Pith. sign in

REVIEW 4 major objections 4 minor 42 references

When GNNs Fail: Quantifying and Overcoming Temporal Correlation Volatility in Time Series

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Forecasting accuracy under shifting correlations is governed by a measurable volatility; a new GLIDE layer beats 18 baselines on eight benchmarks.

desk verdict The dynamic adjacency in Eqs. (2)-(3) leaks future data as written, so GLIDE's headline gains are not trustworthy until the authors confirm a causal window; the D2 static/dynamic split is still worth a look. read the letter →

arxiv 2608.07333 v1 pith:WCK6DXAG submitted 2026-08-07 cs.LG

Chen Shao , Yue Wang , Zhenyi Zhu , Zhanbo Huang , Tobias Käfer , Zonghan Wu , Danai Koutra This is my paper · ORCID
classification cs.LG
keywords multivariatetimeseriesforecastinggraphneuralnetworkstemporalcorrelationvolatilitydynamicstructuremessagepassingstructuralbreakstopology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that multivariate time-series forecasting models which assume a fixed correlation graph—spatial and spectral GNNs and Transformers—systematically degrade when pairwise correlations shift rapidly over time, and that the degradation is predictable. To quantify this, it proposes Temporal Correlation Volatility (TCV), the average change between consecutive correlation matrices, scaled to fall between 0 and 1. It then introduces GLIDE, a graph layer combining path-based message passing with separate static and dynamic adjacency channels, and reports that GLIDE ranks first among 18 baselines on all eight benchmarks, with average RMSE reductions up to 45.6% and a maximum MAE reduction of 85.7% on Exchange Rate at horizon 12. The practical upshot is that a measurable diagnostic plus a two-channel layer can replace the static-topology assumption in volatile settings.

What carries the argument

Three components carry the argument. First, TCV, defined as the time-averaged Frobenius distance between consecutive correlation matrices, quantifies structural volatility and is shown empirically to predict the performance collapses of existing models. Second, path-based message passing replaces the usual one-hop adjacency with a weighted polynomial reachability, so that signals propagate over multi-hop paths that are more stable than individual edges. Third, static/dynamic propagation separation computes two adjacency matrices—a learned time-invariant one and a short-window gradient-based one—whose design is motivated by a kernel-weighted graphical lasso consistency theorem, and combines their polynomial expansions additively.

What would settle it

Run GLIDE with the dynamic adjacency computed strictly from past observations (causal masking) and compare the errors with the paper's reported numbers; if the large gains collapse, the centered window in Equations (2) and (3) is leaking future information.

Watch

Extended reading notes

Core claim

The paper's central claim is that forecasting performance under shifting inter-variable dependencies is controlled by the volatility of the correlation graph, not merely by the strength of the correlations themselves. Static-graph models, and especially Transformers whose self-attention implicitly induces a graph, lose up to 36.6% accuracy on synthetic dynamic data, while a structure-free MLP degrades only 14.6%. GLIDE is designed to survive this regime: it aggregates along weighted paths of the correlation graph, which remain stable when single edges flicker, and it splits the adjacency into a static backbone and a gradient-driven dynamic part. The authors report that GLIDE achieves the best MAE/RMSE in every tested setting, including static and dynamic synthetic regimes and all six real-world benchmarks, with the largest relative gains on the high-TCV Germany and France energy datasets.

Load-bearing premise

The load-bearing premise is that the dynamic graph at the last observed time step is built only from past observations, but the paper's centered-window formulas include future observations at that step and no causal masking is described, so the reported gains could be inflated if future values leak into the adjacency.

Editorial extensions

If this is right

  • If the TCV diagnosis is correct, practitioners can measure a dataset's correlation volatility and decide in advance whether a static-graph forecaster will be reliable or needs a volatility-aware method.
  • If GLIDE's reported results hold, the static/dynamic split plus path-based aggregation is a practical recipe for forecasting under structural breaks while remaining competitive on near-static benchmarks.
  • The paper's claims imply that models learning an implicit graph through attention are particularly fragile under high TCV, losing up to 36.6% on synthetic dynamic data.
  • A direct corollary is that benchmark selection should control for TCV, since average performance across mixed static and dynamic datasets can hide volatility-induced failures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The centered-window construction of the dynamic adjacency may leak future observations into the forecast if no causal masking is applied; running a strictly causal version of GLIDE would test whether the headline gains survive.
  • TCV as defined captures only linear, Pearson-type dependence; extending it to rank or kernel correlations could reveal whether the same diagnostic applies under nonlinear coupling.
  • Separating persistent and transient structure, while weighting multi-hop paths, could transfer to other dynamic-graph prediction tasks beyond time-series forecasting, such as traffic or financial network prediction.
  • One testable extension is to use TCV as an automatic selector: compute TCV on rolling segments and adaptively re-weight the static and dynamic branches of GLIDE to see whether accuracy improves in real time.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Temporal Correlation Volatility (TCV), a metric intended to quantify how much the pairwise correlation structure of multivariate time series changes over time, and claims that high TCV predicts performance degradation of graph-based and Transformer forecasting models. To address this degradation, the authors introduce GLIDE, a GNN layer with path-based message passing (D1) and a static/dynamic propagation separation (D2), and report state-of-the-art results across eight benchmarks with up to 45.6% average RMSE reduction and 85.7% MAE reduction. The paper includes a theoretical discussion in Appendix B, ablation studies, and a public code repository.

Significance. If the empirical results survive scrutiny, the paper would make two useful contributions: a model-agnostic diagnostic metric for topology drift (TCV) and a concrete architectural mechanism for separating stable from transient cross-series dependencies. The design of D2, which learns a dynamic adjacency from local windows while keeping a static backbone, is a plausible and timely idea for nonstationary multivariate forecasting. The paper also ships code and reports extensive comparisons against 18 baselines, which are helpful for reproducibility. However, the central empirical claim is currently undermined by a likely causal-leakage issue in the dynamic adjacency construction, and the theoretical justification in Theorem 1 is incomplete.

major comments (4)
  1. [§3.2, Eqs. (2)–(3)] The dynamic adjacency A_d is computed from a centered box window [t-B/2, t+B/2], which at forecast time t includes B/2 future observations. The manuscript nowhere states that the window is masked or shifted to be causal at inference; since A_d enters every aggregation in Eq. (5), the reported rank-1 results and the 45.6%/85.7% improvements could be inflated by test-time leakage. Please either implement and document a causal window (e.g., [t-B+1, t]) or justify why future values are not used, and rerun all experiments under that protocol.
  2. [§3.2, Theorem 1 and Appendix B] Theorem 1 is stated without a proof; the 'Justification' paragraph is informal and does not connect the kernel-weighted similarity in Eq. (2) to the claimed Graph-LASSO identification result. The Appendix B proof concerns a different estimator (a kernel-weighted covariance matrix) and the lemmas are not shown to imply the zero-pattern recovery stated in Theorem 3. As written, the theoretical grounding for D2 is not established. Please provide a complete proof or clearly label these statements as conjectures or heuristics.
  3. [§2.2, Table 1] The claimed 'clear connection between TCV and performance degradation' is supported by only two synthetic configurations (TCV=0.296 and TCV=0.874) and qualitative dataset anecdotes. There is no correlation coefficient, confidence interval, or statistical test across a range of TCV values, so the paper does not actually quantify the relationship. Add a systematic sweep over TCV levels and report the empirical correlation between TCV and the degradation of each baseline.
  4. [§3.6, Table 3 and Appendix F, Tables 10–11] The experimental reporting contains inconsistencies that need resolution: GLIDE's RMSE on Electricity is 0.051 while its MAE is 0.010, a ratio of about 5, whereas all baselines show ratios near 1.5; and Appendix F presents two tables (Table 10 and Table 11) with identical content, which appears to be a copy-paste error. Please audit the tables, clarify whether metrics are computed in normalized or original space, and correct the duplicated table.
minor comments (4)
  1. [§3.2, Eq. (3)] The gradient variant uses ∇~x_k = ~x_k - ~x_{k-1}; when k = t - B/2, the index k-1 lies outside the stated window and is undefined. Clarify the boundary handling in the implementation.
  2. [Appendix B, Lemma 2] The proof uses quantities Φ2, Φ3, Φ4, and C1 that are never defined, and states that the 'full proof' will be given later without actually providing it. Please complete the proof or explicitly mark these as sketch-level arguments.
  3. [Table 1] The 'Relative ΔMAE (%)' column shows negative values labeled as degradation; the sign convention should be defined explicitly so the reader knows whether negative means worse or better.
  4. [§3.7, Table 4] The text states GLIDE 'consistently achieves' rank 1, but Table 4 shows that on ETTh1 at h=3, the MAE of the runner-up is 0.133 versus GLIDE's 0.135; please make the presentation of near-ties precise and report confidence intervals or significance tests for such close margins.

Circularity Check

1 steps flagged · score 7.0 of 10

Centered dynamic-graph window in Eqs. (2)-(3) makes short-horizon forecasts a function of the target values themselves.

  1. self definitional [Section 3.2, Eq. (3); used in Section 3.3, Eq. (5); window size B=24 (Table 7)]
    "The gradient variant, defined in Eq. (3) below, replaces ˜xk with the first-order temporal difference ∇˜xk = ˜xk − ˜xk−1 ... ˆA(t) d = ReLU( 1 B Pt+B/2 k=t−B/2 (∇˜xk)(∇˜xk)⊤ Md),(3) ... H=σ( XK k=0 wk s Ak s Ws H + XK k=0 wk dAk d Wd H,(5)"

    In a forecast made at the last observed time t, the centered box window [t−B/2, t+B/2] includes observations x_{t+1},...,x_{t+B/2}. With B=24 and horizons h=3,6,12 ≤ B/2, the ground-truth future value x_{t+h} is one of the terms inside the sum defining A_d^(t) (via ∇˜x_k). Eq. (5) then aggregates H using A_d, and the prediction layer maps the resulting H^(f) to Y_hat. Hence Y_hat_{t+h} is, by construction, a function of x_{t+h}. The paper never states that this window is masked or shifted to be causal at inference, so the reported 'forecast' at short horizons is not a function of the past alone; it is a function of the very target values it claims to predict.

full rationale

The paper's TCV metric, the empirical TCV-degradation connection, and the GLIDE architecture are otherwise not circular: TCV is defined from the data independently of the performance numbers, the synthetic connection is a controlled empirical comparison, and the model is trained end-to-end on the target loss. The self-citations (e.g., the Real-E benchmark [24]) and the kernel-smoothing theorem do not reduce the central claim to their own inputs. However, the dynamic adjacency in Eqs. (2)-(3) uses a centered box window that extends B/2 steps past the last observed time. For the main experiments (B=24), this places the forecast targets themselves inside the receptive field of A_d for all reported horizons (h=3,6,12). Since A_d is used in every GLIDE layer update (Eq. 5) and the final prediction is a linear projection of the propagated H, the model output is a function of the ground-truth future values at those horizons. The paper does not describe any causal masking or truncated-window implementation, so the equations as written define an input that depends on the output. This is a self-definitional circularity that plausibly inflates the reported improvements. The severity is high because the rank-1 performance and the headline percentage gains are the paper's central claims; yet the rest of the pipeline is standard supervised learning, so the circularity is partial rather than total.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central performance claim depends on learned polynomial weights, a learned dynamic mixing matrix, hand-chosen window and polynomial order, plus the unverified assumption that the dynamic adjacency can be formed without future data. TCV is a new metric but is not a free parameter; it is a defined diagnostic.

free parameters (4)
  • Polynomial coefficients w_k (k=0..K) = Learned from data; no values reported
    In Eq. (5), weights on A^k control the path-based aggregation; they are optimized on the training loss, and Fig. 2 shows their learned magnitude shifts with TCV.
  • Dynamic mixing matrix M_d = Learned from data; no values reported
    Eq. (3) multiplies the local gradient outer product by M_d before ReLU; this is a data-fitted re-scaling that defines the dynamic adjacency.
  • Sliding window size B = 24 (used for TCV and dynamic adjacency in Table 7)
    TCV and the box kernel in D2 depend on this hand-chosen window; no sensitivity analysis for the hybrid metric/architecture is reported.
  • Polynomial order K = Not reported; Fig. 2 illustrates orders 0-3
    Maximum path length in D1 is a hyperparameter; its choice affects the composition of static and dynamic branches.
assumptions (4)
  • domain assumption Pairwise inter-series dependencies are adequately represented by Pearson correlation of standardized signals.
    Def. 2 and TCV rest on this; the paper acknowledges only linear relationships in Sec. 3.9.
  • ad hoc to paper The evolving topology follows A_t = A_{t-1} + Z_t with independent zero-mean Gaussian increments.
    Theorem 1 assumption; the appendix also states '{A_t} independent random adjacency structures', which is inconsistent with the random-walk model. Not validated on real data.
  • ad hoc to paper The dynamic adjacency at forecast time t can be computed from a centered window [t-B/2, t+B/2] without using future observations.
    Eqs. (2)-(3) use centered windows; for a forecast at the last observed time this includes B/2 future steps unless causal masking or padding is described, which it is not.
  • ad hoc to paper Sparsity and smoothness conditions in Theorem 3 hold for the benchmark datasets.
    Conditions like (N+s)=o(n^{2/3}/log N) and twice-differentiable covariance are not verified.
invented entities (1)
  • Temporal Correlation Volatility (TCV) metric independent evidence
    purpose: Quantifies temporal volatility of pairwise correlations in multivariate time series; used as a diagnostic for when static graph assumptions fail and to motivate GLIDE.
    TCV is fully specified by Eq. (1) and can be computed on any dataset, so it provides a falsifiable quantitative handle outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When GNNs Fail: Quantifying and Overcoming Temporal Correlation Volatility in Time Series." pith.science (2026). https://pith.science/paper/WCK6DXAG

@misc{pith2026260807333,
  author       = {Pith},
  title        = {Pith review of: When GNNs Fail: Quantifying and Overcoming Temporal Correlation Volatility in Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WCK6DXAG}},
  note         = {Machine review of arXiv:2608.07333}
}
read the original abstract

Modeling multivariate time series by representing them as graphs, where individual series act as nodes and pairwise temporal corre- lations serve as edges, has gained significant traction. Recent advances in Graph Neural Networks (GNNs) have demonstrated strong perfor- mance by assuming a static graph topology and aggregating information from neighboring series. In this work, we investigate the representa- tional power of GNNs for forecasting under both static and dynamic settings (i.e., when pairwise correlations evolve drastically over time) and identify critical limitations in current architectures. To formalize this, we first propose Temporal Correlation Volatility (TCV), a model- agnostic metric designed to quantify the distributional evolution of these latent structures. We establish a clear connection between TCV and performance degradation, demonstrating that many popular models, including Transformers, generalize poorly in high-TCV settings and are often outperformed by simple structure-agnostic baselines. To address these limitations, we propose Graph Layer for Inference in Dynamic En- vironments (GLIDE), a novel GNN layer enhanced by two theoretically grounded design mechanisms: (D1) Path-based Message Passing, which captures path-based neighborhoods and (D2) Static and Dynamic Propagation Separation, which identifies optimal dynamics via local static approximation. These components significantly improve learning under dynamic topology while preserving robustness in static scenarios. Ex- tensive experiments on synthetic and real-world benchmarks show that GLIDE improves average performance by up to 45.6% across static and dynamic settings, with the largest gain reaching 85.7%. The source code is available at https://github.com/ChenS676/GLIDE.

Figures

Figures reproduced from arXiv: 2608.07333 by the authors.

Figure 1
Figure 1. Core motivation of our study: GNN/Transformers fail when the underlying [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of learned coefficient magnitudes [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The framework of GLIDE. It consists of K TC-layers, 1-GLIDE layer and one prediction layer. The inputs are first transformed by a normalization layer and then passed to the TC-Layer followed by the GLIDE. Each layer has residual connections and is skip-connected to the output layer. difference ∇x˜k = x˜k − x˜k−1 so the dynamic branch isolates transient shocks while the static branch absorbs the slow-varying correlat… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Model complexity comparison on the Germany dataset. Bubble areas [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Ablation study of GLIDE. (a–b) Comparison with base ablations (w.o. D1 and D2). (c–d) Sensitivity analysis of D2 removing dynamic topology (D2- NoGraph) and a random matrix baseline (D2-Random). Both D1 and D2 con￾tribute to GLIDE ’s gains, with D2 most critical under …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 32 canonical work pages

  1. [1]

    In: Advances in Neural Information Processing Systems 33 (NeurIPS 2020) (2020)

    Bai, L., Yao, L., Li, C., Wang, X., Wang, C.: Adaptive graph convolutional recurrent network for traffic forecasting. In: Advances in Neural Information Processing Systems 33 (NeurIPS 2020) (2020)

    work page 2020

  2. [2]

    ArXivabs/1803.01271(2018), https://api.semanticscholar.org/CorpusID:4747877

    Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. ArXivabs/1803.01271(2018), https://api.semanticscholar.org/CorpusID:4747877

    arXiv 2018

  3. [3]

    In: Advances in Social Networks Analysis and Mining 2013, ASONAM

    Berlingerio, M., Koutra, D., Eliassi-Rad, T., Faloutsos, C.: Network similarity via multiple social theories. In: Advances in Social Networks Analysis and Mining 2013, ASONAM. pp. 1439–1440. ACM (2013)

    work page 2013

  4. [4]

    ACM Comput

    Brugere, I., Gallagher, B., Berger-Wolf, T.Y.: Network structure inference, a survey: Motivations, methods, and applications. ACM Comput. Surv.51(2) (Apr 2018)

    work page 2018

  5. [5]

    Campbell, A., Zippo, A.G., Passamonti, L., Toschi, N., Lio, P.: Dyndepnet: Learning time-varying dependency structures from fmri data via dynamic graph structure learning (2023),https://arxiv.org/abs/2209.13513

    work page Pith review arXiv 2023

  6. [6]

    Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting

    Cao, D., Wang, Y., Duan, J., Zhang, C., Zhu, X., Huang, C., Tong, Y., Xu, B., Bai, J., Tong, J., Zhang, Q.: Spectral temporal graph neural network for multivariate time-series forecasting. CoRRabs/2103.07719(2021)

    work page Pith review arXiv 2021

  7. [7]

    In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada

    Hallac, D., Park, Y., Boyd, S.P., Leskovec, J.: Network inference via the time- varying graphical lasso. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada. ACM (2017). https://doi.org/10.1145/3097983.3098037

    work page arXiv 2017

  8. [8]

    In: International Conference on Artificial Intelligence and Statistics, AISTATS

    Ito, M., Koutra, D., Wiens, J.: Understanding gnns and homophily in dynamic node classification. In: International Conference on Artificial Intelligence and Statistics, AISTATS. Proceedings of Machine Learning Research, vol. 258, pp. 2764–2772. PMLR (2025)

    work page 2025

Show all 42 references
  1. [9]

    ArXiv abs/2001.04451(2020) Overcoming Temporal Correlation Volatility in GNNs 17

    Kitaev, N., Kaiser, L., Levskaya, A.: Reformer: The efficient transformer. ArXiv abs/2001.04451(2020) Overcoming Temporal Correlation Volatility in GNNs 17

    2020 arXiv

  2. [10]

    The Annals of Applied Statistics4(1), 94–123 (2010)

    Kolar, M., Song, L., Ahmed, A., Xing, E.P.: Estimating time-varying networks. The Annals of Applied Statistics4(1), 94–123 (2010). https://doi.org/10.1214/09- AOAS308

    2010 doi

  3. [11]

    In: Proceedings of the 13th SIAM International Conference on Data Mining

    Koutra, D., Vogelstein, J.T., Faloutsos, C.: DELTACON: A principled massive-graph similarity function. In: Proceedings of the 13th SIAM International Conference on Data Mining. pp. 162–170. SIAM (2013)

    2013

  4. [12]

    The 41st International ACM SI- GIR Conference on Research & Development in Information Retrieval (2017), https://api.semanticscholar.org/CorpusID:4922476

    Lai, G., Chang, W.C., Yang, Y., Liu, H.: Modeling long- and short-term tem- poral patterns with deep neural networks. The 41st International ACM SI- GIR Conference on Research & Development in Information Retrieval (2017), https://api.semanticscholar.org/CorpusID:4922476

    2017

  5. [13]

    ACM Transactions on Knowledge Discovery from Data17, 1 – 21 (2021)

    Li, F., Feng, J., Yan, H., Jin, G., Jin, D., Li, Y.: Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution. ACM Transactions on Knowledge Discovery from Data17, 1 – 21 (2021)

    2021

  6. [14]

    In: International Conference on Learning Represen- tations (ICLR ’18) (2018),https://openreview.net/forum?id=SJiHXGWAZ

    Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. In: International Conference on Learning Represen- tations (ICLR ’18) (2018),https://openreview.net/forum?id=SJiHXGWAZ

    2018

  7. [15]

    In: Annual Conference on Neural Information Processing Systems (NeurIPS) 2022, New Orleans, LA, USA

    Liu, Y., Liu, Q., Zhang, J., Feng, H., Wang, Z., Zhou, Z., Chen, W.: Multivariate time-series forecasting with temporal polynomial graph neural networks. In: Annual Conference on Neural Information Processing Systems (NeurIPS) 2022, New Orleans, LA, USA

    2022

  8. [16]

    CoRRabs/2110.11017(2021), https://arxiv.org/abs/2110.11017

    Natali, A., Isufi, E., Coutino, M., Leus, G.: Learning time-varying graphs from online data. CoRRabs/2110.11017(2021), https://arxiv.org/abs/2110.11017

    2021 arXiv

  9. [17]

    https://doi.org/10.1088/1367-2630/ac54c9

    Nauck, C., Lindner, M., Schürholt, K., Zhang, H., Schultz, P., Kurths, J., Isenhardt, I., Hellmann, F.: Predicting basin stability of power grids using graph neural networks (2022). https://doi.org/10.1088/1367-2630/ac54c9

    2022 doi

  10. [18]

    In: The Eleventh International Conference on Learning Representations, ICLR, Kigali, Rwanda (2023),https: //openreview.net/forum?id=Jbdc0vTOcol

    Nie, Y., Nguyen, N.H., Sinthong, P., Kalagnanam, J.: A time series is worth 64 words: Long-term forecasting with transformers. In: The Eleventh International Conference on Learning Representations, ICLR, Kigali, Rwanda (2023),https: //openreview.net/forum?id=Jbdc0vTOcol

    2023

  11. [19]

    Ortega, A., Frossard, P., Kovacevic, J., Moura, J.M.F., Vandergheynst, P.: Graph signal processing: Overview, challenges, and applications. Proc. IEEE106(5), 808–828 (2018). https://doi.org/10.1109/JPROC.2018.2820126

    2018

  12. [20]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)

    Pareja, A., Domeniconi, G., Chen, J., Ma, T., Suzumura, T., Kanezashi, H., Kaler, T., Schardl, T.B., Leiserson, C.E.: EvolveGCN: Evolving graph convolutional networks for dynamic graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence (2020)

    2020

  13. [21]

    In: 2017 IEEE International Conference on Data Mining, ICDM

    Safavi, T., Sripada, C.S., Koutra, D.: Scalable hashing-based network discovery. In: 2017 IEEE International Conference on Data Mining, ICDM. pp. 405–414. IEEE Computer Society (2017)

    2017

  14. [22]

    Sezer, O.B., Gudelek, M.U., Ozbayoglu, A.M.: Financial time series forecasting with deep learning : A systematic literature review: 2005-2019 (2019), https: //arxiv.org/abs/1911.13288

    2019 arXiv

  15. [23]

    In: 9th International Conference on Learning Representations, ICLR, Virtual Event, Austria, May 3-7 (2021)

    Shang, C., Chen, J., Bi, J.: Discrete graph structure learning for forecasting multiple time series. In: 9th International Conference on Learning Representations, ICLR, Virtual Event, Austria, May 3-7 (2021)

    2021

  16. [24]

    CIKM ’25, Association for Computing Machinery (2025)

    Shao, C., Färber, M., Pütz, S., Schäfer, B., Wang, Y., Käfer, T., Huang, Z., Zhu, Z.: Real-e: A foundation benchmark for advancing robust and generalizable electricity forecasting. CIKM ’25, Association for Computing Machinery (2025). https://doi.org/10.1145/3746252.3761637 18...

    2025

  17. [25]

    In: The Twelfth In- ternational Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024.https://openreview.net/forum?id=7oLshfEIC2

    Wang, S., Wu, H., Shi, X., Hu, T., Luo, H., Ma, L., Zhang, J.Y., Zhou, J.: Timemixer: Decomposable multiscale mixing for time series forecasting. In: The Twelfth In- ternational Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024.https://openrevi...

    2024

  18. [26]

    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 38: Annual Confer- ence on Neural Information Proc...

    2024

  19. [27]

    In: The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., Long, M.: Timesnet: Temporal 2d- variation modeling for general time series analysis. In: The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda. https: //openreview.net/forum?id=ju\_Uqw384Oq

    2023

  20. [28]

    In: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, virtual

    Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, virtual. https://doi.org/10.48550/arXiv.2106.13008

    2021 doi arXiv

  21. [29]

    KDD ’20, Association for Computing Machinery, New York, NY, USA

    Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., Zhang, C.: Connect- ing the dots: Multivariate time series forecasting with graph neural net- works. KDD ’20, Association for Computing Machinery, New York, NY, USA. https://doi.org/10.1145/3394486.3403118

  22. [30]

    In: Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI, Macao, China, August 10-16, 2019

    Wu, Z., Pan, S., Long, G., Jiang, J., Zhang, C.: Graph wavenet for deep spatial- temporal graph modeling. In: Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI, Macao, China, August 10-16, 2019. https://doi.org/10.24963/IJCAI.2019/264

    2019 doi

  23. [31]

    IEEE Internet Things J

    Xu, Y., Lu, Y., Ji, C., Zhang, Q.: Adaptive graph fusion convolutional recurrent network for traffic forecasting. IEEE Internet Things J. (2023). https://doi.org/10.1109/JIOT.2023.3244182

    2023

  24. [32]

    CoRR (2025)

    Yang, W., Wu, S., Zhou, Y., Luo, C., He, X., Xie, W., Shen, L., Song, S.: Bench- marking graph representations and graph neural networks for multivariate time series classification. CoRR (2025). https://doi.org/10.48550/ARXIV.2501.08305

    2025 doi arXiv

  25. [33]

    Yi, K., Zhang, Q., Fan, W., He, H., Hu, L., Wang, P., An, N., Cao, L., Niu, Z.: Fouriergnn: Rethinking multivariate time series forecasting from a pure graph perspective. In: Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processi...

    2023 doi arXiv

  26. [34]

    In: Ben- gio, Y., LeCun, Y

    Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. In: Ben- gio, Y., LeCun, Y. (eds.) 4th International Conference on Learning Representations, ICLR San Juan, Puerto Rico, May 2-4 (2016),http://arxiv.org/abs/1511.07122

    2016 arXiv

  27. [35]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum?id=DAyKP1tvwI

    Yue, W., Liu, Y., Wang, H., Li, H., Ying, X., Guo, R., Xing, B., Shi, J.: OLinear: A linear model for time series forecasting in orthogonally transformed domain. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum...

    2025

  28. [36]

    Zeng, A., Chen, M., Zhang, L., Xu, Q.: Are transformers effective for time series forecasting? (2023)

    2023

  29. [37]

    Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2017) Overcoming Temporal Correlation Volatility in GNNs 19

    Zhang, L., Aggarwal, C.C., Qi, G.J.: Stock price prediction via discovering multi- frequency trading patterns. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2017) Overcoming Temporal Correlation Volatility in GNNs 19

    2017

  30. [38]

    In: Thirty-Fifth Conference on Artificial Intelligence

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: In- former: Beyond efficient transformer for long sequence time-series forecast- ing. In: Thirty-Fifth Conference on Artificial Intelligence. AAAI Press (2021). https://doi.org/10.1609/AAAI.V35I12.17325

    2021 doi

  31. [39]

    In: Proc

    Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., Jin, R.: FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. In: Proc. 39th International Conference on Machine Learning (ICML 2022) (2022) 20 C. Shao et al. A Notation Table 5: Summary of notatio...

    2022

  32. [40]

    Temporal Smoothness:The covariance Σ(t)is twice differentiable with respect tot, such thatmax i,j supt |σ′′ ij(t)| ≤Cfor someC >0

  33. [41]

    Sparsity:The number of active edges|Et| = s satisfies( N +s) =o(n2/3/logN )

  34. [42]

    Optimal Bandwidth:The kernel weights wst = K( |s−t| h )utilize a bandwidth h≍n −1/3. The time-varying topology at timet is identified via theℓ1-regularized objective: bΘt = arg min Θ≻0 n tr(Θ bAt)−log|Θ|+λ n∥Θ∥1 o (18) where bAt = P s wstAsP s wst is the kernel-weighted Global...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.