acceptodds
Under review as a conference paper at ICLR 2027

JELI: Joint Modeling of Latent and Input-Space Evidence for Time Series Anomaly Detection

Abstract

Time-series anomaly detection is a crucial task, commonly formulated as learning characteristic temporal patterns from normal observations and identifying abnormal deviations. Prior work has broadly explored time-series anomaly detection in two spaces, with input-space approaches detecting anomalies through discrepancies between inputs and their reconstructions, while latent-space approaches identify deviations from normal patterns in learned representations. However, the two approaches have largely evolved as separate research directions, even though they can serve complementary roles in anomaly detection. In this paper, we propose **JELI** (**J**oint **E**vidence from **L**atent and **I**nput Spaces), a unified framework that integrates contextual anomaly evidence from both spaces for time-series anomaly detection. Specifically, we first train a JEPA-based backbone to learn contextual representations from surrounding input context and predict target representations at arbitrary temporal positions, forming a normal representation space. Second, we finetune the pretrained backbone for contextual reconstruction with an attention-masking bottleneck that prevents each target from attending to its own representation while preserving alignment with the pretraining objective. Lastly, to bridge the two sources of anomaly evidence, we design a joint-distribution-based scoring method that explicitly accounts for their dependence. Through extensive experiments, we demonstrate that our framework substantially outperforms prior state-of-the-art baselines on the time-series anomaly detection benchmarks.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Anomaly Detection, Self-supervised Learning, Attention, Transformer

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2TRMOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an anomaly detector that is a combination of two anomaly detectors: one based on some reconstruction error from a transformer, and the other based a Mahalanobis distance to non-anomaly data distribution in the latent representation. These two measures are combined using the joint CDF to obtain a final anomaly probability score, which is a smart idea, because the two anomaly scores have inherently different scales.

Although it is applied to a Transformer-style encoder-decoder timeseries model, it is trivially generalizable to any encoder-decoder model.

Critical strengths

The idea is clear and sound. Combining different scores into one is always problematic, and using the 2D CDF is a very nice solution for this.

Critical weaknesses

I'm not sure if it is a critical weakness, but this approach requires that the training data is relatively clean. For both training the model, and fitting the Mahalanobis mean+covariance, you cannot have too much pollution of outlier data. When the Mahalanobis distance is used in high-dimensional spaces (say, more than 100D?) it becomes more important to have enough clean target data.

Questions for authors

To honor past research, the idea of combining different anomaly scores using the 2D CDF is not new. So it would be nice to cite: Li, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., & Chen, G. H. (2022). Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. IEEE Transactions on Knowledge and Data Engineering, 35(12), 12181-12193.

Next, could the authors give an idea on how sensitive this approach is to the fraction of anomalies in the training data. In particular, how sensitive is the Mahalanobis distance to additional outliers?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2TRM

Dear Reviewer 2TRM,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Please refer to the revised manuscript, as all our responses follow its updated content.


[W1-1] I'm not sure if it is a critical weakness, but this approach requires that the training data is relatively clean. For both training the model, and fitting the Mahalanobis mean+covariance, you cannot have too much pollution of outlier data.

[Q2] Next, could the authors give an idea on how sensitive this approach is to the fraction of anomalies in the training data. In particular, how sensitive is the Mahalanobis distance to additional outliers?

[Response to W1-1, Q2] Thank you for this insightful question. As stated in the Introduction, we would first like to clarify that our paper follows the standard unsupervised TSAD setting that assumes normal training data. However, as you pointed out, evaluating robustness to anomalies in training data is also important. To this end, we conduct additional experiments to examine the sensitivity of our approach to training data contamination.

Specifically, we evaluate training contamination ratios of 0%, 1%, 5%, 10%, and 30% by adding Gaussian noise with a magnitude of three times the training standard deviation to selected training windows. As shown below, JELI maintains strong anomaly detection performance even at 30% contamination, achieving a VUS-PR of 0.5742 and a VUS-ROC of 0.9006. Notably, these results still outperform all baselines evaluated without additional training contamination (see Table 1), demonstrating that our approach maintains strong performance even under substantial training contamination. (Note that we report results in terms of VUS-PR / VUS-ROC)

The Mahalanobis-based latent-space score also maintains stable performance under contamination, achieving a VUS-PR of 0.5143 and a VUS-ROC of 0.8752 at 30% contamination. By modeling the normal distribution based on semantic information captured in learned latent representations, this design can help reduce the influence of outliers in the training data [1]. The observed stability of latent-space scoring supports this interpretation. We have incorporated these experimental results and the corresponding discussion into the revised manuscript (Appendix A.2).

Contamination (%)JELI (Latent-only)JELI (Joint)
00.5588 / 0.89170.6493 / 0.9202
10.5345 / 0.88160.6293 / 0.9141
50.5193 / 0.87650.6152 / 0.9115
100.5068 / 0.87220.6015 / 0.9086
300.5143 / 0.87520.5742 / 0.9006

[W1-2] When the Mahalanobis distance is used in high-dimensional spaces (say, more than 100D?) it becomes more important to have enough clean target data.

[Response to W1-2] Thank you for this insightful comment. We agree that reliable covariance estimation is important when using Mahalanobis distance in high-dimensional latent spaces. We have considered this issue in our design and adopted the Ledoit-Wolf shrinkage estimator [2], as described in Appendix A.1. This estimator regularizes the covariance matrix to improve numerical stability, allowing Mahalanobis distance to be computed even when the number of reference samples is smaller than the latent dimension. This addresses the concern regarding high-dimensional covariance estimation, while sensitivity to anomalous training samples is examined separately in our response to [W1-1, Q2].


[Q1] To honor past research, the idea of combining different anomaly scores using the 2D CDF is not new. So it would be nice to cite ECOD [3].

[Response to Q1] Thank you for suggesting this valuable reference. We agree that ECOD [3] is an important prior work on CDF-based anomaly scoring. Specifically, our approach models the joint distribution of complementary anomaly evidence from the input and latent spaces, whereas ECOD computes empirical CDFs for individual input features and combines their tail probabilities. We believe this is an important methodological distinction and have added the suggested reference along with a clarification in the revised manuscript (Section 2).


References

[1] Assran et al., Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture, CVPR 2023
[2] Ledoit et al., A Well-Conditioned Estimator for Large-Dimensional Covariance Matrices, Journal of Multivariate Analysis 2004
[3] Li et al., ECOD: Unsupervised Outlier Detection Using Empirical Cumulative Distribution Functions, IEEE Transactions on Knowledge and Data Engineering 2023


We hope that our additional experiments and explanations have addressed your concerns and provided further evidence supporting our framework. If you have any further questions or would like to discuss any remaining points, we would be glad to continue the discussion.

Thank you for your valuable time.

Best regards,
Authors

Dear authors, Thank you for your additional experiments. I think that (in contrast to "the standard unsupervised TSAD setting") assuming clean training data is not very realistic in practice, and I'm therefore happy to see you used a robust estimator. Thanks.

AuthorsOpenReview ↗

Additional Response to 2TRM

Dear Reviewer 2TRM,

Thank you very much for suggesting an important experiment and for taking the time to review our manuscript. We are glad that our additional experiments have helped address your concerns about the practical applicability of our approach.

We hope our work will inspire further research toward practical time-series anomaly detection.

If you have any further questions, we would be happy to hear them.

Best regards,
Authors

Reviewer_pzZBOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes JELI, a time-series anomaly detection framework that combines anomaly evidence from the latent and input spaces. The method uses JEPA-based pretraining to learn contextual representations, followed by contextual reconstruction with an attention-masking mechanism. Anomaly scores from both spaces are then combined using a joint distribution that accounts for their statistical dependence. Experiments on TSB-AD-U and TSB-AD-M show strong performance compared with recent baselines, supported by ablation studies and case analyses.

Critical strengths
  1. I like the simplicity and clarity of the overall framework. The method is clean and easy to understand.

  2. Given its relatively simple design, the strong performance on both TSB-AD-U and TSB-AD-M is quite impressive.

Critical weaknesses
  1. The complementarity between input-space and latent-space anomaly detection is not sufficiently justified, either in the motivation or in the experiments. I am not convinced that the two anomaly types discussed naturally motivate this distinction. It remains unclear whether the improvement simply comes from combining two strong detectors.

  2. The additional improvement from joint anomaly scoring over independent-tail scoring appears relatively small. Also, I could not find the code for reproducing the ablation studies in the supplementary ZIP file.

Questions for authors

See weaknesses. I might adjust my final scores based on the answers of these two weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer pzZB

Dear Reviewer pzZB,

We sincerely appreciate your careful review and thoughtful comments. In what follows, we address your concerns one by one. Please refer to the revised manuscript, as all our responses follow its updated content.


[W1] The complementarity between input-space and latent-space anomaly detection is not sufficiently justified, either in the motivation or in the experiments. I am not convinced that the two anomaly types discussed naturally motivate this distinction. It remains unclear whether the improvement simply comes from combining two strong detectors.

[Response to W1] We respectfully disagree that the improvement of JELI can be explained simply by combining two strong detectors.

Prior research has shown that input-space and latent-space representations capture different levels of information [1], and our approach focuses on leveraging these differences to obtain complementary anomaly evidence. To demonstrate how this complementarity appears in anomaly detection, we present representative cases in the score distributions across multiple datasets in Figure 7 (Appendix A.4) to further support this interpretation. For example, in Daphnet (row 1) and SMD (row 2), the input- and latent-space scores exhibit different distributions, while their joint modeling provides clearer separation between normal and anomalous samples.

These observations support our view that the key benefit lies in jointly modeling complementary anomaly evidence at different levels, rather than simply combining two strong detectors.


[W2] The additional improvement from joint anomaly scoring over independent-tail scoring appears relatively small. Also, I could not find the code for reproducing the ablation studies in the supplementary ZIP file.

[Response to W2] Thank you for your thoughtful comment. We would first like to clarify that both Independent-tail and Dependent-tail (previously Joint-tail) in Table 6 are our joint scoring methods that combine input- and latent-space anomaly evidence.

The key difference between the two methods is whether the covariance between the anomaly scores is considered. Independent-tail scoring can be effective when simultaneously high scores are characteristic of anomalies, whereas dependent-tail scoring can avoid overestimating such combinations when they are common even under normal conditions. Therefore, the performance difference between these two methods reflects the additional benefit of modeling their dependence, rather than the benefit of joint scoring itself. We have clarified these distinctions and renamed “Joint-tail” to “Dependent-tail” in the revised manuscript (Section 4.3). We have also added the code (see script folder and README.md for more details) for ablation studies to the supplementary material.


References

[1] Yang et al., From Observations to States: Latent Time Series Forecasting, ICML 2026


We hope that our responses have addressed your concerns and clarified the key aspects of our framework. We would be happy to address any additional questions you may have.

Thank you for your valuable time.

Best regards,
Authors

Thank you for your response and the additional experiments. However, I feel that my first concern (W1) has not been fully addressed.

First, capturing different information does not necessarily imply complementarity in anomaly detection. The manuscript claims that input-space scores are more sensitive to fine-grained changes, while latent-space scores better capture window-level shifts. I think this is an important claim, as time-series anomalies can broadly involve local outliers or contextual deviations (e.g., periodic anomalies).

However, neither Figure 7 in Appendix A.4 nor the other experiments explicitly examine this distinction. An analysis using synthetic anomalies with different morphologies, or visualizations of how the two branches respond to different anomaly types, would help support this claim.

This also leaves me wondering whether the improvement comes from the specific complementarity of these two spaces, or simply from combining two effective anomaly scores. If the latter is true, why not combine more scores? Prior work [1] has also shown that simple mean-based ensembles can achieve strong results, which further motivates this concern.

I appreciate the authors' efforts, but I still find the evidence for the claimed complementarity insufficient.

[1] Liu, Q., Lee, S., and Paparrizos, J. "TSB-AutoAD: Towards Automated Solutions for Time-Series Anomaly Detection." PVLDB, 2025.

AuthorsOpenReview ↗

Additional Response to Reviewer pzZB

First, capturing different information does not necessarily imply complementarity in anomaly detection. The manuscript claims that input-space scores are more sensitive to fine-grained changes, while latent-space scores better capture window-level shifts. I think this is an important claim, as time-series anomalies can broadly involve local outliers or contextual deviations (e.g., periodic anomalies). … This also leaves me wondering whether the improvement comes from the specific complementarity of these two spaces, or simply from combining two effective anomaly scores. If the latter is true, why not combine more scores?

[Response] We sincerely appreciate your thoughtful follow-up and constructive suggestions. Following your recommendation, we conduct additional experiments with synthetic anomalies to directly examine the complementarity between input- and latent-space anomaly scores.

As illustrated in Figure 8 (Appendix A.5 in the revised manuscript), we consider three types of synthetic anomalies to verify whether the two spaces respond differently to fine-grained and coarse-grained changes. The results are summarized as follows:

  • Rapid oscillation. The rapid fluctuations introduce fine-grained changes that produce large reconstruction errors in the input space, while the latent-space score shows a limited response (see Figure 8a).

  • Gradual rise. The gradual change produces a sustained deviation from normal patterns, which is consistently captured by the latent-space score, whereas the input-space score shows a less consistent response (see Figure 8b).

  • Abrupt level shift. The sudden changes at the boundaries produce sharp peaks in the input-space score. In contrast, the latent-space score captures the sustained shift in the signal level, remaining elevated throughout the anomaly region (see Figure 8c).

Importantly, the joint score (last row) captures both fine-grained and coarse-grained changes by combining the complementary responses from the two spaces. These results provide further evidence that the two spaces capture complementary anomaly information and that our joint modeling effectively integrates their strengths.

For further details on the experimental setup, please refer to Appendix A.5. We also include the generated synthetic dataset and visualization code in the supplementary material for reproducibility.

We hope these additional experiments help clarify the complementarity between the two spaces and address your concern. We would be happy to discuss any remaining questions or suggestions you may have.

Best regards,
Authors

Reviewer_PoDfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The study proposes a new model for time-series anomaly detection, named JELI, that combines anomaly scores from the input space (reconstruction errors) and the latent space (deviations from learned representations of normal behavior). For the latent-space anomaly score, the authors use a JEPA-based Transformer to learn representations of normal temporal patterns and measure deviations using the Mahalanobis distance from the distribution of normal representations. For the input-space anomaly score, the pretrained model is fine-tuned for reconstruction using a masked attention mechanism that prevents each patch from attending to its own representation. This forces the model to reconstruct each patch using information from other patches, with reconstruction errors serving as anomaly scores. Finally, the two scores are combined by modeling their log-transformed joint distribution as a bivariate Gaussian, fitted on normal training data. The final anomaly score is derived from the joint upper-tail probability of this distribution.

Critical strengths
  • (1) The paper demonstrates strong anomaly detection performance across both univariate and multivariate benchmarks. I also appreciate that the authors adopt appropriate evaluation metrics and avoid point adjustment, which has been shown to produce misleadingly optimistic results in time-series anomaly detection.

  • (2) I find the idea of modeling the joint distribution of the latent-space and reconstruction-space anomaly scores interesting. However, I would like to see a more thorough investigation of this component.

Critical weaknesses
  • (1) The methodological novelty appears somewhat limited. The central idea of combining complementary anomaly evidence from different representation spaces has already been explored in prior work. In addition, the proposed model largely builds on an existing JEPA-based Transformer framework, without introducing a substantially new architectural component or modification that would constitute a clear methodological advance for the field. Therefore, the main contribution appears to lie primarily in the integration of existing ideas.

  • (2) The Related Work section is relatively short and should include a broader discussion of recent studies, particularly from ICLR 2026 and NeurIPS 2025. Since a central motivation of the paper is the benefit of jointly exploiting latent-space and reconstruction-space anomaly evidence, the authors should more explicitly discuss prior methods that already combine or leverage complementary anomaly signals. This would help readers better understand what has already been explored in the field and how JELI differs from existing approaches.

  • (3) Missing recent strong baselines. More recent publications should be considered, e.g., ICLR 2026, NeurIPS 2025, etc.

  • (4) There is no computational analysis.

  • (5) In the multivariate setting, the model processes each channel independently and only averages the resulting channel-wise anomaly scores. This may be limiting for real-world multivariate systems, where anomalies can arise from complex cross-variable dynamics and abnormal relationships between channels that cannot be captured by independent univariate modeling.

Questions for authors
  • (1) The construction of the point-wise anomaly score requires further clarification. The input-space anomaly score is first computed by averaging the reconstruction error over the entire window, and the final score for a time point (t) is then obtained by averaging the scores of all overlapping windows containing (t). This raises an important question regarding real-time applicability, since depending on how the windows are constructed, the score at (t) may rely on future observations and introduce detection latency. I suggest that the authors clearly describe the sliding-window construction, stride, and alignment, and explicitly state how much past and future context is used to compute the anomaly score of each individual time point.

Minor comments:

  • I suggest moving the Related Work section immediately after the Introduction to improve the overall flow of the paper.
  • A more precise title for the appendix section “ERROR BAR” would be “Standard Deviations.”
7 comments
AuthorsOpenReview ↗

Response to Reviewer PoDf [1/3]

Dear Reviewer PoDf,

We sincerely appreciate your thoughtful feedback and constructive suggestions. In what follows, we address your concerns one by one. Please refer to the revised manuscript, as all our responses follow its updated content.


[W1] The methodological novelty appears somewhat limited. The central idea of combining complementary anomaly evidence from different representation spaces has already been explored in prior work. In addition, the proposed model largely builds on an existing JEPA-based Transformer framework, without introducing a substantially new architectural component or modification that would constitute a clear methodological advance for the field. Therefore, the main contribution appears to lie primarily in the integration of existing ideas.

[Response to W1] We respectfully disagree with the concern regarding methodological novelty.

First, the key contribution of JELI is not simply combining two anomaly detectors, but jointly modeling complementary anomaly evidence from the latent and input spaces. Although AnomalyTransformer [1] and MEMTO [2] leverage information from both spaces, their final anomaly scores are based on reconstruction errors, with latent-space information used only as auxiliary weights. Consequently, even substantial latent-space deviations can be suppressed when reconstruction errors are small. More recently, OracleAD [4] has also explored combining input- and latent-space anomaly evidence, but simply multiplies the two scores without modeling their joint distribution. In contrast, JELI directly utilizes anomaly evidence from both spaces and explicitly models their joint distribution. Moreover, as highlighted by a recent representative study [3], even latent-space anomaly detection remains relatively underexplored compared with forecasting- and reconstruction-based approaches. Jointly modeling anomaly evidence from both spaces has received even less attention, further highlighting the novelty of our proposed method.

Second, our contribution lies not in introducing a new JEPA architecture, but in extending its role for time-series anomaly detection. JEPA itself is a representation learning framework rather than an anomaly detection method. JELI goes beyond applying JEPA as a backbone by investigating how to learn representations of normal patterns, how to utilize these representations for anomaly detection, and how to derive complementary anomaly evidence from both the latent and input spaces. Through this design, we extend JEPA into a unified framework for unsupervised time-series anomaly detection.


[W2] The Related Work section is relatively short and should include a broader discussion of recent studies, particularly from ICLR 2026 and NeurIPS 2025. Since a central motivation of the paper is the benefit of jointly exploiting latent-space and reconstruction-space anomaly evidence, the authors should more explicitly discuss prior methods that already combine or leverage complementary anomaly signals. This would help readers better understand what has already been explored in the field and how JELI differs from existing approaches.

[W3] Missing recent strong baselines. More recent publications should be considered, e.g., ICLR 2026, NeurIPS 2025, etc.

[Minor Comment 1] I suggest moving the Related Work section immediately after the Introduction to improve the overall flow of the paper.

[Response to W2, W3, Minor Comment 1] Thank you for your thoughtful suggestions regarding recent baselines and related work.

First, we would like to clarify that our original experiments already included recent baselines published at NeurIPS 2025 (CrossAD [6]), ICLR 2026 (PaAno [3]), and ICML 2026 (TimeRCD [5]), as reported in Tables 1 and 2. Other evaluated baselines include ICLR 2025 (DADA [7], CATCH [8]), ICLR 2024 (FITS [9]), ICLR 2023 (PatchTST [10]), NeurIPS 2023 (MEMTO [2]), and ICLR 2022 (Anomaly Transformer [1]). To further strengthen the comparison, we additionally include OracleAD (NeurIPS 2025) [4] as a recent and closely related baseline in the revised manuscript (Section 4.2). For clarity, we also add the publication venues and years to Tables 1 and 2.

Second, we have moved the Related Work section immediately after the Introduction. We believe this organization helps readers better understand the differences between JELI and prior approaches before presenting the methodology. We have also reorganized the Related Work section around input-space and latent-space approaches and clarified the discussion of methods that combine evidence from both spaces. These comparisons further highlight the methodological differences between JELI and prior approaches in the revised manuscript (Section 2).

We sincerely appreciate your suggestions and would also welcome any recommendations for additional baselines that you consider particularly relevant. We would be happy to conduct further comparisons to strengthen our evaluation.

AuthorsOpenReview ↗

Response to Reviewer PoDf [2/3]

[W4] There is no computational analysis.

[Response to W4] Thank you for this valuable suggestion. Following your recommendation, we conducted an additional computational analysis comparing the inference time and GPU memory usage of JELI with all baselines in Table 1. For a fair comparison, we use the same GPU, input length, and batch size across all methods.

As shown below, JELI achieves an inference time of less than 2 ms per batch (faster than seven of the ten baselines) and GPU memory usage below 40 MiB (lower than eight baselines). Compared with AnomalyTransformer [1], MEMTO [2], and OracleAD [4], which utilize information from both spaces, JELI achieves the fastest inference time and the lowest GPU memory usage.

Importantly, JELI also achieves the best VUS-PR and VUS-ROC among all compared methods, demonstrating that our framework combines state-of-the-art anomaly detection performance with high computational efficiency. We have incorporated this analysis into the revised manuscript (Appendix A.3).

MethodVUS-PRVUS-ROCInference Time (ms/batch)GPU Memory (MiB)
AnomalyTransformer [1]0.10380.540711.507660.6
MEMTO [2]0.10600.56146.284270.3
PatchTST [10]0.47840.83821.24344.6
FITS [9]0.39030.76380.2598.4
DADA [7]0.32350.77045.018353.7
OracleAD [4]0.35050.80322.152144.7
CATCH [8]0.36990.82072.419206.2
CrossAD [6]0.42640.82342.40975.4
TimeRCD [5]0.52240.87114.142174.5
PaAno [3]0.52570.88330.41331.5
JELI (Ours)0.64930.92021.86337.3

[W5] In the multivariate setting, the model processes each channel independently and only averages the resulting channel-wise anomaly scores. This may be limiting for real-world multivariate systems, where anomalies can arise from complex cross-variable dynamics and abnormal relationships between channels that cannot be captured by independent univariate modeling.

[Response to W5] Thank you for your valuable comment. The primary focus of JELI is to improve anomaly detection by jointly modeling complementary evidence from the input and latent spaces, rather than explicitly modeling cross-channel dependencies. Accordingly, we focus on capturing anomaly evidence within each channel and adopt a simple averaging of channel-wise anomaly scores for multivariate time series. Despite this simple approach, JELI substantially outperforms recent methods that explicitly model cross-channel dependencies, such as OracleAD [4] and CATCH [8], as shown below. These results demonstrate that jointly modeling input- and latent-space evidence is effective even in the multivariate setting. We view incorporating cross-channel dependencies as a potential direction for future work.

VUS-PRVUS-ROCRange-F1AUC-PRAUC-ROCPoint-F1Avg Rank
OracleAD [4]0.32780.74430.33790.34040.71960.38884.33
CATCH [8]0.32640.73470.25750.31490.70450.36487.00
JELI (Ours)0.46840.82150.46870.44210.81410.49841.00
AuthorsOpenReview ↗

Response to Reviewer PoDf [3/3]

[Q1] The construction of the point-wise anomaly score requires further clarification. The input-space anomaly score is first computed by averaging the reconstruction error over the entire window, and the final score for a time point (t) is then obtained by averaging the scores of all overlapping windows containing (t). This raises an important question regarding real-time applicability, since depending on how the windows are constructed, the score at (t) may rely on future observations and introduce detection latency. I suggest that the authors clearly describe the sliding-window construction, stride, and alignment, and explicitly state how much past and future context is used to compute the anomaly score of each individual time point.

[Response to Q1] Thank you for raising this important question regarding real-time applicability. While JELI is evaluated in the offline anomaly detection setting commonly adopted in TSAD research, we would like to clarify the sliding-window construction and temporal alignment of point-wise anomaly scores.

Clarification: We construct sliding windows with window length LL and stride 1. For each time point (t), the final anomaly score is obtained by averaging the scores of all overlapping windows containing (t). These windows collectively cover observations from t−L+1t-L+1 to t+L−1t+L-1, requiring at most L−1L-1 past and L−1L-1 future observations.

Real-time Applicability: Importantly, an initial anomaly score can be computed as soon as time point (t) is observed, using the window ending at (t). While this initial score relies on the current and past observations, subsequent overlapping windows incorporate additional future observations, allowing the score to be progressively updated using richer temporal context. This progressive scoring process naturally aligns with the sequential nature of time-series data, allowing anomaly assessments to be updated as new temporal context becomes available.

We have clarified the window construction, temporal alignment, and required future context in the revised manuscript (Appendix A.1).


[Minor Comment 2] A more precise title for the appendix section “ERROR BAR” would be “Standard Deviations.”

[Response to Minor Comment 2] Thank you for your valuable suggestion. Taking your comment into account, along with the additional analysis presented in our response to Reviewer 2TRM [W1-1, Q2], we have renamed the appendix section to “Stability Estimations” to better reflect its expanded scope (Appendix A.2). We would be happy to hear any further feedback on this revision.


References

[1] Xu et al., Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy, ICLR 2022
[2] Song et al., MEMTO: Memory-guided Transformer for Multivariate Time Series Anomaly Detection, NeurIPS 2023
[3] Park and Kang, PaAno: Patch-based Representation Learning for Time-series Anomaly Detection, ICLR 2026
[4] Cho et al., Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection, NeurIPS 2025
[5] Lan et al., Towards Foundation Models for Zero-Shot Time Series Anomaly Detection: Leveraging Synthetic Data and Relative Context Discrepancy, ICML 2026
[6] Li et al., CrossAD: Time Series Anomaly Detection with Cross-Scale Associations and Cross-Window Modeling, NeurIPS 2025
[7] Shentu et al., Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders, ICLR 2025
[8] Wu et al., CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching, ICLR 2025
[9] Xu et al., FITS: Modeling Time Series with 10k Parameters, ICLR 2024
[10] Nie et al., A Time Series is Worth 64 Words: Long-term Forecasting with Transformers, ICLR 2023


We appreciate your valuable feedback and hope that our responses have addressed your concerns. We would be happy to continue our discussion to further improve our paper, so please do not hesitate to share any additional thoughts.

Thank you for your valuable time.

Best regards,
Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers and AC,

We sincerely appreciate the reviewers’ time and effort, as well as their insightful feedback and constructive suggestions. We are also grateful for their thoughtful assessments of our work, including its interesting and sound joint distribution-based scoring approach (Reviewers PoDf, 2TRM), simple or clear framework (Reviewers pzZB, 2TRM), strong anomaly detection performance across univariate and multivariate benchmarks (Reviewers PoDf, pzZB), appropriate evaluation metrics without point adjustment (Reviewer PoDf), and general applicability to encoder-decoder architectures (Reviewer 2TRM).

To further strengthen our work, we have conducted additional experiments and revised the manuscript accordingly. The key revisions are summarized below:

  • Expand Related Work and Baseline Comparisons

    • Expand the discussion of input-space, latent-space, and hybrid anomaly detection approaches (Section 2).
    • Expand baseline comparisons to include a recent hybrid anomaly detection method (Section 4.2).
  • Clarify Joint Anomaly Scoring

    • Clarify the distinctions between joint anomaly scoring strategies with different treatments of statistical dependence (Section 4.3).
  • Analyze Robustness and Computational Efficiency

    • Conduct additional experiments on sensitivity to training data contamination (Appendix A.2).
    • Evaluate inference time and GPU memory usage against baseline methods (Appendix A.3).
  • Improve Organization and Experimental Clarity

    • Move and reorganize the Related Work section to better contextualize our approach (Section 2).
    • Clarify sliding-window construction and temporal alignment for point-wise anomaly scoring (Appendix A.1).
  • Validate Complementarity Between Input and Latent Spaces

    • Conduct synthetic anomaly experiments to examine complementary responses to fine-grained and coarse-grained changes (Appendix A.5).

All revisions are highlighted in magenta\textcolor{magenta}{magenta} in the revised manuscript for ease of reference.

We hope that these revisions further strengthen our work and address the reviewers’ concerns. We would be happy to answer remaining questions or discuss further suggestions.

Thank you for your time and effort.

Best regards,
Authors

AuthorsOpenReview ↗

A Gentle Reminder to AC and Reviewers

Dear AC and Reviewers,

We hope this message finds you well. We sincerely appreciate your valuable feedback and would like to kindly follow up on our rebuttal.

We have made a sincere effort to address the reviewers' concerns through detailed clarifications, further analysis, and additional experimental results. We have revised the manuscript accordingly and summarized the key revisions in our General Response. We hope these additions have helped strengthen our paper and clarify important aspects of our work.

If you have any remaining concerns or would like to discuss any aspect of our work further, we would be happy to address them.

Thank you for your time and thoughtful consideration you have devoted to our work.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.