Pith. sign in

REVIEW 3 major objections 4 minor 54 references

A Multi-Task Learning Approach to Linear Multivariate Forecasting

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Grouping correlated variates into separate tasks, each re-weighted by its forecast error, makes a purely linear forecaster at least as accurate as transformer-based state-of-the-art models on standard multivariate benchmarks.

desk verdict The method is a reasonable heuristic, but the paper's central 'superior' claim collapses when ILI is removed from the average; the theory overstates the gradient-angle link. read the letter →

arxiv 2502.03571 v2 pith:J7LO67KL submitted 2025-02-05 cs.LG cs.AI

Liran Nochumsohn , Hedi Zisling , Omri Azencot This is my paper · ORCID
classification cs.LGcs.AI
keywords multivariatetimeseriesforecastingmulti-tasklearninglinearvariategroupingcorrelation-basedclusteringgradientbalancingconflictsMTLinear
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 claims that multivariate time series forecasting should be treated as a multi-task learning problem: each variate (one channel of the series), or each group of correlated variates, is an optimization task in its own right, and the quality of the solution depends on how those tasks are grouped and balanced. By differentiating the loss of a linear forecaster, the authors show that each variate contributes a gradient along its own input direction and scaled by its prediction error, which motivates two interventions: group variates by absolute Pearson correlation, and re-weight each per-element loss by inverse error statistics. The resulting model, MTLinear, gives each variate group its own linear head and applies both interventions together. On seven standard benchmarks spanning electricity, traffic, weather, exchange rates, and epidemiology, the paper reports that MTLinear is on par with or better than transformer-based state-of-the-art baselines, with a best global MSE of 0.550 and MAE of 0.423. If correct, this means simple linear architectures with principled task grouping can replace channel-independent transformer decoders in practical forecasting systems.

What carries the argument

The central object is the per-variate gradient decomposition of the linear forecasting loss, Eq. (6): for each horizon $j$, $\nabla_{\theta_j} F(\theta_j) = \frac{1}{k} \sum_{i=1}^k 2 x_i (x_i^T \theta_j - y_{j,i})$, meaning variate $i$ presses on the shared weights along its own input direction $x_i$ with strength proportional to its residual $e_{i,j} = x_i^T \theta_j - y_{j,i}$. This identity does two jobs: it turns task alignment into a geometric statement — the angle between task gradients parallels the angle between variates, which on centered data is the Pearson correlation — and it turns task imbalance into an error-scaling statement: variates with larger residuals push harder. Two mechanisms convert the identity into a model: agglomerative (bottom-up) clustering of the absolute correlation matrix $R_X$ under a threshold $\bar\alpha$, assigning each resulting group its own linear head; and the error-scaled weights $w^a_{i,j} = 1/(K_j H_i)^a$, treated as constants during training so the balancing step costs $O(1)$ rather than the per-task backward passes of gradient-manipulation baselines. Each head is built from the DLinear or NLinear linear modules, and the heads train in parallel, giving MTLinear its multi-head structure.

What would settle it

On a dataset whose cross-variate correlations drift (check first by comparing rolling Pearson correlation matrices on the training portion of ECL or Weather), train MTLinear with clusters frozen from the first half of the data and compare against a version whose clusters are recomputed from later windows; if re-clustering improves MSE by more than the seed-to-seed standard deviations the paper reports in its appendix, the fixed-cluster premise is wrong. A second, cheaper observation: log the signs of the residuals $e_{i,j}$ during training, and if the signs flip frequently within correlated variate pairs, then Eq. (6) implies gradient directions do not actually track variate directions, so the theoretical motivation for correlation-based grouping would need to be revisited.

Watch

Extended reading notes

Core claim

The paper's central claim is that the standard practice of forecasting each variate independently with one shared model leaves real performance on the table, and that the remedies for multivariate forecasting can be read off the gradients of a linear forecaster. For a linear layer applied along the time axis, the gradient for a forecast horizon $j$ decomposes per variate as $\nabla_{\theta_j} F(\theta_j) = \frac{1}{k} \sum_{i=1}^k 2 x_i (x_i^T \theta_j - y_{j,i})$, so each variate's update points along its own input vector $x_i$ and is scaled by twice its signed residual. From this the paper derives two interventions: cluster the variates by absolute correlation (agglomeratively, under an angle threshold $\bar\alpha$) so that strongly related variates share one linear head and conflicting variates are separated; and multiply each per-element loss by $w^a_{i,j} = 1/(K_j H_i)^a$, where $K_j$ averages errors across variates at horizon $j$ and $H_i$ across horizons for variate $i$, so that harder variates no longer dominate the gradient. The combined model, MTLinear, is reported to match or improve on transformer-based state of the art on the ETT family (four variants), ECL, Exchange, Traffic, Weather, and ILI, and the ablations attribute the gain to the two components working together rather than either alone.

Load-bearing premise

The method clusters variates only once, from the correlations visible in the input window, and never revisits those clusters during training, so everything depends on that one-time grouping staying valid across the whole training run and at every forecast horizon; the gradient calculation meant to justify the grouping does not fully prove it, because it contains a signed error term whose sign can change which way a variate's update points.

Editorial extensions

If this is right

  • If the headline result holds, channel-independent transformer forecasting is not the only route to state-of-the-art accuracy on these benchmarks: a purely linear model with correlation-based grouping and error-scaled losses matches or beats it.
  • The grouping and error-scaling steps are cheap and model-agnostic, and the linear-probing experiments show MTLinear can replace PatchTST's decoder, so the module is a candidate building block for other forecasting architectures.
  • The paper's ablation indicates the improvement comes from grouping and balancing in combination, so adopting just one of the two interventions in an existing model is not expected to reproduce the gains.
  • Longer lookbacks improve MTLinear consistently, a property the paper notes is not typical of transformer-based models, which suggests linear models may make better use of extended context on these benchmarks.
  • MTLinear's heads train in parallel and share the depth of a single layer in memory, which makes the approach practical even for high-dimensional variate sets such as Traffic's 862 channels.

Reading between the lines

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

  • The clusters are computed once from input-window correlations and never updated, so the method implicitly bets that this correlation geometry persists; updating the groupings during training — which the paper lists as future work — is the natural experiment to test that bet on datasets with drifting dependence structure.
  • Because Eq. (6) scales each variate's update by a signed residual, the angle between task gradients is only strictly the angle between variates when residual signs are consistent; directly measuring gradient angles during training would show how much of the grouping benefit the analysis explains versus the empirical conflict counts in Fig. 1.
  • Both the grouping threshold $\bar\alpha$ and the penalty exponent $a$ are selected per dataset by grid search, so the transferability of a single fixed configuration across datasets is not yet established; evaluating MTLinear with one fixed setting everywhere would separate the method's contribution from its tuning.
  • Because MTLinear never shares weights across clusters, it cannot exploit predictive information in weakly correlated variates even when such information exists; combining grouped heads with a shared channel-independent head is a concrete extension that the paper's own limitation discussion invites.
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

3 major / 4 minor

Summary. The paper proposes a multi-task learning view of multivariate time-series forecasting, in which each variate (or group of correlated variates) is treated as a separate task. The authors analyze the gradient of a linear forecast model, conclude that task alignment can be deduced from input-variate correlation, and build MTLinear: a multi-head linear forecaster that clusters variates by absolute Pearson correlation and reweights per-element losses by inverse error statistics. They evaluate MTDLinear and MTNLinear on nine standard benchmarks and report that the method 'has an overall superior performance' with a best global MSE of 0.550 and MAE of 0.423 in Table 1, alongside extensive ablations, gradient-manipulation comparisons, and a convexity proof for the weighted linear objective.

Significance. If the empirical and theoretical claims were fully supported, the paper would be a useful contribution: it would show that a simple, parallelizable multi-head linear model with correlation-based task grouping and error-scaled gradients can compete with transformer-based forecasters, and it would offer a reusable decoder-style module. The ablation study (Table 2), the comparisons to gradient manipulation baselines (Table 3), and the availability of code are concrete strengths. However, the headline aggregate claim is fragile because it is driven by a single dataset, and the analytic motivation in Sec. 4.1 is not established as stated. Both issues are central to the paper's message and require correction before the contribution can be assessed reliably.

major comments (3)
  1. [Sec. 5, Table 1] The claim of 'overall superior performance' rests on an unweighted average over nine datasets in which ILI dominates. Recomputing the Table 1 MSE averages without ILI gives approximately MTDLinear 0.367, MTNLinear 0.373, iTransformer 0.345, and PatchTST 0.354, so the two transformer baselines beat the MTLinear variants on the eight remaining datasets. The ILI advantage of MTNLinear over iTransformer is roughly 0.77 MSE, which, divided by nine, is the main driver of the reported 0.06 global gap. The paper reports no paired significance tests and no standard deviations for the imported baseline numbers (Table 9 gives standard deviations only for MTLinear), so the aggregate comparison is fragile. The authors should either replace the aggregate claim with per-dataset conclusions, use a statistically grounded aggregation, or substantially weaken the claim.
  2. [Sec. 4.1, Eq. (6)] The statement that 'the angle between gradients is equivalent to the angle between variates x_a and x_b' is not implied by Eq. (6). Each variate's contribution to the gradient is g_i = 2 e_{i,j} x_i, where e_{i,j} = x_i^T θ_j - y_{j,i}, so the direction of g_i is sign(e_{i,j}) times the direction of x_i. The angle between g_a and g_b equals the angle between x_a and x_b only when the residuals e_{a,j} and e_{b,j} have the same sign; with opposite signs the gradient angle is π minus the variate angle. Moreover, the full gradient in Eq. (6) is a sum over variates and horizons, so the pairwise 'task gradient' notion needs a precise definition. The empirical conflict-count experiments in Sec. 5.2 and Fig. 6 may still support correlation-based grouping, but the analytic motivation as written is not established.
  3. [Sec. 4.2] The variate clustering is computed once from the absolute correlation matrix R_X of the input lookback and is never updated during training. The paper's own conclusion states that learning the clusters during training is future work, which acknowledges that fixed clusters are a limitation. If cross-variate linear structure is non-stationary, or if the lookback correlation is not representative of the forecast-horizon behavior, the fixed groups may be misspecified and the reported gains may not transfer. The authors should report cluster stability across seeds and horizons, or evaluate a variant that recomputes or updates the clusters, before claiming that the grouping mechanism is generally reliable.
minor comments (4)
  1. [Eqs. (7)-(8)] The notation for the weight matrix is inconsistent: Eq. (7) uses W^a while the text and Eq. (8) define w^a_{i,j}, and the relationship between the matrix W^a and the scalar weights should be stated explicitly.
  2. [Appendix B] The Hessian 2 X_W X_W^T is positive semi-definite, not necessarily strictly positive definite, so the text should not call it an SPD matrix without qualification; this affects the precision of the convexity statement, though not the convergence conclusion.
  3. [Tables 1 and 7] The reported global MAE for MTNLinear differs between Table 1 (0.423) and Table 7 (0.422); the discrepancy should be resolved.
  4. [Abstract] The statement that recent state-of-the-art works 'ignore the inter-relations between variates, using their model on each variate independently' is overbroad, since iTransformer and Crossformer explicitly model variate dependencies; the introduction later cites these works, so the abstract should be qualified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MTLinear's grouping and gradient-scaling components are heuristic design choices validated against external benchmarks, not results derived from their own assumptions.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The proposed variate grouping by absolute correlation (Sec. 4.2) and the error-based gradient penalty (Sec. 4.3) are algorithmic design choices, motivated heuristically by the gradient analysis in Sec. 4.1; they are not fitted to the test benchmarks and then relabeled as predictions. The central claim of 'overall superior performance' is an empirical comparison against externally reported baseline numbers, with per-horizon results, ablations, and standard deviations for the proposed method. The gradient-angle statement in Sec. 4.1 is imprecise because each task gradient carries the sign of the residual, so the angle between gradients is not generally equal to the angle between variates; however, that is a soundness or correctness issue, not circularity, since the clustering heuristic stands or falls on the empirical ablation evidence. Self-citations in the related-work section are background references and are not load-bearing for the method's validity. The paper also explicitly acknowledges its limitations, including the fixed-cluster assumption and reliance on specific linear layer types. No fitted constant is disguised as a derived quantity, and no load-bearing result is imported from a self-citation chain.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. Its free parameters are two validation-tuned hyperparameters. The conceptual contributions rest on domain assumptions about correlation stability and error weighting; the analytic motivation has a sign error that is a soundness issue, not a circularity issue.

free parameters (2)
  • alpha_bar (grouping threshold) = grid {pi/2, pi/3, pi/4, pi/6, 0}; selected per dataset on validation
    Controls how many variate groups and linear heads are created; changes model capacity and results (Fig. 4, Tab. 6).
  • a (gradient penalty exponent) = grid {0, 1, 2}; a=2 selected in most cases
    Controls the strength of error-based loss reweighting; ablated in App. E.
assumptions (4)
  • standard math The gradient of the unweighted MSE linear forecaster is 2/k times the sum over variates of x_i times the residual (x_i^T theta_j - y_{j,i}).
    Elementary calculus; derived in Sec. 4.1 and App. A.
  • domain assumption Correlated variates with high absolute Pearson correlation share similar optimization trajectories, so grouping them reduces gradient conflicts.
    Empirical Fig. 1 and Sec. 5.2 suggest this, but the analytic claim in Sec. 4.1 ignores the sign of residuals and is not proven.
  • domain assumption The cross-variate correlation structure in the lookback window persists into the forecast horizon.
    Clusters are computed once from R_X in Sec. 4.2 and never updated; no stationarity test is provided.
  • domain assumption Weighting the loss by w^a_{i,j} = 1/(K_j H_i)^a and treating the weights as constants still optimizes the unweighted MSE that is reported.
    The surrogate loss differs from the evaluation metric; the convergence proof in App. B covers the weighted loss only, not equivalence to unweighted MSE.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Multi-Task Learning Approach to Linear Multivariate Forecasting." pith.science (2026). https://pith.science/paper/J7LO67KL

@misc{pith2026250203571,
  author       = {Pith},
  title        = {Pith review of: A Multi-Task Learning Approach to Linear Multivariate Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7LO67KL}},
  note         = {Machine review of arXiv:2502.03571}
}
read the original abstract

Accurate forecasting of multivariate time series data is important in many engineering and scientific applications. Recent state-of-the-art works ignore the inter-relations between variates, using their model on each variate independently. This raises several research questions related to proper modeling of multivariate data. In this work, we propose to view multivariate forecasting as a multi-task learning problem, facilitating the analysis of forecasting by considering the angle between task gradients and their balance. To do so, we analyze linear models to characterize the behavior of tasks. Our analysis suggests that tasks can be defined by grouping similar variates together, which we achieve via a simple clustering that depends on correlation-based similarities. Moreover, to balance tasks, we scale gradients with respect to their prediction error. Then, each task is solved with a linear model within our MTLinear framework. We evaluate our approach on challenging benchmarks in comparison to strong baselines, and we show it obtains on-par or better results on multivariate forecasting problems. The implementation is available at: https://github.com/azencot-group/MTLinear

Figures

Figures reproduced from arXiv: 2502.03571 by the authors.

Figure 1
Figure 1. The total number of conflicts as a function of epochs. Colored lines represent variate pairs. Pairs with [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Our pipeline consists of three steps: estimating variate correlations, variate clustering, and assigning a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. MSE results for different lookback lengths with a forecast horizon of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: MSE measures for different clustering α¯. The red dashed line is the mean for iTransformer. The results suggest that MTLinear is comparable or better in comparison to iTransformer [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: We plot the error ei,j of a given loss vs. its gradient’s magnitude. These results highlight the clear positive correlation between the two for both DLinear and NLinear. errors corresponding to larger gradients—motivating our penalty in Sec. 4.3. 5.4 Model Variations a…
Figure 6
Figure 6. Figure 6: The Pearson correlation matrix (rightmost) and conflict matrices, where each element represents the [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: The dendrograms for the ILI and ETTm2 datasets. The red lines represent the cut points [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: A performance comparison between different values of the sensitivity parameter [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: 2D (left) and 3D (right) PCA applied to ETTm2, ECL, and Weather datasets. PCA assists with [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: 2D (left) and 3D (right) PCA applied to Exchange, Traffic, and ILI datasets. PCA assists with [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Comparison of the distribution of the correlation coefficient between variates between ECL and Traffic. [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 47 canonical work pages

  1. [1]

    Boyd and L

    S. Boyd and L. Vandenberghe.Convex optimiza- tion. Cambridge university press, 2004

    work page 2004

  2. [2]

    Chang, W.-Y

    C. Chang, W.-Y. Wang, W.-C. Peng, and T.- F. Chen. LLM4TS: Aligning Pre-Trained LLMs as Data-Efficient Time-Series Forecasters.arXiv preprint arXiv:2308.08469, 2023

    arXiv 2023

  3. [3]

    Z. Chen, V. Badrinarayanan, C.-Y. Lee, and A. Ra- binovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794–803. PMLR, 2018

    work page 2018

  4. [4]

    Z. Chen, E. Jiaze, X. Zhang, H. Sheng, and X. Cheng. Multi-task time series forecasting with shared attention. In2020 International Confer- ence on Data Mining Workshops (ICDMW), pages 917–925. IEEE, 2020

    work page 2020

  5. [5]

    K. Cho, B. van Merrienboer, Ç. Gülçehre, D. Bah- danau, F. Bougares, H. Schwenk, and Y. Ben- gio. Learning phrase representations using RNN encoder-decoder for statistical machine transla- tion. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Process- ing, EMNLP, pages 1724–1734. ACL, 2014

    work page 2014

  6. [6]

    Cohen, Y

    I. Cohen, Y. Huang, J. Chen, J. Benesty, J. Ben- esty, J. Chen, Y. Huang, and I. Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1–4, 2009

    work page 2009

  7. [7]

    Fifty, E

    C. Fifty, E. Amid, Z. Zhao, T. Yu, R. Anil, and C. Finn. Efficiently Identifying Task Groupings for Multi-Task Learning.Advances in Neural In- formation Processing Systems, 2021

    work page 2021

  8. [8]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville.Deep Learning. MIT press, 2016

    work page 2016

Show all 54 references
  1. [9]

    C. W. Granger. Investigating causal relations by econometric models and cross-spectral methods. Econometrica: journal of the Econometric Society, pages 424–438, 1969

    1969

  2. [10]

    Groenendijk, S

    R. Groenendijk, S. Karaoglu, T. Gevers, and T. Mensink. Multi-Loss Weighting with Coefficient of Variations. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1469–1478, 2020

    2020

  3. [11]

    Guangyuan, Q

    S. Guangyuan, Q. Li, W. Zhang, J. Chen, and X.- M. Wu. Recon: Reducing Conflicting Gradients From the Root For Multi-Task Learning. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  4. [12]

    Hastie, R

    T. Hastie, R. Tibshirani, J. H. Friedman, and J. H. Friedman. The Elements of Statistical Learning: Data mining, Inference, and Prediction, volume 2. Springer, 2009

    2009

  5. [13]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber. Long Short- term Memory. Neural computation, 1997

    1997

  6. [14]

    Ishihara, A

    K. Ishihara, A. Kanervisto, J. Miura, and V. Hau- tamaki. Multi-task learning with attention for end-to-end autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2902–2911, 2021

    2021

  7. [15]

    Kaufman and O

    I. Kaufman and O. Azencot. Data representations’ study of latent image manifolds. InInternational Conference on Machine Learning, pages 15928– 15945. PMLR, 2023

    2023

  8. [16]

    Kaufman and O

    I. Kaufman and O. Azencot. Analyzing deep trans- former models for time series forecasting via man- ifold learning. Transactions on Machine Learning Research, TMLR, 2024

    2024

  9. [17]

    Kaufman and O

    I. Kaufman and O. Azencot. First-order mani- fold data augmentation for regression learning. In Forty-first International Conference on Machine Learning, ICML, 2024

    2024

  10. [18]

    H. Li, Z. Xu, G. Taylor, C. Studer, and T. Gold- stein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018

    2018

  11. [19]

    Z. Li, S. Qi, Y. Li, and Z. Xu. Revisiting long-term time series forecasting: An investigation on linear mapping. arXiv preprint arXiv:2305.10721, 2023

    2023 arXiv

  12. [20]

    B. Liu, X. Liu, X. Jin, P. Stone, and Q. Liu. Conflict-Averse Gradient Descent for Multi-task Learning. Advances in Neural Information Pro- cessing Systems, 2021

    2021

  13. [21]

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

    2023 arXiv

  14. [22]

    J. Ma, Z. Zhao, X. Yi, J. Chen, L. Hong, and E. H. Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD interna- tional conference on knowledge discovery & data mining, pages 1930–1939, 2018

    1930

  15. [23]

    Misra, A

    I. Misra, A. Shrivastava, A. Gupta, and M. Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994–4003, 2016. Liran Nochumsohn, Hedi Zisling, Omri Azencot

    2016

  16. [24]

    Naiman, N

    I. Naiman, N. Berman, I. Pemper, I. Arbiv, G. Fad- lon, and O. Azencot. Utilizing image transforms and diffusion models for generative modeling of short and long time series.Advances in Neural Information Processing Systems, 2024

    2024

  17. [25]

    Naiman, N

    I. Naiman, N. B. Erichson, P. Ren, M. W. Ma- honey, and O. Azencot. Generative modeling of regular and irregular time series data via koopman vaes. InThe Twelfth International Conference on Learning Representations,ICLR, 2024

    2024

  18. [26]

    Navon, A

    A. Navon, A. Shamsian, I. Achituve, H. Maron, K. Kawaguchi, G. Chechik, and E. Fetaya. Multi- task learning as a bargaining game. InInterna- tional Conference on Machine Learning, ICML, volume 162 ofProceedings of Machine Learning Research, pages 16428–16446. PMLR, 2022

    2022

  19. [27]

    Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. InThe Eleventh International Conference on Learning Representations, ICLR, 2023

    2023

  20. [28]

    Nochumsohn, M

    L. Nochumsohn, M. Moshkovitz, O. Avner, D. Di Castro, and O. Azencot. Beyond data scarcity: A frequency-driven framework for zero- shot forecasting.arXiv preprint arXiv:2411.15743, 2024

    2024 arXiv

  21. [29]

    B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio. N-BEATS: neural basis expansion anal- ysis for interpretable time series forecasting. In 8th International Conference on Learning Repre- sentations, ICLR, 2020

    2020

  22. [30]

    Pascanu, T

    R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training recurrent neural networks. In International conference on machine learning, pages 1310–1318. Pmlr, 2013

    2013

  23. [31]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Brad- bury, G.Chanan, T.Killeen, Z.Lin, N.Gimelshein, L. Antiga, et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. Ad- vances in Neural Information Processing Systems, 32, 2019

    2019

  24. [32]

    P. Ren, R. Nakata, M. Lacour, I. Naiman, N. Nakata, J. Song, Z. Bi, O. A. Malik, D. Mo- rozov, O. Azencot, et al. Learning physics for unveiling hidden earthquake ground motions via conditional generative modeling.arXiv preprint arXiv:2407.15089, 2024

    2024 arXiv

  25. [33]

    Salinas, V

    D. Salinas, V. Flunkert, J. Gasthaus, and T. Januschowski. DeepAR: Probabilistic forecast- ing with autoregressive recurrent networks.Inter- national Journal of Forecasting, 36(3):1181–1191, 2020

    2020

  26. [34]

    Schreiber

    T. Schreiber. Measuring information transfer. Physical review letters, 85(2):461, 2000

    2000

  27. [35]

    Sener and V

    O. Sener and V. Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018

    2018

  28. [36]

    X. Song, S. Zheng, W. Cao, J. Yu, and J. Bian. Efficient and Effective Multi-task Grouping via Meta Learning on Task Combinations.Advances in Neural Information Processing Systems, 2022

    2022

  29. [37]

    Standley, A

    T. Standley, A. Zamir, D. Chen, L. Guibas, J. Ma- lik, and S. Savarese. Which tasks should be learned together in multi-task learning? InInternational Conference on Machine Learning, pages 9120–9132. PMLR, 2020

    2020

  30. [38]

    Vandenhende, S

    S. Vandenhende, S. Georgoulis, W. Van Gans- beke, M. Proesmans, D. Dai, and L. Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614–3633, 2021

    2021

  31. [39]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  32. [40]

    H. Wu, J. Xu, J. Wang, and M. Long. Auto- former: Decomposition Transformers with Auto- Correlation for Long-Term Series Forecasting.Ad- vances in Neural Information Processing Systems, 2021

    2021

  33. [41]

    W. Xue, T. Zhou, Q. Wen, J. Gao, B. Ding, and R. Jin. Make transformer great again for time series forecasting: Channel aligned robust dual transformer. arXiv preprint arXiv:2305.12095, 2023

    2023 arXiv

  34. [42]

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Haus- man, and C. Finn. Gradient Surgery for Multi- Task Learning.Advances in Neural Information Processing Systems, 2020

    2020

  35. [43]

    A. R. Zamir, A. Sax, W. Shen, L. J. Guibas, J. Ma- lik, and S. Savarese. Taskonomy: Disentangling task transfer learning. InProceedings of the IEEE conference on computer vision and pattern recog- nition, pages 3712–3722, 2018

    2018

  36. [44]

    A. Zeng, M. Chen, L. Zhang, and Q. Xu. Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023. A Multi-T ask Learning Approach to Linear Multivariate F orecasting

    2023

  37. [45]

    Zhang and J

    Y. Zhang and J. Yan. Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multi- variate Time Series Forecasting. InThe Eleventh International Conference on Learning Representa- tions, 2022

    2022

  38. [46]

    Zhang and Q

    Y. Zhang and Q. Yang. A survey on multi-task learning. IEEE Transactions on Knowledge and Data Engineering, 34(12):5586–5609, 2021

    2021

  39. [47]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang. Informer: Beyond Effi- cient Transformer for Long Sequence Time-Series Forecasting. InProceedings of the AAAI confer- ence on artificial intelligence, 2021

    2021

  40. [48]

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin. FEDformer: Frequency Enhanced De- composed Transformer for Long-term Series Fore- casting. In International Conference on Machine Learning. PMLR, 2022

    2022

  41. [49]

    T. Zhou, P. Niu, L. Sun, R. Jin, et al. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems, 36:43322–43355, 2023. Liran Nochumsohn, Hedi Zisling, Omri Azencot A Linear Analysis Proof Below, we provide the full d...

    2023

  42. [50]

    For all models and algorithms presented, check if you include: (a) A clear description of the mathematical setting, assumptions, algorithm, and/or model. [Yes. Mentioned in 2] (b) An analysis of the properties and complexity (time, space, sample size) of any algorithm. [Yes. M...

  43. [51]

    For any theoretical claim, check if you include: (a) Statements of the full set of assumptions of all theoretical results. [Yes. See App. A] (b) Complete proofs of all theoretical results. [Yes. See App. A] (c) Clear explanations of any assumptions. [Yes. See App. A]

  44. [52]

    For all figures and tables that present empirical results, check if you include: (a) The code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL). [No. The code will be provided upon acceptance, but all th...

  45. [53]

    If you are using existing assets (e.g., code, data, models) or curating/releasing new assets, check if you include: (a) Citations of the creator If your work uses existing assets. [Yes. Citation are present for relevant models and datasets.] (b) The license information of the ...

  46. [54]

    [Not Applicable] (b) Descriptions of potential participant risks, with links to Institutional Review Board (IRB) approvals if applicable

    If you used crowdsourcing or conducted research with human subjects, check if you include: (a) The full text of instructions given to participants and screenshots. [Not Applicable] (b) Descriptions of potential participant risks, with links to Institutional Review Board (IRB) ...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.