acceptodds
Under review as a conference paper at ICLR 2027

Contrastive Learning of Data-Quality-Invariant Representations for Time Series Classification

Abstract

Data quality issues are pervasive in time series collected from real-world sensing systems. Such issues can induce distribution shift and degrade downstream task performance. In this paper, we propose a self-supervised contrastive learning framework to learn data-quality-invariant representations for time series classification. First, we propose anomaly injection to generate negative samples for contrast with normal samples in the latent space, thereby mitigating data-quality-induced distribution shift. Second, we use autoencoder-based reconstruction as the pretext task to learn discriminative representations. Lastly, we freeze the pretrained backbone and employ a prototypical network for few-shot time series classification using a small number of labeled samples. Extensive experiments on real-world datasets demonstrate that our proposed model consistently outperforms state-of-the-art self-supervised contrastive learning frameworks, improving accuracy by an average of 4.09 percentage points across the three datasets compared with the strongest baseline on each dataset. An ablation study also shows that anomaly injection is effective at improving classification performance compared with current data augmentation techniques.

Primary area:
learning on time series and dynamical systems
Keywords:
Contrastive Learning, Time-Series Classification, Data-Quality Robustness, Self-Supervised Representation Learning, Few-Shot Learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_9wbsOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses time-series classification under data-quality-induced distribution shifts and limited labels. It proposes a two-stage framework that combines anomaly injection, reconstruction-based contrastive learning, and prototypical few-shot classification. Experiments on Steel Frame, IASC-ASCE, and Z24 datasets show consistent improvements over ten self-supervised baselines, supported by ablation study and t-SNE analysis.

Critical strengths
  1. Representation learning on data with quality issues is a real need.

  2. The balance between contrastive learning and the reconstruction information loss is plausible and reasonable.

Critical weaknesses
  1. The writing of the Related Work is quite confusing. "self-supervised time series representation" (subsection 2.1) and "contrastive representation learning" (subsection 2.2) are somewhat overlapping. For example, TNC and T-Loss in 2.1 also fail to the category of contrastive representation learning, while TS2Vec can also be seen as a self-supervised time series representation method. This will cause readers outside the relevant literature to be quite confused.

  2. There are quite a few places where conflicts arise between one another. Please refer to Questions 1 to 3.

  3. The Ablation is too rough. The authors should also prove the effectiveness of their detailed designs, like positive/negative pair selection strategy, anomaly injection, etc.

Questions for authors
  1. In subsection 3.3, the phrase "neighbouring windows" conflicts with the following formula whose meaning is to choose 'all' windows except j, but not limited to neighbouring ones.

  2. In subsection 3.3, the author mentioned "This ensures that samples affected by data-quality-induced distribution shifts yield representations similar to those of their original classes.". However, if the original time-series and the anomaly-injected ones were considered as negative pairs, the result should be as "This ensures that samples affected by data-quality-induced distribution shifts yield representations dissimilar to those of their original classes.", as anomaly injection transforming the original series to low-quality ones.

  3. In subsection 4.1, the samples are categorized into four states. But if the aim is to get a data-quality-invariant representation, as the title of this paper suggests, there should only be two categories: healthy or damaged.

  4. If a shift occurs in the neighbouring window, using them as positive pairs will not be reasonable. This has been discussed in detail in TS2Vec. Can you give me the rationale behind your positive pair selection strategy?

  5. In the supplemental material, I can only see the Appendix. Can I have a look at the code? This will strengthen the reproducibility and will make the experimental details clearer to reviewers.

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

The paper proposes a two-stage framework combining reconstruction and contrastive learning with anomaly-injected negatives, followed by frozen-encoder prototypical classification. Experiments on three structural vibration datasets report an average accuracy improvement of 4.09 percentage points over the strongest baseline per dataset.

Critical strengths
  1. Practical relevance. The paper addresses structural monitoring under sensor disturbances and limited labeled data.

  2. Empirical improvements. Results improve over the reported baselines across three datasets, with ablations examining reconstruction, anomaly injection, and backbone choice.

Critical weaknesses
  1. Unsubstantiated invariance claim. Equation (3) explicitly treats corrupted counterparts as negatives. Without demonstrating a separation of structural and quality information, this objective does not establish the claimed data-quality invariance.

  2. Evaluation does not validate the central claim. The four-class task distinguishes both structural health and sensor quality, rewarding sensitivity to quality differences. Missing clean-to-corrupted health-state evaluations leave the claimed robustness unverified.

  3. Unclear attribution of improvements. Backbone, training-budget, and downstream-classifier matching across baselines is insufficiently documented. The substantial backbone effect in Table 2 makes it difficult to isolate gains attributable to the proposed learning strategy.

  4. Insufficient evaluation reliability. Results use one reported seed without variability across five-shot support-set selections. Record-level separation between pretraining, support, and test data is also unclear despite overlapping windows, limiting confidence in generalization.

Questions for authors

See the weaknesses.

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

This paper proposes a self-supervised representation learning framework to improve time-series classification performance under data-quality-induced distribution shifts caused by sensor data degradation.

The proposed method adopts a two-stage learning framework that combines anomaly injection, contrastive learning, and autoencoder-based reconstruction. During the pretraining stage, different segments from the same time series are used as positive samples, while negative samples are generated by artificially injecting five types of anomalies (Outlier, Drift, Bias, Minor, and Noise). The model is trained by jointly optimizing the InfoNCE loss and reconstruction loss.

This paper proposes a self-supervised representation learning framework to improve time-series classification performance under data-quality-induced distribution shifts caused by sensor data degradation.

The proposed method adopts a two-stage learning framework that combines anomaly injection, contrastive learning, and autoencoder-based reconstruction. During the pretraining stage, different segments from the same time series are used as positive samples, while negative samples are generated by artificially injecting five types of anomalies (Outlier, Drift, Bias, Minor, and Noise). The model is trained by jointly optimizing the InfoNCE loss and reconstruction loss.

After pretraining, the learned encoder is frozen, and few-shot classification is performed using class prototypes in the latent space. The proposed method is evaluated on three structural vibration datasets—Steel Frame, IASC-ASCE, and Z24—and compared against 10 existing methods. It achieves classification accuracies of 84.55%, 90.45%, and 70.45%, respectively. These results represent improvements of 6.37, 3.63, and 2.27 percentage points over the best-performing baseline for each dataset, with an average accuracy improvement of 4.09 percentage points. Furthermore, ablation studies and t-SNE visualizations are conducted to analyze the effectiveness of the proposed learning components.

Critical strengths

S1. This paper addresses the impact of distribution shifts caused by sensor failures, noise, and data quality degradation on time-series classification performance. The research problem is particularly relevant to practical applications such as Structural Health Monitoring (SHM), where reliable sensor measurements are critical and labeled data are often limited.

S2. The proposed framework adopts a relatively straightforward two-stage architecture that combines contrastive learning, autoencoder-based reconstruction, and prototype-based few-shot classification. In particular, the encoder is frozen after self-supervised pretraining, and classification is performed using only a small amount of labeled data. This design allows the effectiveness of the learned representations to be evaluated without requiring additional large-scale fine-tuning.

Critical weaknesses

W1. The most significant concern with this paper is that it does not sufficiently explain how the proposed learning objective encourages the model to learn data-quality-invariant representations. In Section 3.3 and Equation (3), time-series samples with injected anomalies are used as negative samples. Consequently, minimizing the InfoNCE loss encourages the model to distinguish between representations of original data and anomaly-injected data. However, data-quality invariance generally implies that representations of the same structural state should remain consistent despite changes in sensor data quality. From this perspective, there is a conceptual inconsistency between the proposed negative sampling strategy and the claimed data-quality invariance. Furthermore, the reconstruction loss in Equation (1) is formulated using only the original signals as inputs, rather than reconstructing the original signals from anomaly-injected data. Therefore, its connection to learning data-quality-invariant representations is also unclear.

W2. In Section 4.1, the authors define four states: Healthy-Normal, Healthy-Anomalous, Damaged-Normal, and Damaged-Anomalous. Classifying these four states as distinct classes requires the learned representations to distinguish not only the structural condition but also the sensor data quality condition. In contrast, data-quality-invariant representations generally aim to suppress the effects of sensor quality variations while preserving information about the actual structural condition. Therefore, high classification accuracy on this four-class task does not necessarily demonstrate that the learned representations are invariant to data quality variations.

W3. The central motivation of this paper is to achieve robust classification performance under distribution shifts caused by sensor data quality degradation. However, the current experiments primarily compare overall classification accuracy against existing methods across three datasets. While the presented t-SNE visualizations and ablation studies demonstrate the discriminative capability of the learned representations and the contributions of individual components, they do not directly establish invariance to changes in data quality. Similarly, the hyperparameter sensitivity analysis in Appendix B does not adequately address this central validation concern.

Questions for authors

Q1. I would like the authors to clearly define what is meant by data-quality invariance and explain how the proposed contrastive loss and reconstruction loss encourage the learning of such representations. In particular, Equation (3) uses anomaly-injected time-series samples as negative samples, encouraging the model to distinguish them from the original data. I would appreciate theoretical or empirical justification for why this strategy promotes representations that remain consistent for the same structural condition, rather than representations that are sensitive to changes in data quality.

Q2. The authors should clarify how Classes 0–3 in Figure 2 correspond to the four states defined in Section 4.1: Healthy-Normal, Healthy-Anomalous, Damaged-Normal, and Damaged-Anomalous. Furthermore, some of the visualizations appear to show that Classes 0 and 1 form one larger group, while Classes 2 and 3 form another, with substantial overlap within each group rather than clear separation among all four classes. I would like to know whether these groups correspond to data with different sensor quality conditions but the same structural condition. The authors should also explain how these observations support the claimed class separability and data-quality invariance.

Q3. To more directly validate the central claim of data-quality invariance, I would like to see an analysis of how consistently the learned representations are maintained between normal and anomalous data corresponding to the same structural condition. Additionally, I would appreciate a quantitative evaluation of robustness to data quality degradation by analyzing how classification accuracy changes with different anomaly injection intensities.

Q4. The current experiments report results using only a single random seed (42). While the proposed method achieves higher classification accuracy than existing methods under this setting, few-shot classification performance can vary depending on randomness during training and the selection of support samples. I would like to know whether similar performance improvements can be observed with different random seeds. If possible, I encourage the authors to report the mean classification accuracy and standard deviation across multiple random seeds and different support set configurations to assess the consistency and reliability of the reported improvements.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.