, in which each node’s workload state is jointly characterized through temporal, spectral, and static-context views, and reusable knowledge is selected conditioned on this state. Building on this formulation, we propose SCPEC, which replaces the historical prototype pool with state conditioned composition of learnable pattern experts, thereby improving both generalization and forecasting accuracy. Experiments on eight real-world workload, electricity, and traffic datasets show that SCPEC achieves the best or tied-best results on 15 of 16 MSE/MAE evaluation entries. Our code is publicly available at https://anonymous.4open.science/r/test-F170/.

Primary area:
learning on time series and dynamical systems
Keywords:
workload forecasting, time series forecasting

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_yY8LOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SCPEC, a unified workload forecasting framework designed to address the limitations of history-derived pattern reuse under heterogeneous and evolving workload conditions. The authors identify two challenges: insufficient historical pattern coverage and changes in the predictive applicability of similar historical patterns after application events. To address these challenges, SCPEC introduces a learnable pattern-expert bank shared across nodes and a state-conditioned sparse routing mechanism. The routing state integrates temporal dynamics, spectral information, and optional static node context. Selected pattern experts are composed and incorporated into Transformer representations through gated residual connections, followed by a horizon-query readout for multi-step forecasting.

Critical strengths
  1. The research motivation is supported by a well-grounded quantitative analysis. The paper provides a well-motivated analysis of the limitations of history-based pattern reuse under dynamic workload conditions. By quantifying historical pattern coverage and pattern-transfer ambiguity (PTA), the authors demonstrate how application-state changes can undermine the reliability of similarity-based retrieval. The additional historical-composition diagnostic and node-level bootstrap confidence intervals for the ECW–NewAPP comparisons further strengthen this analysis. These observations provide meaningful empirical motivation for the proposed state-conditioned pattern composition framework.
  2. The paper demonstrates strong forecasting performance, supported by informative ablation studies. SCPEC demonstrates strong forecasting performance across diverse datasets, achieving the best or tied-best results on 15 of 16 reported MSE/MAE entries at three-decimal precision against 20 baselines. The zero-shot evaluation on NewAPP and Switch using models trained on ECW without retraining further indicates its potential to handle application-state changes. The ablation studies consistently show performance degradation when removing pattern-expert composition or spectral information, while replacing the proposed knowledge module with Meta-pattern Pooling also reduces performance. These results provide encouraging evidence for the effectiveness of the overall framework.
Critical weaknesses
  1. The independent contributions of different method components have not yet been fully distinguished. Although the ablation studies demonstrate the effectiveness of the overall design, they do not fully disentangle the contributions of the learnable pattern bank, state-conditioned routing, and gated composition. In particular, the w/ MPP variant replaces all three mechanisms simultaneously, making it difficult to determine which component accounts for the observed improvement over history-derived pattern reuse. Furthermore, the expert bank employs an EMA warm-start based on observed training segments, and the influence of this initialization on generalization remains unclear. More fine-grained ablations would help substantiate the claimed advantages of learnable, state-conditioned pattern composition.
  2. The scope of generalization evaluation and experimental statistical stability still has limitations. The generalization evidence remains limited to a fixed 48-to-24 forecasting configuration with single-channel inputs. Moreover, the two application-event scenarios, NewAPP and Switch, are derived from ECW, and Switch contains only 12 valid forecasting windows under the reported protocol. This relatively small evaluation set limits the strength of conclusions about robustness to application switching. In addition, the main forecasting and ablation results are averaged over only three random seeds without reporting variability or statistical significance, making it difficult to assess the reliability of smaller improvements over competitive baselines.
Questions for authors

Q1. Could the authors provide more fine-grained ablations that separately evaluate the learnable pattern bank, state-conditioned routing, and gated composition? In particular, comparisons between learnable pattern experts with similarity-based routing and history-derived patterns with state-conditioned routing would help clarify the source of the improvements. Additionally, how does the EMA warm-start affect forecasting accuracy and generalization to unseen workload patterns? Q2. How does SCPEC perform under different look-back lengths and prediction horizons? Could the authors also provide additional application-event evaluations beyond NewAPP and Switch, particularly given the limited number of valid windows in Switch? Such experiments would help determine whether the observed generalization improvements extend to a broader range of workload dynamics. Q3. Could the authors report standard deviations or confidence intervals for the results in Tables 1 and 2 and conduct statistical significance tests against the strongest baselines? Since several reported improvements are relatively small, additional statistical evidence would help establish whether these gains are consistent across random seeds.

1 comment
AuthorsOpenReview ↗

Response to Reviewer yY8L: component ablations and EMA warm-start, other lengths and unseen nodes, and significance tests

We thank the reviewer for the constructive questions. All results below are in the revised PDF (App. B.3–B.6).

Q1 (W1): Component contributions and the EMA warm-start. We ran the requested ablations, which separate the three mechanisms that w/ MPP replaces at once, in the setting of Table 2 (same backbone, hyperparameters and budget; 3 seeds). Similarity selection scores each pattern by its Pearson correlation with the block's input segment; a history-initialized bank starts from E randomly drawn block-aligned training segments without warm-up and is then learned or kept fixed. MSE, mean of 3 seeds (MAE and stds in App. B.3, Table 8):

BankSelectionECWNewAPPSwitchECL270
SCPEC (learnable, warm-started)state0.06210.02920.05420.0382
learnable, warm-startedsimilarity0.06370.03040.05770.0394
history-init., learnedstate0.06480.03130.05830.0401
history-init., learnedsimilarity0.06610.03150.06020.0419
history-init., fixedstate0.06620.03180.05930.0421
history-init., fixedsimilarity0.06820.03310.06120.0438
SCPEC, fixed gate σ(6)state0.06550.03070.05980.0414
SCPEC, no warm-startstate0.06360.03010.05610.0394
  • Learnable experts + similarity vs. history-derived patterns + state routing (as requested): the former has 2.7%–6.4% lower MSE on all four datasets, and neither reaches SCPEC. With SCPEC's own bank, state routing beats similarity by 2.5%–6.1%, most on Switch (the ECW/ECL270 gaps are within seed variation).
  • Bank learning and routing are complementary. In the 2×2 block, each reduces MSE when the other is fixed (bank learning 1.6%–4.8%, state routing 0.6%–4.3%; 3 of these 16 gaps are within seed variation), and both together reduce it by 4.7%–8.4%.
  • Gated composition. Fixing the gate at its initial value σ(6) ≈ 0.9975 (each block almost replaced by its recovered patterns) raises MSE on all four datasets; the learned gate reduces it by 4.9%–9.4%.
  • EMA warm-start. It reduces MSE by 2.4% (ECW) and 3.0% (ECL270), and by 3.0% and 3.4% on NewAPP and Switch, evaluated without retraining; each gap exceeds the sum of the two seed stds. Without it, the bank never copies a training segment, yet it still beats every history-initialized variant and w/ MPP (Table 2) on all four datasets, so SCPEC's advantage does not rely on seeding the bank with observed segments.

Q2 (W2): Other lengths and evaluation beyond NewAPP/Switch.

  • Look-back and horizon (App. B.6, Table 14). We ran SCPEC, MetaEformer and FreqMoE with L ∈ {32, 48, 96} (H = 24) and H ∈ {24, 48, 72} (L = 48), trained on ECW and applied to NewAPP without retraining. SCPEC is best on ECW in all five settings (vs. the stronger baseline, MSE 4.2%–18.4% and MAE 2.9%–12.1% lower; 8 of 10 gaps p < 0.05) and beats MetaEformer in all 20 ECW/NewAPP entries (Wilcoxon p = 0.002). On NewAPP, FreqMoE is slightly better at 96→24, 48→48 and 48→72 (≤ 4.5% in MSE), never significantly (p ≥ 0.23), consistent with App. B.1's spectral analysis.
  • Unseen nodes (App. B.5, Table 13). As a further zero-shot target (new infrastructure rather than a new event type), on ECW-New Infras (3,906 windows from 651 edge nodes absent from ECW), ECW-trained checkpoints applied without retraining give MSE/MAE 0.0587/0.1361 for SCPEC vs. 0.0619/0.1379 for FreqMoE and 0.0651/0.1454 for MetaEformer; SCPEC's MSE is below every learned baseline at every seed.
  • Switch (12 windows; its 120-step series leave no window at 96→24 or 48→72) is read descriptively. Multichannel workloads and other platforms remain outside our scope (Sec. 6).

Q3 (W2): Variability and significance (App. B.4). Table 2 now reports mean ± std, and App. B.4 gives the full Table 1 with stds (Tables 9–10), per-seed values and Welch tests (Tables 11–12). While compiling them, we re-ran SAN on NewAPP/Switch and UniTime on ECL270 and added the missing Crossformer row; no strongest baseline changed. On all 14 entries that SCPEC leads, it wins every seed, and its worst run beats the baseline's best run. Ten of 16 entries reach Welch p < 0.05; ECW MSE (p = 0.218) and Switch MSE (p = 0.108) miss because the baseline has one high-error seed, and ECL270/Traffic852 MAE are borderline (p = 0.054/0.052). Across the eight datasets, an exact Wilcoxon signed-rank test gives p = 0.016 (MSE) and 0.023 (MAE). On NewAPP, SCPEC and FreqMoE are statistically tied (p = 1.000/0.082), and SCPEC beats the next-strongest Fremer on both metrics (p ≤ 0.035). In Table 2, every variant is worse than SCPEC by more than its own std, with p < 0.05 except w/ MPP on Switch (12 windows; p = 0.166/0.072) and on ECL270 MSE (p ≈ 0.05).

Reviewer_qG9zOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified workload forecaster that combines temporal, spectral, and optional static-context representations. A sparse router selects learnable temporal patterns, which are fused into Transformer representations before a direct multi-step readout. The paper motivates this design through historical-coverage and future-transfer diagnostics, and reports best or tied-best performance on 15 of 16 MSE/MAE entries across eight dataset settings.

Critical strengths

None

Critical weaknesses
  1. The novelty of the learnable pattern bank is not sufficiently established. For example, DPAD already learns temporal pattern banks with selective retrieval, while ProtoTS learns temporal prototypes selected using input representations.

  2. The motivating diagnosis is not directly connected to the forecasting gains. Appendix B.2 shows that historical composition already restores coverage to 97.8% on ECW and 97.0% on NewAPP, while the remaining ambiguity concerns direct future transfer rather than the baseline's learned decoder.

  3. The evidence for generalization across application events remains limited. NewAPP contains eleven nodes, while Switch has only twelve overlapping windows from three nodes, and every experiment uses the same 48-to-24 forecasting setting.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer qG9z: novelty vs. DPAD/ProtoTS, link between the diagnosis and the gains, and broader generalization evidence

We thank the reviewer for the careful reading. All new results are in the revised PDF (Secs. 1–3, 5–6 and App. B.2–B.6).

W1: Novelty relative to DPAD and ProtoTS. We agree that learnable prototype banks already exist; the revision cites both and states that we do not claim learnability per se (Secs. 1–2). What differs is how the learned patterns are selected and where they enter the forecast:

  • Selection. DPAD retrieves prototypes by Pearson correlation with the (projected) input, i.e., by waveform resemblance, which relies on assumption (A2) that Sec. 3.2 finds unreliable after application events. ProtoTS weights all prototypes by their distance to one pooled input embedding. SCPEC routes each local block sparsely (Top-K) from the encoded temporal–spectral–context workload state, with a router learned end to end with the forecaster.
  • Integration. DPAD concatenates retrieved prototypes with the final representation before the output layer; ProtoTS outputs a convex combination of prototype curves. SCPEC composes patterns inside every encoder stage through a gated residual and decodes with horizon queries, so its forecast is not confined to combinations of stored curves.
  • Setting. One bank is shared by heterogeneous nodes and must transfer zero-shot to post-event workloads, the regime that Sec. 3.2 characterizes.

We implemented both mechanisms inside SCPEC with the same state, encoder, hyperparameters and seeds (App. B.3, Table 7). MSE, mean of 3 seeds:

ECWNewAPPSwitch
SCPEC0.06210.02920.0542
DPAD-style selection (same bank)0.06580.03070.0633
ProtoTS-style prototype head0.06620.03240.0642

With the identical learnable bank, state routing lowers MSE by 5.6%/4.9%/14.4% (ECW/NewAPP/Switch), so learnability alone does not explain the gains; the gap to the ProtoTS-style head widens under application events (6.2% on ECW vs. 9.9%/15.6% on NewAPP/Switch). In all 12 comparisons (incl. MAE), the gap exceeds the sum of the two seed stds. A 2×2 ablation (Table 8) further shows that bank learning and state routing each reduce MSE when the other is fixed.

W2: Link between the diagnosis and the gains. We agree that composition restores coverage (Cov(0.95): 97.8% on ECW, 97.0% on NewAPP) and that PTA transfers the matched future directly. Composition does not restore future consistency, however: the composed PTA ratio on NewAPP (11.78%) is still 3.8× that of ECW (3.08%). We therefore tested a learned decoder directly (App. B.2, Table 6), splitting the 594 NewAPP windows by the diagnostic labels, which are computed from the data alone, independently of any model:

  • For MetaEformer, trained end to end on its history-derived pool, PTA windows have 1.49× the MSE (1.16× the MAE) of consistent windows, and windows still ambiguous after composition 1.42× (1.22×), whereas windows without a close single match have only 1.10× the MSE. The ambiguity is thus not specific to direct transfer, and coverage alone is only weakly related to error. This matches the reviewer's reading: composition largely closes (A1), while (A2) persists and is more strongly associated with error (now stated in Sec. 3.2).
  • SCPEC beats MetaEformer in every group, and its MSE reduction is largest on the 70 still-ambiguous windows (13.9% vs. 8.2% on the other 524), i.e., where (A2) persists. We read these group differences descriptively.

W3: Evidence for generalization. We broadened the evaluation in two directions.

  • Unseen nodes (App. B.5, Table 13). On ECW-New Infras (3,906 windows from 651 edge nodes absent from ECW; this adds nodes, not event types), ECW-trained checkpoints are applied without retraining (3 seeds). MSE/MAE: SCPEC 0.0587/0.1361, FreqMoE 0.0619/0.1379, MetaEformer 0.0651/0.1454, Dynamic-TMoE 0.0675/0.1421, DynEformer 0.0708/0.1594, seasonal naïve 0.0796/0.1458. SCPEC's MSE is below every learned baseline at every seed (5.2% below FreqMoE, 9.9% below MetaEformer).
  • Other lengths (App. B.6, Table 14). With L ∈ {32, 48, 96} (H = 24) and H ∈ {24, 48, 72} (L = 48), trained on ECW and applied to NewAPP without retraining, SCPEC is best on ECW in all five settings (MSE 4.2%–18.4% below the stronger of MetaEformer and FreqMoE; 8 of 10 MSE/MAE gaps p < 0.05) and beats MetaEformer in all 20 ECW/NewAPP entries (Wilcoxon p = 0.002). On NewAPP, FreqMoE is slightly better at 96→24, 48→48 and 48→72 (≤ 4.5% in MSE), never significantly (p ≥ 0.23).

Switch is not in Table 14 (its three 120-step series leave no window at 96→24 or 48→72), and with 12 windows we read it descriptively. Sec. 6 states that the evidence covers two event types and unseen nodes on one platform. We would be glad to address any further questions.

Reviewer_6K5JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SCPEC, a workload forecasting model that combines temporal, spectral, and static-context information with a learnable pattern-expert bank. Unlike existing approaches that retrieve patterns from historical observations, SCPEC learns reusable patterns and selects them based on the current workload state. Experiments on eight datasets show improvements over a wide range of forecasting baselines.

Critical strengths

The paper addresses an important problem in dynamic workload forecasting. The idea of replacing historical pattern retrieval with state-conditioned learnable experts is interesting and well motivated. The experimental results are strong across multiple datasets, and the ablation study provides useful evidence for the effectiveness of the main components. I also appreciate the comparison with MetaEformer using the same model backbone.

Critical weaknesses

My main concern is whether the method can generalize to unseen workload patterns. Although NewAPP and Switch are designed to evaluate changing application states, both are derived from the same ECW dataset. In particular, the Switch scenario contains only 12 valid evaluation windows, which seems insufficient to support strong conclusions about generalization.

The novelty of the proposed architecture is also somewhat limited. The method combines existing ideas, including frequency-domain representations, mixture-of-experts routing, and Transformer-based forecasting. Although the results are encouraging, it is unclear which specific design choices are most responsible for the improvements. For example, the MPP ablation replaces the expert bank, routing, and fusion mechanism together, making their individual contributions difficult to assess.

Finally, all experiments use the same 48-step input and 24-step prediction horizon. It would be useful to understand whether the improvements hold under different forecasting horizons. Some performance differences are also small, and the main results would benefit from reporting variability across runs.

Questions for authors

Can you provide additional evidence of generalization to unseen application states, ideally using independent workload traces?

Can you separate the contributions of the learnable pattern bank and state-conditioned routing through additional ablations?

How does SCPEC perform with different input lengths and prediction horizons?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 6K5J: generalization to unseen nodes, separating the bank from routing, and other input lengths and horizons

We thank the reviewer for the positive assessment and the concrete questions. All results below are in the revised PDF (App. B.3–B.6).

W1 / Q1: Generalization to unseen application states and independent traces. We added a further zero-shot target, ECW-New Infras (App. B.5, Table 13): 3,906 test windows from 651 edge nodes absent from ECW, evaluated with the ECW-trained checkpoints without retraining (3 seeds).

ECW-New InfrasMSEMAE
SCPEC0.0587±.00050.1361±.0023
FreqMoE0.0619±.00060.1379±.0021
MetaEformer0.0651±.00040.1454±.0003
Dynamic-TMoE0.0675±.00110.1421±.0021
DynEformer0.0708±.00770.1594±.0210
Seasonal naïve0.07960.1458

SCPEC's MSE is below every learned baseline at every seed (5.2% below FreqMoE, 9.9% below MetaEformer); its MAE gap to FreqMoE is within seed variation. On NewAPP, its zero-shot advantage over MetaEformer also holds at all five input/horizon settings (W3/Q3 below). We agree that independent traces would be stronger evidence: ECW-New Infras comes from the same platform and adds unseen nodes rather than new event types, and we read Switch (12 windows, 3 nodes) descriptively. Sec. 6 states this scope.

W2 / Q2: Separating the learnable bank from state-conditioned routing. We crossed the two factors in a 2×2 ablation (setting of Table 2; 3 seeds). The bank starts from randomly drawn block-aligned training segments (no warm-up) and is learned or kept fixed; experts are selected by input–pattern Pearson similarity or by state routing. MSE as ECW / NewAPP / Switch / ECL270:

BankSimilarityState routing
Fixed0.0682 / 0.0331 / 0.0612 / 0.04380.0662 / 0.0318 / 0.0593 / 0.0421
Learned0.0661 / 0.0315 / 0.0602 / 0.04190.0648 / 0.0313 / 0.0583 / 0.0401

Each factor reduces MSE on all four datasets when the other is fixed (bank learning 1.6%–4.8%, state routing 0.6%–4.3%; 3 of 16 gaps within seed variation), and both together by 4.7%–8.4%. SCPEC, which warm-starts its learnable bank, is lower still (0.0621 / 0.0292 / 0.0542 / 0.0382). A fixed-gate variant (the learned gate lowers MSE by 4.9%–9.4%) completes the separation of the three mechanisms that w/ MPP replaced at once (App. B.3, Table 8). Together with Table 2, this shows which design choices drive the gains: removing the pattern experts raises MSE by 28%–69% and removing the spectral view by 20%–50%, and within the knowledge module, bank learning, state routing and the learned gate each reduce MSE on all four datasets.

W3 / Q3: Input lengths and prediction horizons. MSE (3 seeds) as SCPEC / MetaEformer / FreqMoE, trained on ECW, NewAPP without retraining (Table 14 adds MAE and stds):

L→HECWNewAPP
32→240.0720 / 0.0882 / 0.08830.0301 / 0.0324 / 0.0315
48→240.0621 / 0.0680 / 0.08110.0292 / 0.0323 / 0.0292
96→240.0732 / 0.0832 / 0.07640.0292 / 0.0303 / 0.0289
48→480.0782 / 0.0882 / 0.08650.0392 / 0.0420 / 0.0375
48→720.0894 / 0.1048 / 0.09800.0452 / 0.0486 / 0.0439

SCPEC is best on ECW in every setting (vs. the stronger baseline, MSE 4.2%–18.4% and MAE 2.9%–12.1% lower; 8 of 10 gaps p < 0.05) and beats MetaEformer in all 20 ECW/NewAPP entries (Wilcoxon p = 0.002); FreqMoE's small NewAPP edge in three settings is never significant (p ≥ 0.23), consistent with NewAPP's stronger spectral concentration (App. B.1). Longer windows leave fewer test windows, so values compare within a setting only.

W3: Variability across runs. Table 2 now reports mean ± std, and App. B.4 gives stds, per-seed results and tests for Table 1. On all 14 entries SCPEC leads, it wins every seed; 10 of 16 entries reach Welch p < 0.05, and a Wilcoxon test across the eight datasets gives p = 0.016 (MSE) and 0.023 (MAE). We would be glad to clarify anything further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.