Pith. sign in

REVIEW 4 major objections 5 minor 17 references

TFGformer claims that multivariate time series forecasting improves when inter-variable dependencies are learned as a sparse graph in the time-frequency domain and fused with historical and future covariates, reporting state-of-the-art erro

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

2026-08-03 06:39 UTC pith:KIZWIJUO

load-bearing objection A promising architecture buried under a mismatched abstract and an unverifiable covariate setup — not ready for acceptance, but not a waste of referee time. the 4 major comments →

arxiv 2607.29459 v1 pith:KIZWIJUO submitted 2026-07-31 cs.LG cs.AI

TFGformer: Multivariate Time Series Forecasting via Time-Frequency Graph Learning and Covariate Fusion

Yu Sun , Yuan Chang , Xiaohou Shi , Yan Sun This is my paper
classification cs.LG cs.AI
keywords multivariate time series forecastingtime-frequency graph learningShort-Time Fourier TransformMahalanobis distanceGumbel-Softmaxcovariate fusionattention maskTransformer
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 argues that multivariate forecasting improves when inter-variable dependencies are first learned in the joint time-frequency domain and converted into a sparse binary attention mask, and when historical and future covariates are fused through MLP residual modules. On seven standard benchmarks, the resulting TFGformer claims the best average MSE and MAE on six datasets, with a 3.6% average MSE reduction over iTransformer, 6.4% over PatchTST, 5.2% over TiDE, and 16.4% over FEDformer. If correct, this points to two design lessons: frequency-domain structure learning can filter noisy channel correlations better than full channel mixing, and explicit covariate fusion is a measurable source of accuracy. A sympathetic reader would care because the components are modular and the claimed gains are consistent across prediction horizons.

Core claim

TFGformer's central claim is that a learned, sparse, time-frequency graph mask — built from STFT magnitude spectrograms, a learnable Mahalanobis distance, and Gumbel-Softmax sampling — improves long-term multivariate forecasting by suppressing irrelevant variable connections, while an MLP-based covariate fusion module injects historical and future context at input and output stages. The paper reports state-of-the-art results on six of seven benchmarks and attributes the gains to these two modules, with ablations showing 6.0% average MSE degradation when the graph module is removed and 3.0% when covariate fusion is removed.

What carries the argument

The Time-Frequency Graph (TFG) module: each variable's sequence is transformed by STFT into a magnitude spectrogram, flattened into a feature vector; pairwise Mahalanobis distances yield a similarity matrix that is max-normalized into probabilities; Gumbel-Softmax turns those probabilities into a binary adjacency graph; a Kronecker product with a causal time mask produces the attention mask that zeroes out weak inter-variable paths. The Covariate Fusion Module has two MLP-based residual branches — I-CFM for historical covariates at the input and O-CFM for future covariates at the output — both concatenating projections and adding back to the main representation.

Load-bearing premise

The reported gains over baselines assume the baselines were not given the same future covariates that TFGformer fuses; if they were, the benchmark comparison would be unfair and the central SOTA claim could collapse.

What would settle it

Run iTransformer, PatchTST, and TiDE on the same seven benchmarks while supplying them exactly the future covariates TFGformer receives via O-CFM; if the reported MSE gap disappears, the gain is covariate access, not graph learning. Alternatively, randomize the learned graph while preserving edge density: if accuracy holds, the STFT-driven structure is not the mechanism.

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

If this is right

  • Sparse, learned time-frequency graphs for inter-variable attention could become a standard component in transformer-based forecasters, replacing dense channel mixing.
  • Explicit future-covariate fusion via residual MLP branches appears to be worth a few percent MSE across datasets, so benchmark protocol should report covariate usage precisely.
  • The learned graph provides a per-dataset heatmap of variable dependencies, giving an interpretability tool for electricity, traffic, and weather networks.
  • The same mask-and-fusion design may transfer to anomaly detection and imputation, which the paper names as future directions.

Where Pith is reading between the lines

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

  • The paper never specifies which covariates are used in the seven benchmarks or whether baselines received the same future covariates; if they did not, the reported gains may reflect an information advantage rather than the architecture.
  • The manuscript's abstract describes a different method (CrossRAG) than the TFGformer presented in the body, suggesting a misassembled submission; readers should verify the intended contribution before relying on the claims.
  • Because the Mahalanobis distance is implemented with a learnable diagonal weight vector, it reduces to a weighted Euclidean distance, so the 'covariance structure' claim is weaker than stated.
  • A direct ablation that randomizes the learned graph while preserving edge density would isolate whether the STFT-derived structure or merely the sparsity is what helps.

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 TFGformer, a Transformer-based multivariate time series forecasting model. The architecture has three main components: (i) a Time-Frequency Graph (TFG) module that uses STFT to build per-variable time-frequency representations, learns pairwise distances with a Mahalanobis metric, converts distances to probabilities, and applies Gumbel-Softmax to produce a sparse binary graph used as an attention mask; (ii) an MLP-based Covariate Fusion Module with an input-side fusion (I-CFM) for historical covariates and an output-side fusion (O-CFM) for future covariates; and (iii) a Transformer backbone that applies the TFG mask to channel-mixing attention. The authors report multivariate long-term forecasting results on ETT (ETTh1/2, ETTm1/2), ECL, Traffic, and Weather for horizons 96/192/336/720, and claim state-of-the-art average MSE/MAE on 6 of 7 datasets, with ablations showing both TFG and CFM contribute and a sensitivity analysis of the Gumbel-Softmax temperature.

Significance. If the results are reproducible and the comparison is fair, the TFG module's data-driven sparse channel mask—derived from joint time-frequency features—is a plausible and interesting mechanism for suppressing noisy inter-variable dependencies in channel-mixing Transformers. The covariate fusion design also addresses a real gap in many existing multivariate forecasting models, which either ignore covariates or concatenate them naively. However, the current significance is undermined by missing experimental controls: the paper does not specify which covariates are used per dataset, whether baselines received the same covariates, or whether any target-channel information leaks through C_future. The reported gains over the strongest baselines are often small (e.g., average MSE 0.171 vs. 0.178 on ECL, 0.422 vs. 0.428 on Traffic), and without error bars, multiple seeds, significance tests, or code, the SOTA claim is not established. The contribution is therefore promising but not yet sufficiently supported for a top-tier venue.

major comments (4)
  1. [Section III-E and Section IV-A] The experimental protocol never specifies which historical and future covariates (C_history, C_future) are used for each of the seven benchmarks, nor whether the baselines (iTransformer, TiDE, PatchTST, Autoformer, FEDformer, DLinear) were given the same auxiliary information. Since O-CFM (Eq. 13) consumes C_future at inference, the reported gains in Table I could be due to an information asymmetry rather than the proposed architecture. Standard ECL, Traffic, and Weather benchmarks do not have obvious exogenous future covariates when predicting all channels; if C_future includes future values of any target channel, the model would be leaking the prediction targets. The ablation in Section IV-C removes modules internally but does not test whether baselines were granted the same covariate access, so it cannot rule out this explanation. This is load-bearing for the central SOTA claim and mu
  2. [Section III-D, Eq. (6) and complexity paragraph] There is a direct inconsistency in the method description. Eq. (6) defines D_ij = (Z_i - Z_j)^T Q (Z_i - Z_j) with Q ∈ R^{N_feat × N_feat} a learnable symmetric PSD matrix, which is a full Mahalanobis distance. However, the complexity paragraph near Eq. (11) states that the implementation uses a 'learnable diagonal weight vector optimized via tensor broadcasting' and that this reduces complexity to O(C^2 · L). A full PSD matrix is not a diagonal weight vector. The paper must state which formulation is actually implemented. If diagonal, Eq. (6) reduces to a weighted Euclidean distance and the 'Mahalanobis' and covariance-structure claims are overstated; if full, the complexity claim is unsupported. This ambiguity affects the technical novelty of the TFG module.
  3. [Section IV-B and Table I] The empirical claim of consistent SOTA performance is not statistically supported. No error bars, number of random seeds, or significance tests are reported, and many differences against strong baselines are small (e.g., ETTh1 average MSE 0.442 vs. 0.440 for FEDformer; ETTh2 0.375 vs. 0.383; ECL 0.171 vs. 0.178; Traffic 0.422 vs. 0.428). Without variance estimates or code release, these gains could be noise. Given that the paper is first and foremost an empirical benchmarking paper, this omission is a major weakness.
  4. [Abstract (top of manuscript)] The abstract describes CrossRAG, a retrieval-augmented forecasting framework with Shape-Aware Memory, Future-Consistent Contrastive learning, and Cross-Attention Temporal Fusion. None of these components appear in the TFGformer methodology, experiments, or conclusion. The manuscript is internally inconsistent at its very front matter: the reader cannot tell which method is actually being proposed. This must be corrected before any further review.
minor comments (5)
  1. [Eq. (14)] The text says the attention mask is applied 'additively,' but the formula uses elementwise multiplication with M and adds -infinity for (1-M). Please align the wording and the math.
  2. [Eq. (8)] After max-normalization, the formula adds the identity matrix I, which sets the diagonal to 1. This appears to contradict Eq. (7), where diagonal elements are masked to 0. Clarify whether self-attention is allowed or not, and adjust the notation accordingly.
  3. [References [8]-[9]] In Section II-A, 'Informer and Autoformer [8]' is cited, but reference [8] is ChronosX, not Informer or Autoformer. Please fix the citation numbering.
  4. [Figure 2 (ablation)] The ablation results are only presented graphically. Include a small table with the exact average MSE/MAE values for Full Model, W/o TFG, and W/o CFM on Weather and ECL, so the claimed 3.0% and 6.0% degradations are verifiable.
  5. [Section III-D, STFT complexity] The stated STFT complexity O(C·L log L) ignores the dependence on nfft, hop, and window length. Please specify the complexity in terms of the actual STFT parameters used in the experiments.

Circularity Check

0 steps flagged

No circularity found: the empirical SOTA claim is checked against held-out external baselines; the main caveats are unspecified covariates and an abstract/body mismatch, neither of which is a by-construction reduction.

full rationale

The paper is an empirical benchmarking paper rather than a derivation. The forecasting head (Eq. 3) maps transformer outputs to the prediction; the TFG mask (Eq. 11) is a learned function of the input X via STFT, Mahalanobis distance, and Gumbel-Softmax; the CFM modules (Eqs. 12-13) fuse declared covariates. None of these steps defines the target Y in terms of the fitted parameters or uses the target to construct the graph, so there is no self-definitional or fitted-input-called-prediction loop. The temperature tau is tuned on validation, which is standard practice. There are no load-bearing self-citations or imported uniqueness theorems. The central SOTA claim is evaluated against external baselines on standard train/validation/test splits, so it is not equivalent to its inputs by construction. Two concerns are noted but they are not circularity: (1) the paper never specifies what C_his and C_fut are for each of the seven benchmarks nor whether the baselines were given the same covariates, which is a reproducibility/fairness threat to the SOTA comparison; and (2) the Abstract describes a different model, CrossRAG, not TFGformer, which is an internal consistency defect. Both are correctness-validity issues, not reductions of predictions to model inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central performance claim rests on a set of unstated modeling and comparison choices. The free parameters listed are all tuned or unreported and affect the graph sparsity, temporal features, and model capacity. The axioms are domain assumptions about the sufficiency of STFT magnitude features, the validity of the sampling procedure, and the availability and fairness of covariates.

free parameters (6)
  • Gumbel-Softmax temperature tau = 0.5 (best on Weather); tuned per dataset
    Controls graph sparsity; Table II shows sensitivity and says it is a tunable hyperparameter.
  • STFT parameters (nfft, hop, window length) = not reported
    Eq. 4 defines the spectrogram but these values determine F and T; they are never specified in the experiments.
  • learning rate = tuned from {1e-4, 5e-4, 1e-3} per dataset
    Section IV-A states learning rate is tuned from this set, which is a per-dataset free choice.
  • patch length and stride = not reported
    Eq. 2 defines N=(L-l)/s+1 but l and s are never given, affecting the sequence representation.
  • Transformer hyperparameters (layers, heads, d_model) = not reported
    The backbone dimensions are not specified; these affect capacity and results.
  • epsilon in Eqs. 7-8 = not reported
    The small constant prevents division by zero in the distance-to-probability conversion and affects graph density.
axioms (5)
  • domain assumption STFT magnitude spectrograms carry discriminative information about inter-variable dependencies
    The TFG module builds the graph solely from magnitude spectra (Eq. 4-5); no evidence is given that this representation is sufficient.
  • domain assumption The Bernoulli plus Gumbel-Softmax procedure produces a differentiable, sparse, and valid attention mask
    Eq. 9-10 are underspecified; applying GumbelSoftmax to a Bernoulli sample is not a standard reparameterization, so the validity of the discrete mask is assumed.
  • domain assumption Future covariates are known at forecast time and can be imputed by forward-filling
    Section III-E assumes future covariates exist and are imputed; the paper does not state which covariates are used for each benchmark.
  • standard math The Kronecker-product attention mask preserves both causality and channel selectivity
    Eq. 11 is mathematically standard, but its usefulness depends on the learned binary graph G being correct.
  • domain assumption The fixed benchmark splits are free of leakage and distribution shift that would invalidate the comparison
    Section IV-A uses standard ratios (6:2:2 or 7:1:2) with no discussion of leakage or non-stationarity handling beyond instance normalization.

pith-pipeline@v1.3.0-daily-deepseek · 8254 in / 13565 out tokens · 140052 ms · 2026-08-03T06:39:59.454744+00:00 · methodology

0 comments
read the original abstract

Large-scale multivariate time series from heterogeneous IoT sensors demand accurate long-term forecasting for resource scheduling and predictive maintenance. While recent time series foundation models exhibit strong generalization, they rely on static parametric knowledge and lack dynamic access to external historical patterns during inference. Retrieval-Augmented Generation (RAG) offers a potential remedy, yet its application to time series forecasting is challenged by magnitude variations across heterogeneous sources and the mismatch between historical similarity and future consistency. We propose CrossRAG, a retrieval-augmented forecasting framework that integrates Shape-Aware Memory (SAM) with RevIN normalization for magnitude-robust shape-level retrieval, Future-Consistent Contrastive (FCC) learning to distinguish informative references from hard negatives with similar history but divergent futures, and Cross-Attention Temporal Fusion (CATF) to fuse retrieved historical--future reference pairs into the backbone's representations at the representation level. Experiments on seven public benchmarks show that CrossRAG consistently outperforms both parametric-only baselines and existing retrieval-augmented forecasting methods.

Figures

Figures reproduced from arXiv: 2607.29459 by Xiaohou Shi, Yan Sun, Yuan Chang, Yu Sun.

Figure 1
Figure 1. Figure 1: The overall architecture of TFGformer. C. Basic Components • Normalization Per-channel Z-score [14] normalization is applied along the temporal dimension: xˆi,t = xi,t − µi σi (1) • Patching The input sequence is split into fixed-length temporal patches using a sliding window: N = L − l s + 1 (2) • Output Head A linear layer maps the Transformer outputs to the prediction sequence: Yˆ = HW + b (3) D. Time–F… view at source ↗
Figure 2
Figure 2. Figure 2: The ablation study results of TFGformer [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Heatmap of learned variable correlations on the ETTh1 dataset [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] 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

17 extracted references · 8 linked inside Pith

  1. [1]

    Attention is all you need[J]

    Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30

    2017

  2. [2]

    An image is worth 16x16 words: Transformers for image recognition at scale[J]

    Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  3. [3]

    Recurrent neural net- works for time series forecasting: Current status and future directions,

    H. Hewamalage, C. Bergmeir, and K. Bandara, “Recurrent neural net- works for time series forecasting: Current status and future directions,” International Journal of Forecasting, vol. 37, no. 1, pp. 388–427, 2021

    2021

  4. [4]

    A Time Series is Worth 64Words: Long-term Forecasting with Transformers[J]

    Nie Y . A Time Series is Worth 64Words: Long-term Forecasting with Transformers[J]. arXiv preprint arXiv:2211.14730, 2022

    Pith/arXiv arXiv 2022

  5. [5]

    itransformer: Inverted transformers are effective for time series forecasting[J]

    Liu Y , Hu T, Zhang H, et al. itransformer: Inverted transformers are effective for time series forecasting[J]. arXiv preprint arXiv:2310.06625, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Timer-xl: Long-context transformers for unified time series forecasting[J]

    Liu Y , Qin G, Huang X, et al. Timer-xl: Long-context transformers for unified time series forecasting[J]. arXiv preprint arXiv:2410.04803, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Long-term forecasting with tide: Time- series dense encoder[J]

    Das A, Kong W, Leach A, et al. Long-term forecasting with tide: Time- series dense encoder[J]. arXiv preprint arXiv:2304.08424, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    ChronosX: Adapting pre- trained time series models with exogenous variables[J]

    Arango S P, Mercado P, Kapoor S, et al. ChronosX: Adapting pre- trained time series models with exogenous variables[J]. arXiv preprint arXiv:2503.12107, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Informer: Beyond efficient transformer for long sequence time-series forecasting[C]//Proceedings of the AAAI conference on artificial intelligence

    Zhou H, Zhang S, Peng J, et al. Informer: Beyond efficient transformer for long sequence time-series forecasting[C]//Proceedings of the AAAI conference on artificial intelligence. 2021, 35(12): 11106-11115

    2021

  10. [10]

    Are transformers effective for time series forecasting?[C]//Proceedings of the AAAI conference on artificial intelligence

    Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting?[C]//Proceedings of the AAAI conference on artificial intelligence. 2023, 37(9): 11121-11128

    2023

  11. [11]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting[C]//The eleventh international conference on learning representations

    Zhang Y , Yan J. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting[C]//The eleventh international conference on learning representations. 2023

    2023

  12. [12]

    Fedformer: Frequency enhanced de- composed transformer for long-term series forecasting[C]//International conference on machine learning

    Zhou T, Ma Z, Wen Q, et al. Fedformer: Frequency enhanced de- composed transformer for long-term series forecasting[C]//International conference on machine learning. PMLR, 2022: 27268-27286

    2022

  13. [13]

    Timesnet: Temporal 2d-variation modeling for general time series analysis[J]

    Wu H, Hu T, Liu Y , et al. Timesnet: Temporal 2d-variation modeling for general time series analysis[J]. arXiv preprint arXiv:2210.02186, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Reversible instance normalization for ac- curate time-series forecasting against distribution shift[C]//International conference on learning representations

    Kim T, Kim J, Tae Y , et al. Reversible instance normalization for ac- curate time-series forecasting against distribution shift[C]//International conference on learning representations. 2021

    2021

  15. [15]

    Duet: Dual clustering enhanced multivariate time series forecasting[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Qiu X, Wu X, Lin Y , et al. Duet: Dual clustering enhanced multivariate time series forecasting[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1. 2025: 1185- 1196

    2025

  16. [16]

    Categorical reparameterization with gumbel- softmax[J]

    Jang E, Gu S, Poole B. Categorical reparameterization with gumbel- softmax[J]. arXiv preprint arXiv:1611.01144, 2016

    Pith/arXiv arXiv 2016

  17. [17]

    Spatio-temporal graph neural networks for predictive learning in urban computing: A survey[J]

    Jin G, Liang Y , Fang Y , et al. Spatio-temporal graph neural networks for predictive learning in urban computing: A survey[J]. IEEE transactions on knowledge and data engineering, 2023, 36(10): 5388-5408

    2023