pith. sign in

arxiv: 2606.19413 · v1 · pith:DBN3XH47new · submitted 2026-06-17 · 💻 cs.LG

Does Text Actually Help? Uncovering and Resolving Text Collapse in Multimodal Time Series Forecasting

Huu Hiep Nguyen , Minh Hoang Nguyen , Dung Nguyen , Hung Le This is my paper

Pith reviewed 2026-06-26 21:16 UTC · model grok-4.3

classification 💻 cs.LG
keywords text collapsemultimodal time series forecastingresidual supervisionREST-TSnumerical asymmetrymodality utilizationforecasting residuals
0
0 comments X

The pith

Text collapse occurs because numerical inputs dominate multimodal time series models, but exclusive residual supervision on the text branch forces it to extract genuine content from descriptions.

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

The paper identifies text collapse as a systematic failure in multimodal time series forecasting, where the text branch converges to a content-independent output regardless of the input description. This stems from the fundamental asymmetry that numerical sequences are strongly autocorrelated with the forecast target, allowing the numerical backbone to dominate while the text branch remains underutilized. REST-TS counters the asymmetry by first generating an independent numerical forecast and then training the text branch solely to predict the residual components that numbers cannot explain. Because numerical pathways cannot reduce those losses, the text branch must draw on actual descriptive content. Evaluations across domains and architectures show improved performance and higher text utilization under this supervision.

Core claim

Existing multimodal time series forecasting frameworks exhibit text collapse, in which the text branch converges to a content-independent transformation and contributes negligible signal. REST-TS resolves the underlying asymmetry by letting the numerical backbone produce its own independent forecast while exclusively supervising the text branch to predict the structured components of the residual, the prediction gap that numerical inputs cannot reduce.

What carries the argument

REST-TS (Residual-Exclusive Supervision for Text in Time Series), which decouples numerical forecasting from residual prediction and assigns the text branch only to the latter.

If this is right

  • REST-TS achieves state-of-the-art performance across diverse real-world domains and backbone architectures.
  • Text branches demonstrate measurably greater utilization of input descriptions than in prior frameworks.
  • Supervision on residuals prevents the text branch from relying on numerical pathways.
  • The design principle applies consistently when swapping numerical backbones or forecasting tasks.

Where Pith is reading between the lines

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

  • The same residual-exclusive principle could be tested in other multimodal settings such as vision-augmented forecasting to check whether modality collapse is likewise asymmetry-driven.
  • Applying REST-TS at longer horizons might show whether text contributes more when numerical autocorrelation weakens.
  • One could measure whether the learned residual components align with the semantic elements explicitly mentioned in the text reports.
  • Synthetic experiments that inject known predictive text signals would directly test whether residual supervision recovers them.

Load-bearing premise

No numerical pathway can reduce the residual losses, so the text branch must extract genuine content from the input description.

What would settle it

A controlled run of REST-TS in which the text branch outputs remain statistically independent of the input text content, or where end-to-end performance fails to exceed a numerical-only baseline.

Figures

Figures reproduced from arXiv: 2606.19413 by Dung Nguyen, Hung Le, Huu Hiep Nguyen, Minh Hoang Nguyen.

Figure 1
Figure 1. Figure 1: Effective rank of the text branch per backbone. We term this failure mode text collapse: the text branch of a multimodal forecasting model degenerates such that it produces nearly the same output regardless of the in￾put text. Text collapse is not a data problem: we observe it consistently across all domains of Time-MMD [10], a benchmark that pairs time series with high-quality, expert￾curated reports, rul… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of REST-TS. The Online Backbone (bottom, blue) is trained with backprop￾agation and fuses its numerical latent with the text representation ZS to produce the joint forecast Yjoint. The Target Network (top, blue) is a momentum-updated EMA copy that produces a stable unimodal forecast Yts, defining the residual R = Y−Yts. The Text Branch (green) encodes S with a frozen language model and feeds the r… view at source ↗
Figure 3
Figure 3. Figure 3: Robustness to text-branch learning rate lr2. REST-TS (blue) maintains stable effective rank, CKA, and MSE across all settings, while MMTSF (purple) and TaTS (orange) show higher variance, particularly on forecasting error [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effective rank of text-branch output representations ( [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: CKA between text-branch representations and forecasting targets ( [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Text-to-numerical gradient norm ratio ∥∇text∥/∥∇num∥ during training across all nine Time-MMD domains with iTransformer. MMTSF (purple) collapses to near-zero text gradients across every domain. REST-TS (blue) consistently maintains near-parity, demonstrating that exclusive residual supervision prevents gradient suppression regardless of domain. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Text-to-numerical gradient norm ratio during training across all nine Time-MMD domains [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity to text-branch learning rate [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sensitivity to text-branch learning rate [PITH_FULL_IMAGE:figures/full_fig_p032_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sensitivity to text-branch learning rate [PITH_FULL_IMAGE:figures/full_fig_p033_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prediction divergence δ = MSE(yˆreal, yˆno_info)/MSEreal across all nine Time-MMD domains with Informer. A higher δ indicates greater sensitivity of the prediction to the text input. REST-TS (green) consistently achieves the highest divergence, while MMTSF (blue) and TaTS (orange) show near-zero values in several domains, confirming text collapse. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prediction divergence δ across all nine Time-MMD domains with PatchTST. TaTS (orange) yields exactly zero divergence in most domains, indicating that its predictions are entirely insensitive to the text input. REST-TS (green) maintains the highest divergence in every configuration, confirming that the result is backbone-agnostic. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Sensitivity of REST-TS MAE to the moving-average trend-kernel width on all nine [PITH_FULL_IMAGE:figures/full_fig_p036_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Sensitivity of REST-TS MAE to the Fourier component count [PITH_FULL_IMAGE:figures/full_fig_p037_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Prediction comparison across all nine Time-MMD domains with iTransformer. [PITH_FULL_IMAGE:figures/full_fig_p038_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Prediction comparison across all nine Time-MMD domains with PatchTST. [PITH_FULL_IMAGE:figures/full_fig_p038_16.png] view at source ↗
read the original abstract

Multimodal time series forecasting, which pairs numerical sequences with domain-relevant textual reports, promises to inject world knowledge into forecasting pipelines. However, we uncover a critical failure mode in existing frameworks that we term text collapse: the text branch converges to a content-independent transformation, contributing negligible discriminative signal regardless of the input description. We argue that text collapse is a consequence of a fundamental asymmetry in time series forecasting: the numerical input is strongly autocorrelated with the output, making the numerical backbone inherently dominant, while the text branch, despite carrying complementary and often critical information, is insufficiently utilized, leading to its systematic underexploitation. To address this, we propose \textbf{REST-TS} (\textbf{R}esidual-\textbf{E}xclusive \textbf{S}upervision for \textbf{T}ext in \textbf{T}ime \textbf{S}eries), which turns the asymmetry into a design principle: the numerical backbone produces its own independent numerical forecast, and the text branch is exclusively supervised to predict the structured components of the residual, the prediction gap that numbers cannot explain. Because no numerical pathway can reduce these losses, the text branch must extract genuine content from the input description. Evaluated across diverse real-world domains and backbone architectures, REST-TS achieves state-of-the-art performance and consistently demonstrates greater text-branch utilization than existing frameworks, providing strong empirical evidence that supervising the text branch on the residual compels it to extract genuine content from the input.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims that multimodal time series forecasting suffers from 'text collapse,' where the text branch converges to a content-independent transformation due to the dominance of autocorrelated numerical inputs. It proposes REST-TS, which uses residual-exclusive supervision: the numerical backbone makes an independent forecast, and the text branch is supervised only on the residual components that numbers cannot explain, forcing it to extract genuine content from text. Experiments across domains and backbones show SOTA performance and greater text utilization.

Significance. If the results hold, this work identifies a previously unrecognized failure mode in multimodal forecasting and offers a design principle to resolve it by leveraging the inherent asymmetry. The cross-domain and cross-backbone evaluation, along with demonstrations of improved text-branch utilization, provides empirical support. Strengths include the focus on a falsifiable prediction about text utilization.

major comments (1)
  1. [Abstract (REST-TS design principle)] Abstract (paragraph describing REST-TS design principle): The claim that 'Because no numerical pathway can reduce these losses, the text branch must extract genuine content from the input description' is load-bearing for the central contribution but rests on an assumption of complete independence between branches. If the architecture permits any information flow (concatenated features, cross-attention, shared layers, or joint optimization), gradients from the combined output could allow the text branch to collapse while the numerical backbone absorbs adjustments. The manuscript must provide explicit equations or a diagram (likely in §3) confirming no such pathway exists during training.
minor comments (2)
  1. [§3] The term 'structured components of the residual' is used without a precise definition or equation; add a formal decomposition (e.g., in §3.2) to clarify what the text branch is supervised to predict.
  2. [Experiments section] Ensure all tables reporting SOTA results include statistical significance tests and full baseline details for reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need to rigorously substantiate the independence assumption underlying the REST-TS design principle. The concern is well-taken, and we will strengthen the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract (REST-TS design principle)] Abstract (paragraph describing REST-TS design principle): The claim that 'Because no numerical pathway can reduce these losses, the text branch must extract genuine content from the input description' is load-bearing for the central contribution but rests on an assumption of complete independence between branches. If the architecture permits any information flow (concatenated features, cross-attention, shared layers, or joint optimization), gradients from the combined output could allow the text branch to collapse while the numerical backbone absorbs adjustments. The manuscript must provide explicit equations or a diagram (likely in §3) confirming no such pathway exists during training.

    Authors: We agree that the independence claim is central and requires explicit architectural clarification. In REST-TS, the numerical backbone is trained to convergence on its own forecast loss using only the numerical input; the text branch receives gradients exclusively from a separate residual loss computed on the prediction gap after the numerical forecast is fixed. No shared layers, cross-attention, or feature concatenation occurs between the branches during this residual-supervision stage, and the numerical backbone parameters are not updated by the residual loss. To make this fully transparent, we will add (i) a detailed diagram in §3 illustrating the two-stage training flow with separate loss paths and (ii) the corresponding equations showing the numerical loss L_num and the residual loss L_res with their respective gradient flows. This revision will directly address the possibility of indirect information leakage. revision: yes

Circularity Check

0 steps flagged

No significant circularity; REST-TS is an explicit design objective motivated by observed asymmetry

full rationale

The paper identifies text collapse empirically, attributes it to numerical dominance in autocorrelated series, and defines REST-TS as residual-exclusive supervision on the text branch. The statement that this forces genuine content extraction follows directly from the supervision choice by construction and is presented as the intended mechanism, not as a first-principles prediction or fitted result that reduces to the inputs. No self-citations, uniqueness theorems, ansatzes smuggled via prior work, or renamings of known patterns appear in the abstract or described derivation. The central claim remains an implemented training principle evaluated on external data rather than a self-referential reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the domain assumption of numerical autocorrelation dominance and introduces two new concepts (text collapse and REST-TS) without external validation or independent evidence in the abstract.

axioms (1)
  • domain assumption Numerical input is strongly autocorrelated with the output in time series forecasting tasks.
    Invoked in the abstract as the root cause of text collapse and numerical dominance.
invented entities (2)
  • text collapse no independent evidence
    purpose: To describe the failure mode in which the text branch converges to a content-independent transformation.
    Newly coined phenomenon introduced to frame the problem.
  • REST-TS no independent evidence
    purpose: The proposed training procedure that supervises text exclusively on residuals.
    New method introduced to resolve the identified issue.

pith-pipeline@v0.9.1-grok · 5805 in / 1280 out tokens · 31349 ms · 2026-06-26T21:16:56.047395+00:00 · methodology

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

29 extracted references · 4 canonical work pages

  1. [1]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750–15758, June 2021

    2021

  2. [2]

    AutoWebGLM:

    Zihan Dong, Xinyu Fan, and Zhiyuan Peng. Fnspid: A comprehensive financial news dataset in time series. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, page 4918–4927, New York, NY , USA, 2024. Association for Computing Machinery. ISBN 9798400704901. doi: 10.1145/3637528.3671629. URL https: //doi.org/10.114...

    work page doi:10.1145/3637528.3671629 2024

  3. [3]

    Pmr: Prototypical modal rebalance for multimodal learning

    Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. Pmr: Prototypical modal rebalance for multimodal learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20029–20038, June 2023

    2023

  4. [4]

    Classifier-guided gradient modulation for enhanced multimodal learning

    Zirun Guo, Tao Jin, Jingyuan Chen, and Zhou Zhao. Classifier-guided gradient modulation for enhanced multimodal learning. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URLhttps://openreview.net/forum?id=oe5ZEqTOaz

    2024

  5. [5]

    Modality competition: What makes joint training of multi-modal network fail in deep learning? (Provably)

    Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (Provably). In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors,Proceedings of the 39th International Conference on Machine Learning, volume 162...

    2022

  6. [6]

    Multi-modal time series analysis: A tutorial and survey

    Yushan Jiang, Kanghui Ning, Zijie Pan, Xuyang Shen, Jingchao Ni, Wenchao Yu, Anderson Schneider, Haifeng Chen, Yuriy Nevmyvaka, and Dongjin Song. Multi-modal time series analysis: A tutorial and survey. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .2, KDD ’25, page 6043–6053, New York, NY , USA,

  7. [7]

    ISBN 9798400714542

    Association for Computing Machinery. ISBN 9798400714542. doi: 10.1145/3711896. 3736567. URLhttps://doi.org/10.1145/3711896.3736567

    work page doi:10.1145/3711896

  8. [8]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations (ICLR), 2015. URL https://arxiv.org/ abs/1412.6980

    Pith/arXiv arXiv 2015

  9. [9]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 3519–3529. PMLR, 09–15 Jun 2019. URL https://...

    2019

  10. [10]

    Language in the flow of time: Time- series-paired texts weaved into a unified temporal narrative

    Zihao Li, Xiao Lin, Zhining Liu, Jiaru Zou, Ziwei Wu, Lecheng Zheng, Dongqi Fu, Yada Zhu, Hendrik Hamann, Hanghang Tong, and Jingrui He. Language in the flow of time: Time- series-paired texts weaved into a unified temporal narrative. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum? id=a1zBg9cBvt

    2026

  11. [11]

    Sasanur, Megha Sharma, Jiaming Cui, Qingsong Wen, Chao Zhang, and B

    Haoxin Liu, Shangqing Xu, Zhiyuan Zhao, Lingkai Kong, Harshavardhan Kamarthi, Aditya B. Sasanur, Megha Sharma, Jiaming Cui, Qingsong Wen, Chao Zhang, and B. Aditya Prakash. Time-MMD: Multi-domain multimodal dataset for time series analysis. InThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. URLhttps...

    2024

  12. [12]

    Time- bridge: Non-stationarity matters for long-term time series forecasting.International Conference on Machine Learning, 2025

    Peiyuan Liu, Beiliang Wu, Yifan Hu, Naiqi Li, Tao Dai, Jigang Bao, and Shu-Tao Xia. Time- bridge: Non-stationarity matters for long-term time series forecasting.International Conference on Machine Learning, 2025

    2025

  13. [13]

    itransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview. net/forum?id=JePfAI8fah. 10

    2024

  14. [14]

    Are multimodal transformers robust to missing modality? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18177–18186, June 2022

    Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18177–18186, June 2022

    2022

  15. [15]

    Spectral text fusion: A frequency-aware approach to multimodal time-series forecasting

    Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, and Hung Le. Spectral text fusion: A frequency-aware approach to multimodal time-series forecasting. InThe 29th International Conference on Artificial Intelligence and Statistics, 2026. URL https://openreview.net/ forum?id=w4yjJc06zt

    2026

  16. [16]

    A time series is worth 64 words: Long-term forecasting with transformers

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InThe Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum? id=Jbdc0vTOcol

    2023

  17. [17]

    Balanced multimodal learning via on-the-fly gradient modulation

    Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  18. [18]

    Language models are unsupervised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019

    2019

  19. [19]

    The effective rank: A measure of effective dimensionality

    Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European Signal Processing Conference, pages 606–610, 2007

    2007

  20. [20]

    Timemixer: Decomposable multiscale mixing for time series forecasting

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and JUN ZHOU. Timemixer: Decomposable multiscale mixing for time series forecasting. In International Conference on Learning Representations (ICLR), 2024

    2024

  21. [21]

    Tailornet: Predict- ing clothing in 3d as a function of human pose, shape and garment style

    Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12692–12702, 2020. doi: 10.1109/CVPR42600.2020.01271

    work page doi:10.1109/cvpr42600.2020.01271 2020

  22. [22]

    From news to forecast: Integrating event analysis in LLM-based time series forecasting with reflection

    Xinlei Wang, Maike Feng, Jing Qiu, Jinjin Gu, and Junhua Zhao. From news to forecast: Integrating event analysis in LLM-based time series forecasting with reflection. InThe Thirty- eighth Annual Conference on Neural Information Processing Systems, 2024. URL https: //openreview.net/forum?id=tj8nsfxi5r

    2024

  23. [23]

    Mmpareto: boosting multimodal learning with innocent unimodal assistance

    Yake Wei and Di Hu. Mmpareto: boosting multimodal learning with innocent unimodal assistance. InInternational Conference on Machine Learning, 2024

    2024

  24. [24]

    Autoformer: Decomposition transformers with Auto-Correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with Auto-Correlation for long-term series forecasting. InAdvances in Neural Information Processing Systems, 2021

    2021

  25. [25]

    11121–11128

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting?Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):11121–11128, Jun. 2023. doi: 10.1609/aaai.v37i9.26317

    work page doi:10.1609/aaai.v37i9.26317 2023

  26. [26]

    Multimodal representation learning by alternating unimodal adaptation

    Xiaohui Zhang, Jaehong Yoon, Mohit Bansal, and Huaxiu Yao. Multimodal representation learning by alternating unimodal adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27456–27466, June 2024

    2024

  27. [27]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting

    Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum?id=vSVLM2j9eie

    2023

  28. [28]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In The Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Virtual Conference, volume 35, pages 11106–11115. AAAI Press, 2021

    2021

  29. [29]

    FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. InProc. 39th International Conference on Machine Learning (ICML 2022), 2022. 11 Appendix for Does Text Actually Help? Uncovering and Resolving Text Collapse in Multimodal Time Series Forecasting A Im...

    arXiv 2022