acceptodds
Withdrawn from ICLR 2027

Collaborative-enhanced Multi-view Learning for Noisy Multivariate Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multivariate time series forecasting fundamentally relies on cross-channel dependencies. However, the presence of noise may distort such dependencies, posing significant challenges to reliable relation modeling. Existing methods either model temporal patterns independently for each channel or capture cross-channel dependencies within a single shared space. The former overlooks informative channel interactions, while the latter is susceptible to noise-distorted dependencies. As a result, neither is able to fully exploit reliable and complementary relational information, which constrains their forecasting robustness under noisy conditions. To address this challenge, we propose a collaborative-enhanced multi-view framework that constructs multiple overlapping latent relation views through adaptive filtering and dynamic soft routing. Bidirectional view interaction enables mutual guidance and collaborative enhancement by leveraging complementary relational information, reducing reliance on potentially distorted dependencies. Extensive experiments demonstrate improved forecasting accuracy and stable performance under Constant, Missing, and Gaussian noise.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Noise, multi-view, co-training

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ivVxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a collaborative-enhanced multi-view framework for robust multivariate time series forecasting under noisy conditions. The key idea is that single-space relation modeling is vulnerable to corrupted cross-channel dependencies when input data is noisy. To address this, the authors construct multiple overlapping latent relation views through adaptive filtering and dynamic soft routing, enabling diverse cross-channel dependencies to be modeled beyond a single shared relation space. A collaborative enhancement mechanism using bidirectional state-space models enables mutual guidance among views.

Critical strengths

None

Critical weaknesses
  1. The paper claims that multiple overlapping latent views capture complementary relational evidence, but provides no quantitative analysis of view diversity. Do different views learn distinct channel relationships? Is there a risk of view collapse? Without such analysis, the central claim of "complementary evidence" remains empirically unverified.
  2. The design of dynamic soft routing is clearly lacking in explainability. Why is it that this approach can capture the relationships between channels? Furthermore, why is bidirectional interaction effective? These lack both theory and intuition, and have not been further analyzed in experiments.
  3. Baselines do not represent the strongest robust forecasting methods. The paper does not compare against recent robust-specific methods such as CrossGNN, Crossformer, or other robust channel-dependent models.
Questions for authors
  1. How many latent views KK are used in each experiment, and how sensitive is the method to KK?
  2. Do the learned routing assignments actually produce meaningful overlapping views? Can you visualize assignment matrices or provide entropy/overlap statistics?
  3. Why were methods such as CrossGNN, CARD, iTransformer, Crossformer, or other robust channel-dependent models not included?
Reviewer_wZNMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

To address the issue of distorted cross-channel dependencies in the forecasting of noisy multivariate time series, this paper proposes a collaboratively enhanced multi-view framework. The core idea is to avoid compressing all channel relationships into a single shared space; instead, the framework constructs multiple overlapping latent relationship views through adaptive filtering and dynamic soft routing, and subsequently facilitates information exchange and collaborative enhancement across these views via a bidirectional state-space module.

Critical strengths

1、The method described in this article is clear and easy to understand.

Critical weaknesses

1、 Selecting the "best" checkpoint using the test dataset introduces the risk of information leakage. Even if the authors do not designate a validation set, the test set should not be used to determine the optimal epoch. Doing so leads to an overly optimistic report of the "best" results and renders any comparison with the "last" checkpoint results unfair. Overall, this evaluation approach lacks practical utility; while it may employ the same selection strategy as the baseline, the baseline itself might not be well-suited to this paradigm—it is possible that a superior model could have been identified for the baseline had a validation set been used instead.

2、 Key implementation details are missing, hindering reproducibility. The paper fails to explicitly specify the following hyperparameters and module details:What specific statistical measures are included in the channel-wise statistical features scs_c used for routing? How is the number of views KK determined, and what is its default value? What is the architecture of the relation encoder for each view—is it a Transformer, an attention mechanism, or an MLP? How are the adapter tokens initialized? Furthermore, what are the specific implementation details of the Bi-SSM, including its state dimension and whether it is causal?

3、 The validity of the training objective in this paper is questionable. During the training phase, both the input and the output are subjected to noise, whereas in the testing phase, only the input is corrupted. In essence, this means that during training, the model is fitting the noise distribution rather than the original data distribution. Such a design is highly unreasonable, rendering the evaluation metrics entirely meaningless.

4、 First, adding noise to the input training data only to subsequently remove it using an AF (adaptive filtering) module seems redundant. Second, the suitability of adaptive filtering for various types of noise is questionable; the method employs a learnable low-pass mask while preserving the zero-frequency component. However, this approach has drawbacks: "constant" noise often manifests as a low-frequency bias or even a zero-frequency component, meaning the preservation of the zero-frequency component may hinder the removal of such noise. "Missing" noise creates gaps in the time domain, and applying a direct Fourier transform may introduce artifacts. Finally, while low-pass filtering might be effective against Gaussian noise—which acts as a disturbance across the entire frequency spectrum—it risks the loss of authentic high-frequency dynamics.

5、 Multi-view representations may degenerate into a single space, lacking empirical validation of their diversity. In the formulation Vk=Diag(a:,k)EV_k = \text{Diag}(a_{:,k})E, each view retains all channel tokens, merely applying different reweighting. If the routing assignments AA are too similar across different views, the views will exhibit significant overlap—potentially collapsing into a single, unified space. Consequently, the notion of "multi-view complementarity" currently serves more as a design motivation than a concept backed by empirical evidence. Notably, this paper lacks ablation studies on the number of views (KK) and provides no metrics to assess whether these multiple views truly capture distinct perspectives.

6、 Bi-SSM processes data along the view dimension; however, views lack a natural ordering.The views V1,…,VKV_1, \dots, V_K represent latent relational perspectives and possess no inherent fixed sequence. Since Bi-SSM is a sequence-sensitive module, its performance may be unstable depending on the arrangement of these views.

7、 Insufficient baselines and a lack of robust TSF (Time Series Forecasting) methods.The paper only compares its approach against PatchTST, TimeXer, and Dynamic-TMoE—methods that were not originally designed for multivariate forecasting tasks involving noise. Furthermore, although the authors claim to follow the evaluation protocol used for the "RobustTSF" model, they fail to include a performance comparison against that very model; the rationale behind this design choice remains unclear. Consequently, the absence of these relevant baselines undermines the claim of "superior performance."

Questions for authors

See Weaknesses.

Reviewer_DZA9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets robust multivariate forecasting under corrupted data. A learnable low-pass filter comes first. A prototype-based soft router then assigns channel tokens to K overlapping "views", and each view has an attention encoder with an adapter token. A bidirectional SSM over the K adapter tokens exchanges information between views, followed by gated enhancement and fusion. Training uses corrupted data (constant/missing/Gaussian, η_train=0.5) and testing uses noisy inputs (η_test 0–0.3). Experiments cover ETTh1, ETTh2, Exchange and ILI against PatchTST, TimeXer and Dynamic-TMoE.

Critical strengths
  1. The problem is relevant: noise can distort cross-channel dependencies (Fig. 1a), which hurts channel-dependent models.
  2. The evaluation extends RobustTSF to the multivariate setting with three corruption types and several severities. Training inputs and targets are corrupted while scoring is on clean test targets, which is a reasonable protocol.
  3. Gains on ILI are clear (e.g., 13–35% lower MSE under missing noise), and the gap between best and last checkpoints is small.
Critical weaknesses
  1. The reported results rely on test-set model selection. No validation set is used, and the "Best" results are taken from the epoch with the lowest test error, with a single random seed and no standard deviations. On clean data the method is also often outperformed, for example by PatchTST on ETTh2, 0.242 versus 0.230, and on Exchange, 0.241 versus 0.222.
  2. The experimental scope does not match the motivation. The method is built around multiple cross-channel relation views, yet all four datasets have only 7 or 8 channels. Only three baselines are compared and none is designed for noisy data, although robust methods such as RobustTSF, CrossGNN, Merlin and LGA are cited in the paper. Simple denoise-then-forecast baselines are also missing.
  3. The proposed mechanism is not shown to work as described. No analysis shows that the views capture complementary or more reliable dependencies, and the Bi-SSM imposes an order on views that have no natural sequence. The ablation effects are small, and some ablated variants even outperform the full model, for example on Exchange with constant noise. Key settings such as the number of views and the model size are not reported.
Questions for authors

Please see weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.