acceptodds
Under review as a conference paper at ICLR 2027

FACETS: Selective Cross-Band Supervision for Robust Time-Series Representation Learning

Abstract

Noise in time-series self-supervised learning can corrupt not only observations but also the supervisory relation used to learn from them. When nuisance is predictable, reconstructable, or preserved across views, standard pretext objectives may consistently favor such nuisance patterns, causing them to become readily decodable from the learned representation. We call this failure mode supervisory contamination. FACETS treats robustness as a problem of supervision design: it makes selective agreement among frequency components of the same sequence the primary learning signal, down-weights weakly supported components via detached relative-regularity estimates, discounts batch-shared drift through directional rectification, and uses temporal prediction only as an auxiliary local constraint. We characterize when component agreement can improve the deployed original-view encoder by separating observed cross-band dependence from nuisance-mediated dependence and the component-to-deployment gap. A controlled diagnostic provides mechanistic evidence: prediction-only supervision makes injected interference highly decodable, whereas cross-band-only supervision sharply reduces nuisance decodability, at some cost in semantic accuracy. Across five biomedical and industrial datasets, FACETS improves 10% label linear-evaluation accuracy by 2.07–4.44 percentage points over the strongest dataset-specific SSL baselines, achieves the best weighted F1 on all five datasets, leads all six ECG transfer directions, and remains strongest under increasing Gaussian corruption. Together, these results show that robust self-supervised learning depends not merely on producing cleaner views, but on controlling which relations are trusted to organize the representation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, time-series representation learning, robust representation learning, contrastive learning, multiscale learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_yp2AOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a time-series self-supervised learning method. The authors' motivation is that the time-series can be noisy and current pretext tasks can favor the noisy part and learn that. The proposed method, FACETS, learns the representation primarily by the agreement among frequency components of the same sequence. Authors conducted experiments on five datasets and showed that the proposed method increases the linear probe performance by 10%.

Critical strengths

The overall motivation and criticism against the forecasting and reconstruction based SSL methods are valid. The drawbacks of these techniques are well-motivated. The evaluated datasets also fit the purpose to show the claims, for example Sleep-EDF is noisy. And, the SSL pretext task should learn the features that are not noisy but informative for the sleep stage classification.

Critical weaknesses

The paper falls down after the motivation. The presentation of the method is not clear. The evaluation and comparison are also not extensive. The authors do not compare with one of the recent works in the same direction [1]. The datasets can also be extended. The propositions and theory part of the work are very hard to follow. The notations are not clear and some of them are never explained.

[1] Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections. NeurIPS 2025.

The ablations are also very weak. The critical choices of wavelet design is not motivated. There is so much whitespace in the main manuscript while most figures can be presented better. The authors did not do a good job for presenting their evidence.

I believe the current version of the work needs a major revision. The experiments with baselines and ablations should be extended.

Questions for authors

1-How do the authors choose the wavelet parameters? Are they application specific? Are the chosen hyperparameters robust? How much the performance changes with the chosen number of wavelet level?

Reviewer_jq7mOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FACETS, a time-series self-supervised framework that treats frequency components as candidate supervisory relations rather than merely alternative inputs. It combines selective cross-band contrast with detached relative-regularity weighting and directional rectification to control nuisance-mediated agreement, while frequency-guided temporal prediction serves as an auxiliary local constraint. Experiments across multiple datasets evaluate label efficiency, cross-domain transfer, and robustness to corruption.

Critical strengths
  1. This paper delves into the challenge of deciding which supervisory relations should be trusted, which is cinceptually useful.
  2. The clear correspondence between challenged and proposed modules makes the introduction of the method easy to follow.
  3. The broad experimental setting makes the empirical validation of the proposed method quite comprehensive.
Critical weaknesses
  1. Will the 40/20/40 split instead of entity-disjoint split cause information leak, as the same subject may appear in both training and test sets?
  2. Removing a single mini-batch mean direction may fail under biased mini batch. Additional stress tests and a discussion of failure boundaries would strengthen the method.
  3. I cannot find the codes. This in a sense raises concerns about result reliability.
Questions for authors

Please refer to weaknesses.

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

This paper proposes FACETS, a self-supervised learning framework for robust time-series representation learning. The method uses selective cross-band consistency, reliability-based weighting, and directional rectification to reduce the influence of nuisance information during pretraining. An auxiliary temporal prediction objective is also introduced. Experiments on five biomedical and industrial datasets demonstrate improvements in classification, cross-domain transfer, and noise robustness.

Critical strengths
  1. The paper addresses an interesting problem by considering how noisy supervisory relations affect self-supervised representation learning. The motivation is clear, and the proposed framework is reasonably well designed.

  2. The experimental evaluation is extensive. FACETS achieves consistent improvements across five datasets, and the ablation studies provide useful evidence for the contributions of individual components.

Critical weaknesses
  1. Limited generalization evidence. The experiments use sample-level data splits rather than patient- or entity-disjoint splits. This may introduce dependencies between training and test samples, particularly for biomedical datasets, and limits the strength of the generalization claims.

  2. Insufficient validation of the proposed mechanism. The supervisory contamination analysis is based on a single-seed experiment. Moreover, FACETS still retains substantial nuisance information, making it difficult to establish how effectively the proposed mechanisms suppress nuisance-related representations.

  3. Limited theoretical and robustness guarantees. The theoretical results rely on assumptions about component reliability and do not directly guarantee downstream improvements. In addition, the main robustness experiments introduce noise only at test time, leaving robustness under noisy pretraining insufficiently evaluated.

Questions for authors
  1. Could the authors evaluate FACETS using patient- or entity-disjoint splits, particularly on the ECG and EEG datasets, to better assess its generalization?

  2. Could the supervisory contamination experiment be repeated with multiple random seeds and different nuisance levels? How consistently does FACETS reduce nuisance decodability?

  3. How sensitive is the proposed reliability weighting to structured or predictable noise that can also be reconstructed accurately?

  4. Does FACETS remain effective when noise is introduced during self-supervised pretraining, rather than only during downstream testing?

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

This paper proposes FACETS, a self-supervised representation learning framework for robust time-series classification. The authors introduce the concept of supervisory contamination, arguing that conventional self-supervised objectives may unintentionally encourage the encoder to preserve nuisance patterns that are predictable, reconstructable, or shared across views.

To address this issue, FACETS constructs supervisory relations across wavelet frequency bands rather than relying solely on augmented views. The framework combines similarity-based positive pair selection, reconstruction-based band weighting, and directional rectification to reduce potentially unreliable cross-band agreement. An auxiliary temporal prediction objective is also introduced to preserve local temporal information. Experiments on five biomedical and industrial time-series datasets show improvements in downstream classification, cross-domain transfer, and robustness under synthetic corruption.

Critical strengths
  1. An interesting perspective on robust self-supervised learning. The paper focuses on the reliability of supervisory relations rather than only improving input quality through denoising or augmentation. I find this perspective meaningful, as nuisance patterns can indeed become part of the learning signal in self-supervised objectives.

  2. Consistent improvements across multiple datasets. FACETS achieves competitive results on five datasets, and the ablation study shows that the proposed components contribute to downstream classification performance. Another practical advantage is that the additional frequency decomposition and auxiliary modules are only required during pretraining.

Critical weaknesses
  1. The assumptions underlying the proposed supervisory mechanisms are insufficiently validated. In Eq. (4), cross-band positive pairs are selected based on representation similarity. However, high similarity does not necessarily indicate reliable task-relevant agreement, as it may also arise from shared nuisance factors or subject-specific characteristics. Conversely, frequency bands containing complementary discriminative information may exhibit limited similarity. Although selective pairing avoids indiscriminate cross-band alignment, the paper does not directly establish whether the selected pairs provide more reliable semantic supervision than the discarded ones. Similarly, directional rectification in Eq. (12) uses the mini-batch mean embedding to estimate a shared direction. The extent to which this direction represents nuisance information remains unclear, as it may also contain class-discriminative information. Moreover, semantic and nuisance components need not be orthogonal, potentially causing the rectification to suppress task-relevant information. More direct empirical validation is needed to establish whether cross-band pair selection improves supervisory reliability and whether directional rectification suppresses nuisance-related information while preserving discriminative representations.

  2. The experimental evidence is insufficient to attribute the performance gains to reduced supervisory contamination. Although the main results and ablation studies demonstrate improvements in downstream classification, they do not directly establish that these gains result from mitigating nuisance-related supervision. The nuisance-decoding analysis in Section 4.3 and Appendix D.2 is relevant to this claim but lacks comparisons with existing approaches addressing unreliable contrastive supervision, such as DBPM [1]. Comparisons under matched experimental protocols, potentially including related wavelet-based time-frequency methods [2], would provide stronger evidence. Furthermore, Table 9 omits FACETS, preventing a direct assessment of whether directional rectification reduces nuisance decodability. The projection in Eq. (12) and the similarity adjustment in Eq. (13) are not independently evaluated, and the nuisance-decoding analysis is based on a single random seed. Consequently, the contribution of the proposed mechanisms to nuisance suppression remains insufficiently established.

[1] Lan, X., Yan, H., Hong, S., and Feng, M. (2024). Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach. International Conference on Learning Representations (ICLR). https://proceedings.iclr.cc/paper_files/paper/2024/hash/2b04d138128b76c33e30ef73b771d0f4-Abstract-Conference.html

[2] Yang, C., Chen, Y., Li, Z., and Guo, Z. (2026). Contrastive Representation Learning for Multivariate Time Series Classification Based on Temporal-Frequency Consistency. Applied Soft Computing, 195, 115073. https://doi.org/10.1016/j.asoc.2026.115073

Questions for authors
  1. Reliability of cross-band supervision. How does the similarity-based selection in Eq. (4) distinguish task-relevant cross-band agreement from similarity induced by shared nuisance factors? Can the authors provide empirical evidence comparing the semantic relevance of selected and discarded frequency-band pairs?

  2. Validity of directional rectification. To what extent does the mini-batch mean direction in Eq. (12) capture nuisance rather than task-relevant information? Can the authors examine its relationship with semantic and nuisance representations and independently evaluate the effects of Eqs. (12) and (13)?

  3. Mechanism-level validation. Can the authors extend the nuisance-decoding analysis in Table 9 to include FACETS and a relevant baseline such as DBPM under matched noisy-pretraining conditions? Results for both semantic and nuisance decodability across multiple random seeds would help substantiate the claimed mitigation of supervisory contamination.

  4. Evaluation splits. The experiments adopt the sample-level 40/20/40 partitions from DECL, ensuring consistency with the reported baselines but not establishing subject-independent generalization. For physiological datasets, subject-specific information shared across partitions may influence performance. The conclusions should therefore be qualified according to the evaluation protocol.

  5. Cross-domain evaluation protocol. Section 4.4 fine-tunes the pretrained model using 10% labeled target-domain training data, which constitutes supervised cross-domain transfer rather than strict domain generalization. Furthermore, four of the six transfer directions in Table 3 show improvements below one percentage point, without reported standard deviations, making the significance of these gains difficult to assess.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.