Pith. sign in

REVIEW 4 major objections 5 minor 28 references

DSTFView claims that jointly modeling closeness, period, and multi-view spatio-temporal-frequency relations delivers the best workload forecasts on Alibaba CPU and TP datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 12:39 UTC pith:2GFSHFDS

load-bearing objection New-looking assembly of standard modules; the claimed wins can't be checked until graph construction and variance are reported. the 4 major comments →

arxiv 2607.22565 v1 pith:2GFSHFDS submitted 2026-05-30 cs.AI cs.LG

DSTFView: Multi-View Cloud-Edge Workload Forecasting with Dual-Input Spatio-Temporal-Frequency Modeling

Qingzhong Li , Hui Ma , Yajun Zhang , Qingchang Ma , Zhou Long This is my paper
classification cs.AI cs.LG
keywords cloud-edge collaborationworkload forecastingspatio-temporal-frequency modelingadaptive fusiongraph neural networkmulti-step predictiontime series forecastingedge computing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes DSTFView, a collaborative cloud-edge forecasting framework, and claims it consistently beats representative baselines—including generic time-series models and dedicated cloud-edge methods such as ELASTIC and XELASTIC—on two Alibaba datasets across horizons 1, 4, 8, and 12. The central assertion is that no existing approach combines all three needed ingredients: two temporal inputs (closeness and period), spectral transformations, and an adaptive fusion of multiple relational graphs. If correct, this means cloud-edge workload forecasting can be improved by explicitly unifying spatial, temporal, and frequency modeling in both a cloud-side global stage and an edge-side local stage. The paper supports the claim with reported MAE, SMAPE, and R2 tables plus an ablation study showing the full multi-view AMVF fusion outperforms single-view variants.

Core claim

On the CPU and TP datasets, DSTFView reports the lowest MAE and SMAPE and the highest R2 at every horizon tested, beating the next best baseline XELASTIC by a margin that grows with the horizon (e.g., CPU H=1 MAE 0.147 vs 0.159; TP H=12 MAE 7.877 vs 8.374). The authors attribute this to jointly modeling closeness and period inputs, extracting frequency-domain features via FFT, and adaptively fusing spatial, temporal, and self-adaptive graph views at the global stage and physical, logical, and self-adaptive views at the local stage. Ablations indicate that the adaptive multi-view configuration outperforms any single-view setting.

What carries the argument

The load-bearing mechanism is the dual-input spatio-temporal-frequency block combined with Adaptive Multi-View Fusion (AMVF). Each branch (closeness and period) passes through FFT, gated temporal convolutions, and graph convolutions over multiple adjacency matrices; AMVF learns a softmax weight per view per step, and the global and local stages are linked by learnable aggregation/disaggregation layers and a per-site gating tensor lambda.

Load-bearing premise

The paper does not specify how the spatial, temporal, self-adaptive, physical, and logical graphs are constructed; the entire claimed advantage rests on these adjacency matrices encoding the real inter-site and inter-VM dependencies.

What would settle it

Replace each of the five adjacency matrices (spatial, temporal, self-adaptive, physical, logical) with identity matrices while keeping everything else identical; if MAE and SMAPE on the CPU and TP test sets change by less than a few percent, the multi-view graph fusion is not the source of the reported improvement.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If DSTFView's results hold, cloud-edge forecasting systems can adopt the two-stage aggregation/global-model/disaggregation pipeline to reduce uplink communication while improving multi-step accuracy.
  • The reported gains over XELASTIC at longer horizons suggest the frequency and multi-view fusion components specifically help with periodicity and bursty dynamics.
  • The success of adaptive fusion implies that static single-graph spatio-temporal models leave accuracy on the table for edge workloads.
  • The ablation results indicate that combining multiple relational views adds more value than any single predefined view, supporting research into richer graph definitions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves graph construction unspecified; a natural extension is to test learned or dynamic graphs, since the reported advantage may depend on the chosen adjacency definitions.
  • Because the architecture is task-agnostic, the same dual-input STF plus AMVF recipe could transfer to other multi-entity forecasting domains such as traffic or server performance, where relational views and periodicities coexist.
  • The per-entity gating tensor in the final fusion could be interpreted as a confidence signal useful for downstream auto-scaling decisions, though the paper does not explore this.
  • The visual claim of better peak and valley tracking could be turned into a quantitative test by measuring error at local maxima and minima across the test set.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DSTFView, a two-stage cloud–edge workload forecasting framework. It operates in a global cloud stage and a local edge stage, each using a dual-input (closeness/period) spatio-temporal-frequency architecture with graph convolution over multiple relational views and an adaptive multi-view fusion (AMVF) module to reweight those views. The claimed contribution is that jointly modeling temporal, frequency, and multi-graph spatial dependencies, together with the AMVF mechanism, yields consistent improvements over temporal, spatio-temporal, and collaborative cloud–edge baselines on the Alibaba CPU and TP datasets for horizons 1, 4, 8, and 12. The empirical section reports MAE, SMAPE, and R² for all models, followed by ablation studies for the global and local graph configurations.

Significance. If the reported results are substantiated, DSTFView would be a credible new state of the art on two real cloud–edge workload benchmarks, outperforming both general time-series transformers (PatchTST-class, iTransformer, TimesNet), graph-based traffic methods, and the dedicated ELASTIC/XELASTIC pipeline. The problem is practically relevant, the architecture is modular, and the direction of the reported gains is consistent across horizons and datasets. However, the paper currently lacks several load-bearing implementation details and has an internal inconsistency in the ablation tables, so the significance cannot be fully assessed from the manuscript as written. No code, configuration files, random seeds, or variance estimates are provided, which limits reproducibility and makes it impossible to distinguish the reported gains from evaluation noise.

major comments (4)
  1. [§3.3–§3.4] The construction of every graph adjacency used by the GCNs is left unspecified. The global stage uses a spatial graph A_g^(s), a temporal graph A_g^(t), and a self-adaptive graph A_g^(a); the local stage uses physical A_m^(p), logical A_m^(l), and self-adaptive A_m^(a). The paper never states how these graphs are built: which similarity metric, threshold or k-nearest-neighbor rule, whether adjacency is normalized, whether the temporal graph is derived from full-trace correlations or only the training split, and whether the period input is constructed purely from historical lags. This is not a cosmetic omission: if the temporal graph or period sequence uses the test portion or target variable, the GCN could leak future information, and the reported advantage would not reflect the proposed STF/AMVF mechanism. Please provide precise graph-construction formulas and normalization procedures,
  2. [Table 3 vs. Table 1] There is a large unexplained scale mismatch between the global-stage ablation and the main results. In Table 1, the full DSTFView on CPU at H=1 has MAE 0.147; in Table 3, the same configuration labeled 'tim-spa-adaptive (AMVF)' reports MAE 2.348 at H=1 on the same dataset. The TP values also differ by orders of magnitude (Table 3 H=1 MAE 214.357 vs. Table 2 3.496). If the global stage predicts differently normalized or differently aggregated site-level signals, the ablation does not support the stated claims about the contribution of each graph view to the final VM-level metric. The paper must explain the units, normalization, and prediction target used in Table 3, or report ablation results on the same metric and scale as Tables 1–2.
  3. [§4.2 and §5.1] No variance information is reported for any experiment: no random seeds, error bars, confidence intervals, or significance tests. Several headline margins are small relative to what one would expect from seed-to-seed variation, e.g., CPU H=1 MAE 0.147 vs. XELASTIC 0.159 and TP H=12 MAE 7.877 vs. 8.374. The claim of 'consistently outperforms' in the abstract and §5.1 therefore rests on point estimates without a measure of uncertainty. Please report means and standard deviations over at least three independent runs, and, ideally, a paired significance test (e.g., Wilcoxon or paired t-test) over the test windows.
  4. [Eqs. (2)–(3) and (9)–(10)] The frequency-domain branch applies FFT to obtain F_g^r / F_m^r and then feeds this representation directly into real-valued temporal convolution layers (TCN_a and TCN_b). The FFT output is complex-valued, but the manuscript does not specify whether the real part, imaginary part, magnitude, or a concatenation is used, nor how complex numbers are handled by the subsequent GCN and prediction head. This is a concrete architectural gap that prevents replication and affects the interpretation of the frequency-aware component. Please clarify the exact complex-to-real transformation.
minor comments (5)
  1. [Tables 1 and 2] The column header 'SMAPER 2' appears to be a typo for 'SMAPE' and should be corrected throughout.
  2. [§4.2] The hyperparameter settings for the baselines are not described. It is unclear whether each baseline was tuned on the validation set, what input lengths were used, or whether all models received the same closeness/period inputs. This is important for fair comparison.
  3. [Fig. 2] The figure caption does not identify the units for CPU utilization and TP throughput, and the subfigures lack axis labels. Please add them.
  4. [§1 and §5.1] The paper repeatedly claims 'efficiency' and 'communication overhead' advantages, but no runtime, parameter count, or communication cost is measured. Either add such measurements or soften the efficiency claims.
  5. [§3.2] The notation list omits an explicit definition of the prediction target with horizon H, although it is used in Eq. (1). Please define X_{t:t+H-1}^m clearly.

Circularity Check

0 steps flagged

No significant circularity: DSTFView is an empirical benchmark paper whose reported gains are tested on a held-out split; under-specified graph construction and ablation scale mismatches are validity concerns, not circular reductions.

full rationale

The paper's derivation chain is architectural rather than deductive. The model is defined by equations (1)-(17), and the central claim is the empirical result in Tables 1-2 that DSTFView "consistently outperforms representative baselines." The target variable appears only as the output of the forecasting function F in Eq. (1), never as an input used to define the model. The model is trained with MSE on a training split and evaluated on a chronologically held-out test split (7:2:1, §4.2), so the reported numbers are not fitted parameters renamed as predictions. The learned fusion weights α in Eqs. (5)-(6) and (12)-(13) are optimized on training data; they do not encode the test outcome or the evaluation metric. There is no equation that reduces to a fitted constant, no target quantity defined in terms of the model output by construction, and no load-bearing self-citation: ELASTIC and XELASTIC are external baselines, not premises used to derive DSTFView's mechanism. Two concerns are real but are not circularity. First, §3.3-§3.4 state that spatial/temporal/self-adaptive and physical/logical/self-adaptive graphs are constructed but never specify how; if those graphs were built using future or target information, leakage could inflate the reported gains, but the paper does not say this, and absent such specification one cannot claim an equation-level reduction. Second, the global-stage ablation in Table 3 reports CPU H=1 MAE around 2.35-2.95 while the full model in Table 1 reports CPU H=1 MAE 0.147; this scale mismatch undermines the ablations' support for component-wise conclusions, but it is an internal-consistency/reproducibility issue, not a circular derivation. Overall, the empirical claim is self-contained and externally testable against the cited datasets and baselines; no circular step is exhibited.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim is empirical, so the ledger focuses on learned/hand-chosen parameters and unstated domain assumptions that the architecture's success depends on. The main unprotected choices are the unstated graph-construction rules, the undefined FFT-to-TCN interface, and the unspecified window lengths.

free parameters (5)
  • Neural network weights (TCNs, GCNs, FFT feature projections, AMVF scoring, aggregation/disaggregation, gating lambda) = not reported
    All are learned via MSE on training data; the central benchmark claim depends on these fits, but no initialization or weight details are given.
  • Closeness/period window lengths P_c and P_p = not reported
    These dual-input lengths are never specified; forecasting accuracy can depend strongly on the choice.
  • Graph construction hyperparameters (thresholds, neighbor counts, similarity metrics) = not reported
    Spatial/temporal/physical/logical graphs are introduced without construction details; the adjacency choice determines the inductive bias.
  • Model depth/width (hidden dims, TCN layers, GCN layers) = not reported
    No architecture dimensions are provided, so the model cannot be re-implemented at the same capacity.
  • Training hyperparameters (learning rate, batch size, epochs, scheduler) = lr=0.001, batch=64, epochs=50, ExponentialLR
    These are fixed without sensitivity analysis, and no tuning details are given for baselines.
axioms (4)
  • ad hoc to paper FFT output can be fed directly into real-valued TCN/GCN layers without an inverse transform.
    Eq. (2) applies FFT and then TCNs to F(Y); complex-valued FFT outputs are not discussed, leaving the actual feature representation undefined.
  • domain assumption The pre-defined and self-adaptive graphs remain valid across the entire train/test period.
    Sections 3.3 and 3.4 define static graph sets even though the paper itself notes that edge relations are time-varying due to mobility and dynamic deployment.
  • domain assumption Z-score standardization and the chronological 7:2:1 split do not leak test-set statistics into training.
    Section 4.2 states standardization and split but does not say whether normalization statistics are computed only on the training portion.
  • domain assumption Baselines are fairly configured and their reported numbers are comparable to DSTFView's.
    No baseline hyperparameters or code are given, so claims of consistent superiority rest on assumed fair comparison.

pith-pipeline@v1.3.0-alltime-deepseek · 10054 in / 12343 out tokens · 124768 ms · 2026-08-02T12:39:56.619748+00:00 · methodology

0 comments
read the original abstract

With the widespread deployment of edge-side AI inference, edge platforms are increasingly required to support latency-sensitive, highly concurrent, and reliability-critical applications. However, existing methods often struggle to balance multidimensional feature modeling and forecasting efficiency in collaborative cloud-edge environments. To address this issue, we propose DSTFView, a dual-input spatio-temporal-frequency multi-view workload forecasting framework for collaborative cloud-edge environments. It jointly models closeness and period dependencies and extracts spatial, temporal, and frequency-domain dependencies. Besides, it designs an adaptive fusion mechanism and adjusts the contribution of each view to capture abrupt changes. Experimental results on the CPU and TP datasets demonstrate that DSTFView consistently outperforms representative baselines across multiple forecasting horizons and evaluation metrics.

Figures

Figures reproduced from arXiv: 2607.22565 by Hui Ma, Qingchang Ma, Qingzhong Li, Yajun Zhang, Zhou Long.

Figure 1
Figure 1. Figure 1: Structure of DSTFView model. 3 Method 3.1 Framework Overview As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization results on the CPU and TP dataset. 5.2 Ablation Study In this section, we conduct an ablation study to evaluate the contributions of its key components at the Global Stage and the Local Stage. For the global stage, we examine single-view variants, a dual-view variant (spatial-temporal), and the full model integrating temporal, adaptive, and spatial views. As reported in [PITH_FULL_IMAGE:figu… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 2 linked inside Pith

  1. [1]

    In: International Conference on Learning Representations (2024)

    Chen, P., Zhang, Y., Cheng, Y., Shu, Y., Wang, Y., Wen, Q., Yang, B., Guo, C.: Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting. In: International Conference on Learning Representations (2024)

    2024

  2. [2]

    In: International Conference on Learning Representations (2022)

    Fan, W., Zheng, S., Yi, X., Cao, W., Fu, Y., Bian, J., Liu, T.Y.: DEPTS: Deep expansion learning for periodic time series forecasting. In: International Conference on Learning Representations (2022)

    2022

  3. [3]

    Computing107(5), 125 (2025)

    Ghahari-Bidgoli, M., Ghobaei-Arani, M., Sharif, A.: An efficient task offloading and auto-scaling approach for IoT applications in edge computing environment. Computing107(5), 125 (2025)

    2025

  4. [4]

    IEEE Intelligent Transportation Systems Magazine14(1), 153–162 (2022)

    Guo, K., Hu, Y., Qian, Z.S., Sun, Y., Gao, J., Yin, B.: An optimized temporal- spatial gated graph convolution network for traffic forecasting. IEEE Intelligent Transportation Systems Magazine14(1), 153–162 (2022)

    2022

  5. [5]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Guo, K., Hu, Y., Sun, Y., Qian, S., Gao, J.: Hierarchical graph convolution net- work for traffic forecasting. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 151–159 (2021)

    2021

  6. [6]

    IEEE Transactions on Intelligent Transportation Systems22(2), 1138– 1149 (2021)

    Guo, K., et al.: Optimized graph convolution recurrent neural network for traffic prediction. IEEE Transactions on Intelligent Transportation Systems22(2), 1138– 1149 (2021)

    2021

  7. [7]

    In: Proceedings of the AAAI Con- ference on Artificial Intelligence

    Guo, S., Lin, Y., Feng, N., Song, C.: Attention based spatial-temporal graph con- volutional networks for traffic flow forecasting. In: Proceedings of the AAAI Con- ference on Artificial Intelligence. pp. 922–929 (2019)

    2019

  8. [8]

    In: Pro- ceedings of the AAAI Conference on Artificial Intelligence (2023)

    Jiang, R., Wang, X., Jie, Y., Chen, C., Kobayashi, K., Song, X., Fukushima, S., Suzumura, T.: Spatio-temporal meta-graph learning for traffic forecasting. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence (2023)

    2023

  9. [9]

    International Journal of Machine Learning and Cybernetics16, 5079–5112 (2025)

    Kong, X., Chen, Z., Liu, W., Ning, K., Zhang, L., Marier, S.M., Liu, Y., Chen, Y., Xia, F.: Deep learning for time series forecasting: a survey. International Journal of Machine Learning and Cybernetics16, 5079–5112 (2025)

    2025

  10. [10]

    Internet of Things31, 101553 (2025)

    Kreković, D., Krivić, P., Žarko, I.P., Kušek, M., Le-Phuoc, D.: Reducing communication overhead in the iot-edge-cloud continuum: A survey on pro- tocols and data reduction strategies. Internet of Things31, 101553 (2025). https://doi.org/10.1016/j.iot.2025.101553

    arXiv 2025

  11. [11]

    In: Proceedings of the ACM Web Conference

    Li, Y., Yuan, H., Fu, Z., Xu, M., Wang, S.: ELASTIC: Edge workload forecasting based on collaborative cloud-edge deep learning. In: Proceedings of the ACM Web Conference. pp. 3056–3066 (2023)

    2023

  12. [12]

    IEEE Transactions on Mobile Computing24(4), 2861–2875 (2025) 12 Q

    Li, Y., Zhao, P., Ma, X., Yuan, H., Fu, Z., Xu, M., Wang, S.: A collaborative cloud-edge approach for robust edge workload forecasting. IEEE Transactions on Mobile Computing24(4), 2861–2875 (2025) 12 Q. Li, H. Ma, and Y. Zhang, et al

    2025

  13. [13]

    arXiv preprint arXiv:2310.06625 (2023)

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., Long, M.: iTransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625 (2023)

    Pith/arXiv arXiv 2023

  14. [14]

    In: 5G Edge Computing: Technologies, Applications and Future Visions, pp

    Ma, X., Xu, M., Li, Q., Li, Y., Zhou, A., Wang, S.: Edge workload prediction based on deep learning. In: 5G Edge Computing: Technologies, Applications and Future Visions, pp. 45–61. Springer, Singapore (2024)

    2024

  15. [15]

    Nguyen, N., Sinthong, P., Kalagnanam, J.: A time series is worth 64 words: Long-term forecasting with transformers

    Nie, Y., H. Nguyen, N., Sinthong, P., Kalagnanam, J.: A time series is worth 64 words: Long-term forecasting with transformers. In: International Conference on Learning Representations (2023)

    2023

  16. [16]

    ACM Computing Surveys56(1), 1–31 (2023)

    Ning, Z., Hu, H., Wang, X., Guo, L., Guo, S., Wang, G., Gao, X.: Mobile edge computing and machine learning in the Internet of unmanned aerial vehicles: A survey. ACM Computing Surveys56(1), 1–31 (2023)

    2023

  17. [17]

    In: International Conference on Neural Information Processing

    Seo, Y., Defferrard, M., Vandergheynst, P., Bresson, X.: Structured sequence mod- eling with graph convolutional recurrent networks. In: International Conference on Neural Information Processing. pp. 362–373. Springer (2018)

    2018

  18. [18]

    In: Advances in Neural Information Processing Systems

    Wang, Y., Wu, H., Dong, J., Qin, G., Zhang, H., Liu, Y., Qiu, Y., Wang, J., Long, M.: TimeXer: Empowering transformers for time series forecasting with exogenous variables. In: Advances in Neural Information Processing Systems. vol. 37, pp. 469–498 (2024)

    2024

  19. [19]

    The Journal of Su- percomputing81(896) (2025)

    Wang, Z., Hu, R., Zhong, B., Cheng, D., Feng, J., Tan, S.: St-graphormer: spatio- temporal graph transformer for end-to-end traffic forecasting. The Journal of Su- percomputing81(896) (2025)

    2025

  20. [20]

    In: International Conference on Learning Representations (2023)

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., Long, M.: TimesNet: Temporal 2D- Variation modeling for general time series analysis. In: International Conference on Learning Representations (2023)

    2023

  21. [21]

    In: Advances in Neural Informa- tion Processing Systems

    Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Advances in Neural Informa- tion Processing Systems. pp. 22419–22430 (2021)

    2021

  22. [22]

    In: International Conference on Learning Representations (2023)

    Zhang, Y., Yan, J.: Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In: International Conference on Learning Representations (2023)

    2023

  23. [23]

    IEEE Transac- tions on Services Computing (2024)

    Zhao, F., Lin, W., Lin, S., Zhong, H., Li, K.: TFEGRU: Time-frequency enhanced gated recurrent unit with attention for cloud workload prediction. IEEE Transac- tions on Services Computing (2024)

    2024

  24. [24]

    In: International Joint Conference on Artificial Intelligence (2025)

    Zheng, J., Xie, L.: A dynamic stiefel graph neural network for efficient spatio- temporal time series forecasting. In: International Joint Conference on Artificial Intelligence (2025)

    2025

  25. [25]

    In: Proceed- ings of the AAAI Conference on Artificial Intelligence

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: Informer: Beyond efficient transformer for long sequence time-series forecasting. In: Proceed- ings of the AAAI Conference on Artificial Intelligence. pp. 11106–11115 (2021)

    2021

  26. [26]

    arXiv preprint arXiv:2408.10822 (2024)

    Zhou, J., Liu, E., Chen, W., Zhong, S., Liang, Y.: Navigating spatio-temporal heterogeneity: A graph transformer approach for traffic forecasting. arXiv preprint arXiv:2408.10822 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    In: Proceedings of the 39th International Conference on Machine Learning

    Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., Jin, R.: FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. In: Proceedings of the 39th International Conference on Machine Learning. vol. 162, pp. 27268– 27286 (2022)

    2022

  28. [28]

    Zhu, G., Lyu, Z., Jiao, X., et al.: Pushing AI to wireless network edge: An overview onintegratedsensing,communication,andcomputationtowards6G.ScienceChina Information Sciences66(3), 130301 (2023)

    2023