acceptodds
Under review as a conference paper at ICLR 2027

DeDe: Time-Series Deep Decomposition for Explainable Anomaly Detection

Abstract

Industrial applications of time-series anomaly detection (TSAD) require both high accuracy and strong explainability to support reliable business decision-making. However, few existing TSAD methods satisfy these two requirements simultaneously. To address this challenge, we propose an end-to-end model named Time-Series Deep Decomposition (DeDe) for explainable TSAD. DeDe decomposes a time series into trend, seasonal, and residual components while modeling latent normal structure. Each component is encoded with a multi-scale Transformer, and a forecasting objective augmented with contrastive learning encourages a Stable Latent Manifold Structure (SLMS) of normal dynamics. One-step forecasting error and deviations from the component-wise SLMSs are combined into a point-wise anomaly score. Component-level contributions provide interpretable diagnostic attribution. Experiments on 350 real-world evaluation time series from TSB-AD-U show improvements over the evaluated baselines across multiple metrics. DeDe has also been deployed at scale in industrial production, reliably monitoring millions of time series each day while consistently achieving high recall and precision in real-world operations, demonstrating both strong practical effectiveness and operational robustness.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
deep learning, anomaly detection

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_PaJeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DeDe for explainable univariate TSAD. A differentiable decomposition module splits each input window into trend (learnable moving-average filter), seasonal (sine projection at the FFT-dominant period, refined by convolutions), and residual components. Each component is encoded by a multi-scale Transformer. The model is trained with a one-step forecasting loss and a multi-view contrastive loss. At inference, the anomaly score sums the standardized forecasting error and the Mahalanobis distances of the component embeddings to Gaussians fitted on training data. Ratios between these scores are used as view-level and component-level attributions.

Critical strengths
  • Evaluation follows the official TSB-AD-U protocol with many metrics and a broad set of shallow, deep, and foundation-model baselines.
  • The ablation shows that adding the embedding-distance score to the forecasting error gives a clear gain over forecasting error alone.
Critical weaknesses
  • The restriction to univariate series is weakly justified and conflicts with the stated emphasis on growing scale and complexity. Whether the method extends to multivariate data is not discussed.
  • The need for decomposition is unsupported. Decomposition-based detection (e.g., STL residual thresholding) is long-standing, and the paper does not show what existing decomposition-aware methods fail to do beyond noting that their decomposition is fixed or non-differentiable.
  • For univariate series, simple methods already perform well (Liu & Paparrizos, 2024; Sub-PCA in Table 1). There is no argument for why a deep model is needed, and no simple decomposition baseline is included.
  • The method is unnecessarily complex. Learnable decomposition, per-component multi-scale Transformers, forecasting plus contrastive learning, SLMS, and score fusion are combined without linking any of them to a specific challenge.
  • The paper criticizes forecasting methods for large residuals during normal regime transitions, yet uses forecasting error directly in its own score, so that failure case remains.
  • “D-STL” is not STL, it has no Loess, uses a single sine at the FFT-dominant period, cannot capture phase (as acknowledged), and cannot handle multiple periods. Because its weights are learned, the components are not guaranteed to be semantically trend, seasonal, and residual, which undermines component-level attribution.
  • Explainability is neither defined nor properly validated. Attributions are non-causal ratios of distance magnitudes, the discordance logic is narrative, and the only quantitative check is an injection experiment in the appendix. “SLMS” is simply a Mahalanobis distance to a Gaussian, so “manifold” is an overstatement.
  • Removing decomposition, contrastive learning, or multi-scale encoding each costs only about 0.02, without standard deviations. Most of the gain comes from adding the Mahalanobis score, not from decomposition.
  • TSPulse, which the authors acknowledge reports a higher VUS-PR, and PaAno are excluded, yet superiority over all baselines is claimed. Evaluation is limited to TSB-AD-U, and the deployment claims are not backed by any numbers.
Questions for authors

See the weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer PaJe: Additional Experiments and Clarifications

Thank you for your time and constructive feedback on DeDe’s scope, design, explainability, and evaluation. In response, we have added a simple STL-Residual baseline, controlled component-recovery experiments, MD-only scoring, full-benchmark multi-seed results, and a PaAno comparison. We have also clarified the role of each module, defined the attribution rule, renamed the latent reference, and updated the performance discussion. We have uploaded a revised manuscript incorporating the changes described below. Additions and revisions are highlighted in blue to help you locate the changes. Below, we address your concerns point by point.

  1. Univariate scope and industrial use cases (W1, W10b)

Our focus on univariate detection reflects our internal financial monitoring tasks, where channels often operate independently even within the same business scenario. We therefore prioritize detecting and explaining anomalies within each channel. Joint modeling is valuable when cross-channel dependencies provide diagnostic information, as with related metrics in cloud operations. We will consider extending DeDe to such settings.

DeDe runs on several NVIDIA A10 GPUs as an anomaly-detection service in our internal financial forecasting workflows. It handles thousands of requests per minute and typically responds within seconds, including network overhead. Before deployment, our expert team reviewed DeDe’s detection and classification results and confirmed that the system met internal operational acceptance criteria. Experts continue to review results during daily use and guide improvements. We cannot disclose detailed business data or evaluation results for confidentiality reasons.

Component scores support initial assessment of the three broad cause categories used internally and feed into subsequent agentic root-cause analysis.

  1. Decomposition and learned representations (W2–W4)

W2 — Decomposition. We use decomposition to support attribution of anomalous changes to trend, seasonal, or residual behavior. D-STL learns the trend filter and seasonal refinement through forecasting and CL, allowing component extraction to adapt jointly with the representations used for detection. Removing decomposition reduces VUS-PR from 0.495 to 0.477 (Table 2); replacing it with classical STL gives 0.459 with the same encoders, training objective, and scoring (Table A3).

W3 — Comparison with a simpler detector. Table 1 reports AUC-PR/VUS-PR of 0.461/0.495 for DeDe versus 0.375/0.424 for Sub_PCA. We also evaluated a simple STL-Residual baseline. It applies classical STL and scores anomalies using the residual’s absolute deviation from its training median, scaled by the training median absolute deviation (MAD). We selected H = 200 from {50, 100, 200} on the official 48-series Tuning split and evaluated it on all 350 Eval series. STL-Residual achieved AUC-PR/VUS-PR of 0.224/0.278. DeDe scored higher on all nine metrics, showing a substantial performance advantage over classical decomposition followed by residual scoring. The STL-Residual comparison and its settings are reported in Appendix C.1 (Table A2).

W4 — Design rationale. We learn the three-component decomposition jointly with the detector to adapt component extraction to the detection task. Multi-scale Transformers model patterns of different durations, while CL encourages consistent representations across augmented normal windows. Component-wise latent references quantify each component’s deviation from normal behavior and support attribution. Forecasting scores unexpected next observations, while structural scoring measures unusual patterns in the history window; score fusion combines these signals. Table 2 evaluates the contribution of these modules.

  1. Forecasting error during normal transitions (W5)

DeDe combines forecasting error with component-wise structural deviations to assess both unexpected observations and changes in temporal patterns. Table 2 shows that Dual scoring improves VUS-PR from 0.396 with forecasting alone to 0.495. In our internal monitoring, recurring normal transitions represented in training data produce few false alarms.

(To be continued)

AuthorsOpenReview ↗

Response to Reviewer PaJe (continued)

  1. Decomposition and component meaning (W6)

The submitted D-STL retains STL’s trend–seasonal–residual organization and learns the decomposition for anomaly detection. It uses a learned trend filter and a fixed-phase, dominant-frequency sine projection with nonlinear convolutional refinement in place of LOESS. Section 3.3 describes these operations and their gradient path.

We evaluated the complete DeDe model with the submitted single-sine basis on a controlled, trend-dominated sequence with weak Gaussian residuals, without component supervision. Across all 951 overlapping test windows, its trend, seasonal, and residual RMSEs were (0.091, 0.115, 0.145), versus (0.105, 0.132, 0.166) for classical STL supplied with the true period. These results demonstrate more accurate component recovery in this setting. Appendix C.1 provides the training setup, full-test metrics, and illustrative decompositions (Figure A2).

  1. Component attribution (W7)

We define explainability as identifying how a detected event departs from normal temporal behavior. Under our forecasting and decomposition assumptions, DeDe provides four diagnostic signals: prediction error and structural deviations in trend, seasonality, and residuals. These signals help distinguish an unexpected observation from an unusual historical pattern and identify the main structural change to investigate.

Equation (18) compares forecasting and structural score magnitudes. Equation (19) standardizes component distances using normal training statistics; the event source is the component with the largest mean score over the evaluation interval.

Table A4 reports 418/450 correct source predictions (92.89%) for known single-component injections.

Discovering finer anomaly patterns within these diagnostic dimensions through unsupervised learning is a future direction.

  1. Latent reference terminology (W8)

Thank you for the suggestion. Our aim is to measure departures from normal temporal patterns separately for each component. Nonlinear encoders produce the representations; a mean and regularized covariance estimated from training embeddings define each reference for Mahalanobis scoring. We have renamed SLMS “Component-wise Latent Reference” to make this distinction clear.

  1. Statistical reliability and module contributions (W9)

Thank you for highlighting the contribution of the Mahalanobis score. Table 2 and Appendix Table A1 now include MD-only scoring alongside MSE-only and Dual. Appendix B also reports full-model mean ± standard deviation across three additional runs on all 350 Eval series.

  1. Updated baseline comparison (W10a)

We completed the PaAno reproduction on all 350 TSB-AD-U Eval series using the authors’ configuration: patch size 96 and 100 optimizer updates. PaAno obtained AUC-PR 0.4690 and VUS-PR 0.5290, closely matching the published means of 0.4682 and 0.5296. DeDe’s submitted values are 0.4610 and 0.4950. Table 1 now reports all nine metrics for PaAno, and the performance discussion has been updated accordingly.

PaAno achieves strong results on TSB-AD-U, but its scoring pipeline uses future observations as well as historical data (Park and Kang, 2026, Algorithm 2). An industrial monitoring system cannot access those future observations when an immediate alert is needed. In our hourly monitoring workflows, reproducing its 96-point patch aggregation can require waiting up to 95 hours. Avoiding this wait changes the aggregation rule, so its strong benchmark performance may not carry over to real-time deployment. DeDe scores each observation upon arrival using historical context, provides component-level attribution, and achieves competitive detection performance.

We recognize TSPulse-FT’s strong reported VUS-PR of 0.52. We have not included it in the current comparison because it uses substantial external pretraining data, whereas DeDe uses only the prescribed TSB-AD-U training prefixes. We will consider including TSPulse-FT in the final version, with the difference in training resources clearly stated.

Thank you for the response. But some concerns remain.

  • TSB-AD is offline, and the paper does not define an online setting. If online use is the point, please state it and mark which baselines are causal. TSPulse is excluded for external pretraining, but MOMENT, TimesFM, and Lag-Llama are included.
  • Classical STL (0.459) is worse than no decomposition (0.477), and the learned one adds only +0.018. If only learned decomposition helps, the components may not be real trend/seasonal/residual. This conflict with attribution is not discussed. My D-STL points (no Loess, one period, no phase) are not answered.
  • The injection test uses single-component injections under the same decomposition assumption. What about real or mixed anomalies?
Reviewer_h3HQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary:The paper proposes Time-Series Deep Decomposition(DeDe), an end-to-end univariate time-series anomaly detection framework. Contributions:The paper introduces that each window decomposes into trend-seasonal-residual.It encodes each component with multi-scale Transformers.It provides component-level anomaly attribution based on the relative contribution of these scores. Validation:The authors evaluate DeDe on 350 real-world time series from the TSB-AD-U benchmark, comparing it against baseline methods across multiple metrics.

Critical strengths

The method combines observation space residuals with latent space Mahalanobis distance is a sensible approach.It can distinguish magnitude anomalies from structural and contextual anomalies. The two statistics are complementary and well-integrated into a unified framework.

The evaluation uses 350 series from TSB-AD-U with the official protocol and reports multiple metric families . The performance margins over strong baselinesare substantial and suggest practical efficacy.

Critical weaknesses

The Mahalanobis distance relies on a Gaussian assumption for the contrastive. The covariance matrix may be ill-conditioned, and the ridge parameter is not reported.

The hyperparameter sweep is narrow.The number of random seeds is unstated, and single-seed TSAD results are unstable.

Although separating the heterogeneous dynamics into components is a well-motivated idea for structural anomaly detection. In real-world components interact non-linearly. A simple linear summation risks canceling out anomalies.

Questions for authors

Clarification Requests:Given the heterogeneity of the 350 series,could you provide per-series results, a pooled mean is insufficient to confirm that the reported improvements are not benchmark noise. The workflow diagram in the manuscript is not sufficiently clear.Could the authors provide a clearer explanation for the workflow diagram? The current description is insufficient to help readers follow the pipeline.

Suggestions for Future Work:On the learned embeddings,exploring non-parametric alternatives to the Mahalanobis distance could further validate the latent space modeling.

Investigating non-linear interaction terms for the combination of the trend, seasonal, and residual components could mitigate the risk of error cancellation under the strict assumption.

4 comments
AuthorsOpenReview ↗

Response to Reviewer h3HQ: Additional Experiments and Clarifications

Thank you for your time and constructive feedback on covariance estimation, hyperparameter sensitivity, statistical reliability, score aggregation, and the workflow presentation. In response, we have clarified the covariance regularization, added full-benchmark results across three random seeds, and provided a breakdown by source dataset. We have uploaded a revised manuscript incorporating the changes described below. Additions and revisions are highlighted in blue to help you locate the changes. Below, we address your concerns point by point, explain the training and inference stages, and discuss extensions for multiple normal modes and cross-component interactions.

  1. Covariance reference and ridge regularization (W1)

The ridge parameter is ε = 10⁻⁶. We use Σ⁽ᶜ⁾ + εI in each component’s Mahalanobis score. Appendix B already specifies this value; we will also place it beside Eq. (14).

Mahalanobis scoring uses each component’s mean and covariance to define an ellipsoidal normal reference. Adding εI raises each covariance eigenvalue by ε and regularizes the inverse.

  1. Hyperparameter sensitivity (W2a)

Figure 5 (left) shows stable AUC-PR across the tested CL weights λ ∈ [0.05, 0.10] and view counts V ∈ {2, 3, 4}. The right panel shows consistent VUS-ROC, PA-F1, Event-F1, and Affiliation-F across three temporal-scale configurations.

  1. Seeds, uncertainty, and subset-level results (W2b; Clarification 1)

W2b — Random seeds. We repeated the full DeDe model on all 350 Eval series with three random seeds. Across these runs, AUC-PR was 0.461 ± 0.008 and VUS-PR was 0.495 ± 0.006 (mean ± SD; Appendix B).

Clarification 1 — Results by subset. Appendix Table A6 reports all nine metrics as mean ± SD for the 23 TSB-AD-U subsets, covering all 350 Eval series. This breakdown shows performance differences across data sources and follows the dataset-level reporting used by PaAno (Table 15), GBOC (Tables 1–2), and RTDetector (Table 1).

  1. Score aggregation and cancellation (W3)

We use additive scoring to make trend, seasonal, and residual contributions to the structural score directly interpretable, supporting alert diagnosis in our internal monitoring workflows. The trade-off is that nonlinear dependencies between components are not explicitly modeled.

  1. Training and inference workflow (Clarification 2)

We will redraw the workflow as three stages. First, train the decomposition and encoders with forecasting and CL; augmented views are used at this stage. Second, fit the three component references and score-calibration statistics from training data. Third, reuse the trained model and fixed statistics for inference.

Using the target observation’s index t, the input window contains observations from t − W through t − 1. It produces a forecast for the observation at t and three component embeddings. Once the observation at t is observed, its forecasting error is combined with the window’s structural distances to give the score at t. The diagram will show this alignment.

  1. Extensions (Suggestions for Future Work)

We will discuss non-parametric references for multiple normal modes and cross-component interaction scores for joint temporal changes.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal Before the Discussion Closes

Thank you again for your review and positive assessment of our work. We know this is a busy time, and as the rebuttal period draws to a close, we would be grateful if you could spare a few minutes to read our response and share a brief update.

If our response and the revised manuscript have addressed your concerns and you have no further questions, we hope you will consider raising your rating in light of the revisions and additional experimental evidence. If any concerns remain, we would appreciate the opportunity to address them before the discussion closes.

Thank you for your time and consideration.

The author has responded to all the questions I raised, including covariance regularization, hyperparameter sensitivity, uncertainty analysis using multiple random seeds, and the revision of the workflow diagram. I will keep my score.

AuthorsOpenReview ↗

Response to Reviewer h3HQ (Round 2)

Thank you for confirming that our responses have addressed all your questions. We do not understand why your rating remains unchanged, but we thank you for taking the time to reply.

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

This paper proposes DeDe, an explainable time-series anomaly detection framework. It combines forecasting error with MD-based structural deviations to detect anomalies and provide component-level attribution. Experimental results demonstrate competitive performance across multiple metrics.

Critical strengths
  1. Dede addresses both detection accuracy and explainability through decomposition.

  2. The architecture is technical sound. It combines forecasting and contrastive learning with complementary predictive and structural anomaly scores.

  3. The paper is well-structured and easy to follow.

Critical weaknesses
  1. The framework combines several established techniques. The novelty of D-STL and SLMS over existing decomposition-based and contrastive approaches could be better justified.

  2. Sensitivity analysis mainly covers lambda. However, other key hyperparameters, such as the number of views, and scale configurations, window length, contrastive temperature, and so on, are all unexplored. This could largely raise up with issues in terms of the approach's stability and generalizability.

  3. Attribution experiments rely on synthetic, single-component anomalies. Moreover, Equation (19) uses raw distances, whereas the attribution evaluation uses standardized distances. It still has limited explainability validation.

  4. Statistical significance of experimental results, such as confidence intervals and paired statistical tests, are not reported, making it very difficult to assess the consistency of improvements.

  5. Inference latency is reported, but theoretical complexity, memory consumption, and computational comparisons with baselines remain limited.

  6. Another most important concern is that the discrete FFT period selection and single-sine projection raise questions about differentiability and the ability to capture phase-shifted seasonal patterns.

Questions for authors

Please address the main concerns listed in the weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer TgGp: Additional Experiments and Clarifications

Thank you for your time and constructive feedback on DeDe’s contribution, hyperparameter sensitivity, attribution, statistical reliability, computational cost, and differentiability. In response, we have added an STL-Residual baseline, clarified the component-wise latent references, aligned the attribution rule with its evaluation, and expanded statistical and computational reporting. We have also clarified the gradient path through the decomposition module and the seasonal representation. We have uploaded a revised manuscript incorporating the changes described below. Additions and revisions are highlighted in blue to help you locate the changes. Below, we address your concerns point by point.

  1. Contribution of decomposition and component scoring (W1)

DeDe’s design is to learn temporal components and retain a separate normal reference for each at inference. This makes the component distances part of both detection and diagnosis. A trend or seasonal change can then contribute to the anomaly score even when the residual deviation remains small. The classical-STL replacement in Appendix C.1 / Table A3 tests the decomposition choice with the same downstream model: VUS-PR is 0.459 for classical STL and 0.495 for DeDe.

The additional standalone STL-Residual baseline, selected on Tuning and evaluated on all 350 Eval series, achieved AUC-PR/VUS-PR of 0.224/0.278, compared with DeDe’s 0.461/0.495.

DCdetector trains a multi-scale dual-attention model with a purely contrastive objective. DeDe trains forecasting and augmented-view consistency jointly, then scores deviations from the component references. We have renamed SLMS “Component-wise Latent Reference.”

  1. Hyperparameter sensitivity (W2)

Figure 5 (left) shows stable AUC-PR across the tested CL weights λ ∈ [0.05, 0.10] and view counts V ∈ {2, 3, 4}. The right panel shows consistent VUS-ROC, PA-F1, Event-F1, and Affiliation-F across three temporal-scale configurations.

  1. Attribution rules and their evaluation (W3)

We have revised Section 3.7.1 and Eq. (19) to match Appendix C.2: each component distance is standardized using its mean and standard deviation on normal training windows, and the event source has the largest mean score over the evaluation interval.

Table A4 reports 418/450 correct source predictions (92.89%). Detection sums raw distances before branch-level standardization; Eqs. (14)–(17) are unchanged.

In our monitoring workflows, alerts prioritize the dominant component to guide initial investigation. DeDe retains all component scores, so reporting can be extended to multiple components using individual thresholds calibrated on normal training data and adjusted to operational needs. We will add controlled experiments with simultaneous multi-component anomalies to the final manuscript to evaluate whether these thresholds correctly identify the affected components.

  1. Statistical reliability (W4)

Appendix B reports full DeDe results over all 350 Eval series across three random seeds: AUC-PR 0.461 ± 0.008 and VUS-PR 0.495 ± 0.006 (mean ± SD). The ablations are reported in Table 2 and Appendix Table A1.

  1. Computation, memory, and measured latency (W5)

For the implemented univariate model, let W be the window length, n_γ = max{1, floor(γW)} the length at scale γ, L the encoder depth, and d the hidden and component-embedding width. The computation per window at inference is:

O(W log W + 3L Σ_γ (n_γ²d + n_γd²) + 3d²).

The terms correspond to decomposition, the three multi-scale encoders, and component-distance scoring. The normal references require O(3d²) storage and are fitted once after training. Inference uses one unaugmented pass.

Under the training settings in Appendix B, the total training time for DeDe across all 350 Eval series is approximately 3.5 hours, excluding test inference and metric evaluation. Across all 350 series, this involves approximately 82 million training windows of length 50 accumulated over epochs, counting windows before contrastive augmentation.

We measured DeDe and seven baselines on an NVIDIA A100 80 GB using the same 8,192-point synthetic univariate series and inference batch sizes of 16, 32, 64, 256, and 512. The complete scoring procedure is timed, excluding fitting and warm-up. Figure A1 in Appendix B reports the measured runtime and average and peak allocated GPU memory.

  1. Differentiability and phase representation (W6)

Frequency selection uses a discrete FFT argmax. Conditional on the selected frequency, gradients update the projection and learned filters. Section 3.3 now describes this gradient path.

The submitted projection uses a fixed-phase sine basis and estimates its signed amplitude. Section 3.3 describes this dominant-period prior and its learned convolutional refinement.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal Before the Discussion Closes

Thank you again for your review. We know this is a busy time, and as the rebuttal period draws to a close, we would be grateful if you could spare a few minutes to read our response and share a brief update.

If our response and the revised manuscript have addressed your concerns and you have no further questions, we hope you will consider raising your rating in light of the revisions and additional experimental evidence. If any concerns remain, we would appreciate the opportunity to address them before the discussion closes.

Thank you for your time and consideration.

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

This paper proposes DeDe, an end-to-end framework for time-series anomaly detection (TSAD) aiming to reduce the entanglement of heterogeneous temporal dynamics and provide intrinsic component-level attribution. DeDe uses a Differentiable STL (D-STL) to decompose input windows into trend, seasonal, and residual components, encoding each via multi-scale Transformers. The model is trained jointly with forecasting MSE and multi-view contrastive loss. For inference, it fuses observation-space evidence (forecasting error) and latent-space evidence (Mahalanobis distance of embeddings) to compute anomaly scores and component-level attributions. Experiments are conducted on the univariate TSB-AD-U benchmark for detection performance, alongside controlled synthetic anomaly-injection studies to validate the attribution accuracy.

Critical strengths

The paper's primary strengths lie in its well-motivated dual-space framework and rigorous validation of explainability. Fusing observation-space forecasting errors with latent-space Mahalanobis distances effectively addresses the blind spots of single-view detectors. Furthermore, the controlled anomaly-injection study on synthetic data provides compelling and rigorous empirical validation for the model's intrinsic component-level attribution (~93% accuracy), significantly surpassing subjective post-hoc visualizations.

Critical weaknesses
  1. Inconsistent Motivation: The abstract claims "explainable anomaly detection," but the introduction lists three disjointed technical points (reducing entanglement, intrinsic attribution, dual-space fusion). Only "intrinsic attribution" directly serves explainability. The authors fail to define the specific deficiencies of existing explainable TSAD or justify how decomposition and dual-space fusion fundamentally improve explainability, rather than merely boosting detection accuracy.

  2. Insufficient Related Work: Despite extensive prior research on disentanglement, intrinsic attribution, and dual-space fusion, the paper only compares with distant or weak baselines (e.g., Autoformer, TimesNet, LIME/SHAP). It omits highly relevant direct competitors, such as MSHTrans [1] (decomposition), TimeInf [2] and OracleAD [3] (intrinsic attribution), and OracleAD's dual-scoring mechanism. Consequently, DeDe's distinct contribution boundary remains unclear.

  3. Limited Novelty (System Integration rather than Mechanism Innovation): D-STL, multi-scale Transformers, InfoNCE, Mahalanobis distance, and Z-score fusion are all existing components. The paper fails to demonstrate that this specific combination yields "emergent" capabilities that cannot be achieved by individual components. The methodological contribution appears to be a straightforward system integration rather than a novel mechanism.

  4. Unfair Baseline Selection and Tuning: The main table omits highly relevant recent works like TSPulse [4] and PaAno [5] under the pretext of "non-controlled reproduction." Additionally, recent deep models (e.g., RTDetector 0.163, GBOC 0.246) perform drastically worse than the decade-old Sub_PCA (0.375). Given that Sub_PCA is a known strong baseline, the severe underperformance of recent deep models strongly implies unfair tuning budgets or undertraining.

  5. Lack of Multivariate Validation: The evaluation is restricted to the univariate TSB-AD-U subset, ignoring the multivariate TSB-AD-M. It is unclear how D-STL captures inter-channel dependencies (shared vs. independent decomposition) and how the Mahalanobis covariance avoids the curse of dimensionality in multivariate settings.

  6. Insufficient Statistical Rigor and Reproducibility: The paper lacks statistical significance analysis (e.g., multi-seed means, standard deviations, and hypothesis tests on the full benchmark) to support the claimed improvements. Furthermore, the absence of an anonymous code link hinders the independent verification of the complex D-STL and latent-space computations.

References:

[1] Chen Z, et al. MSHTrans: Multi-scale hypergraph transformer with time-series decomposition for temporal anomaly detection. KDD 2025.

[2] Zhang Y, et al. TimeInf: Time series data contribution via influence functions. ICLR 2025.

[3] Cho D, et al. Structured temporal causality for interpretable multivariate time series anomaly detection. NeurIPS 2025.

[4] Ekambaram V, et al. TSPulse: tiny pre-trained models with disentangled representations for rapid time-series analysis. ICLR 2026.

[5] Park J, Kang S. PaAno: Patch-based Representation Learning for Time-Series Anomaly Detection. ICLR 2026.

Questions for authors
  1. Motivation: The abstract claims "explainable anomaly detection," but the introduction lists three disjointed technical points. How do decomposition and dual-space fusion fundamentally improve explainability, rather than merely boosting detection accuracy?
  2. Related Work: You omitted highly relevant direct competitors such as MSHTrans (decomposition), TimeInf and OracleAD (intrinsic attribution/dual-scoring). What is DeDe's distinct contribution boundary compared to these methods?
  3. Novelty: Since D-STL, Transformers, InfoNCE, and Mahalanobis distance are existing components, does this specific combination yield "emergent" capabilities that cannot be achieved by individual components?
  4. Baseline & Tuning Fairness: Why were strong recent baselines like TSPulse and PaAno excluded? Furthermore, given that Sub_PCA is a known strong baseline, why do recent deep models (e.g., RTDetector 0.163) perform drastically worse? Does this imply unfair tuning budgets?
  5. Multivariate Extension: Why was the multivariate TSB-AD-M ignored? If extended, does D-STL share or independently decompose channels, and how does Mahalanobis distance avoid the curse of dimensionality?
  6. Reproducibility: Can you provide statistical significance tests (multi-seed means/std) on the full benchmark, and an anonymous code link to verify the complex computations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer XiVj: Additional Experiments and Clarifications

Thank you for your time and constructive feedback on the motivation, novelty, baseline fairness, and reproducibility of our work. In response, we have clarified how decomposition supports component-level attribution and distinguished DeDe from the related methods you identified. We have also added an STL-Residual baseline, reproduced PaAno, retuned RTDetector and GBOC on the official Tuning split, and expanded statistical reporting to three runs over all 350 Eval series. We have uploaded a revised manuscript incorporating the changes described below. Additions and revisions are highlighted in blue to help you locate the changes. Below, we address your concerns point by point and discuss the scope and possible multivariate extension of DeDe.

  1. Motivation and contribution (W1, W3 / Q1, Q3)

DeDe explains alerts through prediction error and structural deviations in trend, seasonality, and residuals. The two branches help distinguish an unexpected observation from an unusual historical pattern; component attribution identifies the main structural change to investigate. Decomposition makes these changes separately measurable: each component embedding is compared with its own normal reference. The distances are summed for structural detection and standardized separately for attribution (Eq. (19)).

Appendix C.1 / Table A3 isolates the decomposition choice: with the same encoders, CL objective, and Dual scoring, classical STL gives VUS-PR 0.459 versus DeDe’s 0.495. Table A4 evaluates component attribution under known single-component injections, with 418/450 correct source predictions (92.89%).

On all 350 Eval series, the additional STL-Residual baseline selected on Tuning achieved AUC-PR/VUS-PR of 0.224/0.278, compared with DeDe’s 0.461/0.495. This shows the detection benefit of the complete framework over classical decomposition with residual scoring.

  1. Comparison with related work (W2 / Q2)

MSHTrans fuses seasonal and trend features to reconstruct the input and uses reconstruction error as its anomaly score (Eq. 26). DeDe computes a Mahalanobis distance for each trend, seasonal, and residual embedding using that component’s training mean and covariance.

TimeInf attributes predictions to training time points through influence functions and uses self-influence for anomaly detection. DeDe’s diagnostic output concerns changes in temporal components within the observed signal.

OracleAD combines prediction error with deviations of pairwise distances between variable embeddings from a normal reference matrix (Eqs. 11–17). Its diagnosis identifies variables involved in changed relationships. DeDe combines prediction error with each temporal component’s deviation from its own normal reference, showing whether the trend, seasonal, or residual representation departs from normal behavior within a signal.

(To be continued)

AuthorsOpenReview ↗

Response to Reviewer XiVj (continued)

  1. Baseline selection and tuning (W4 / Q4)

We reproduced PaAno with the authors’ released configuration, including patch size 96 and 100 optimizer updates, on all 350 TSB-AD-U Eval series. Using the benchmark training cutoffs, full-series scoring, and evaluation functions, it achieved AUC-PR 0.4690 and VUS-PR 0.5290, closely matching the published means of 0.4682 and 0.5296. DeDe’s submitted values are 0.4610 and 0.4950. Table 1 now reports all nine metrics for PaAno, and the performance discussion has been updated accordingly.

PaAno achieves strong results on TSB-AD-U, but its scoring pipeline uses future observations as well as historical data (Park and Kang, 2026, Algorithm 2). An industrial monitoring system cannot access those future observations when an immediate alert is needed. In our hourly monitoring workflows, reproducing its 96-point patch aggregation can require waiting up to 95 hours. Avoiding this wait changes the aggregation rule, so its strong benchmark performance may not carry over to real-time deployment. DeDe scores each observation upon arrival using historical context, provides component-level attribution, and achieves competitive detection performance.

We reproduced RTDetector and GBOC from the authors’ public code and searched values from their public configurations. For RTDetector, we retained the official architecture, window length 10, batch size 128, optimizer, and learning-rate schedule, and searched learning rates {0.0001, 0.001, 0.006}. We fixed training at 50 epochs for the TSB-AD-U evaluation. For GBOC, we searched window lengths {2, 5, 10, 50} and learning rates {0.0001, 0.001}, retaining the remaining official architecture and training procedure.

We evaluated all three RTDetector configurations and all eight GBOC configurations on the 48 official Tuning series with seed 2024. We selected one global configuration per detector by mean Tuning VUS-PR and fixed it before evaluating all 350 Eval series. Each series had its own independently trained model; the benchmark’s training cutoff, score alignment, and full-series evaluation were preserved.

Tuning selected window 10 / learning rate 0.006 for RTDetector and window 10 / learning rate 0.001 for GBOC. Both completed all 350 Eval series. The updated reproduction results are:

  • RTDetector: AUC-PR increased from 0.1630 to 0.2913; VUS-PR increased from 0.2150 to 0.3298.
  • GBOC: AUC-PR increased from 0.2460 to 0.2505; VUS-PR increased from 0.3010 to 0.3325.

Table 1 now reports all nine metrics for both detectors, averaged over all 350 Eval series (seed 2024). We recognize TSPulse-FT’s strong reported VUS-PR of 0.52. We have not included it in the current comparison because it uses substantial external pretraining data, whereas DeDe uses only the prescribed TSB-AD-U training prefixes. We will consider including TSPulse-FT in the final version, with the difference in training resources clearly stated.

  1. Multivariate extension (W5 / Q5)

Our focus on univariate detection reflects our internal financial monitoring tasks, where channels often operate independently even within the same business scenario. We therefore prioritize detecting and explaining anomalies within each channel.

A multivariate extension could first decompose channels independently, then model cross-channel relationships and map each component to a fixed-width embedding. Regularized covariance references would be fitted from normal training windows, with covariance dimension determined by the embedding width.

  1. Statistical reliability and reproducibility (W6 / Q6)

Appendix B reports full DeDe results over all 350 Eval series across three random seeds: AUC-PR 0.461 ± 0.008 and VUS-PR 0.495 ± 0.006 (mean ± SD). Appendix Table A6 reports all nine metrics as mean ± SD for the 23 TSB-AD-U subsets, covering all 350 Eval series.

We will release the code and reproduction scripts upon acceptance. Appendix B specifies the model configuration and training protocol; Section 3.6 describes score calibration.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal Before the Discussion Closes

Thank you again for your review. We know this is a busy time, and as the rebuttal period draws to a close, we would be grateful if you could spare a few minutes to read our response and share a brief update.

If our response and the revised manuscript have addressed your concerns and you have no further questions, we hope you will consider raising your rating in light of the revisions and additional experimental evidence. If any concerns remain, we would appreciate the opportunity to address them before the discussion closes.

Thank you for your time and consideration.

Reviewer_8EPTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Industrial monitoring wants detectors that explain, not just flag. DeDe's starting observation is that a time series superimposes heterogeneous dynamics — drift in the trend, periodic distortion in the seasonality, spikes in the residual — and that jointly modeling them lets benign variation in one component mask abnormal change in another. DeDe therefore decomposes each series into trend, seasonal and residual components with a differentiable STL module, encodes each with a multi-scale Transformer, and trains with a one-step forecasting objective plus a multi-view InfoNCE contrastive loss to shape a "Stable Latent Manifold Structure" (SLMS) of normal dynamics. The point-wise score combines forecasting error with component-wise Mahalanobis deviations from the SLMS, and component-level contributions yield diagnostic attribution. On 350 univariate series from TSB-AD-U, DeDe reports improvements over 17 baselines across AUC-ROC/PR, VUS-ROC/PR and several F1 variants. The abstract also claims large-scale industrial deployment monitoring millions of series daily.

Critical strengths
  1. The component-wise reference idea is well-motivated and directly serves explainability. Assessing each of trend/seasonal/residual against its own normality, then attributing by relative contribution, is a natural fit for the "which structure broke" question, and the paper argues convincingly that entangled dynamics both hurt detection and destroy explainability.
  2. The decomposition is learned, and the controlled comparison shows it matters. Replacing the D-STL module with classical STL (LOESS + FFT period estimation, same downstream architecture and scoring) drops AUC-PR from 0.461 to 0.404 and VUS-PR from 0.495 to 0.459 (Appendix C.1) — real evidence that task-adaptive decomposition is doing work a fixed decomposition does not.
  3. The evaluation is broad and follows a demanding protocol. TSB-AD-U's 350 series, 17 baselines including time-series foundation models (Lag-Llama, TimesFM, MOMENT zero-shot and fine-tuned, DADA), and the full TSB-AD metric suite (AUC, VUS, and five F1 variants) — all with the official implementations. The main-table ablation cleanly separates decomposition, contrastive learning, multi-scale encoding, and the dual-branch score (dropping the structural score costs the most: AUC-PR 0.461→0.396).
  4. Attribution is evaluated, not just asserted. The injection study (Table A3, 450 attributions over 50 synthetic series × 3 seeds) reports 94.0%/92.0%/92.7% source accuracy for trend/seasonal/residual anomalies, and the value-pattern-consistency scenarios plus the attribution-to-action mapping (Table A4) are concrete attempts to connect scores to operations.
  5. The reporting is candid in places that matter. The paper proactively discloses failure cases (weak anomaly contrast, strong periodicity, repeated wave structures), discloses that TSPulse reports a higher fine-tuned VUS-PR (0.52 vs 0.495) under a different pipeline, and states that the attribution ratios describe score magnitude, not causal effect.
Critical weaknesses
  1. The contribution is an integration of established components, and the paper does not isolate a non-obvious design decision. Differentiable STL, multi-scale Transformers, forecasting + contrastive training, Mahalanobis scoring — each is standard; the contribution is their combination plus component-wise references. There is no new theoretical characterization, no new learning principle, and no "replace only X" study that a simpler pipeline could not replicate. For a top-tier venue this is the central concern, and the thin differentiation from decomposition-aware and frequency-domain baselines leaves the engineering-integration reading unchallenged.
  2. The seasonal term is a single sine, which structurally limits exactly the regime the method targets. The FFT step picks one dominant period, and the projection estimates a signed amplitude but explicitly does not represent arbitrary phase (the paper's own wording). Multi-period, phase-shifted, or non-stationary periodicity cannot be captured, and the refinement network "cannot guarantee recovering the phase information lost in projection." Since the premise is that decomposition quality drives detection, this is a bottleneck inside the load-bearing module, not a peripheral limitation.
  3. "SLMS" is a per-component Gaussian fit, and the name oversells the mechanism. The "stable latent manifold structure" is means and covariances scored by regularized Mahalanobis distance — a unimodal assumption that fails under multi-modal normality (multiple operating regimes), which industrial telemetry commonly has. Calling this a "manifold" is not a wording quibble; it signals a geometric claim the mechanism does not make.
  4. Only univariate data is covered, while the advertised setting is overwhelmingly multivariate. TSB-AD-U by construction; the industrial motivation in the introduction (monitoring, diagnosis, actionable response) presumes multi-sensor context. The gap between motivation and evaluation scope is left unexplained.
  5. Attribution is validated only under controlled injection. The 92.89% headline accuracy is against synthetic single-component injections with known sources; no real fault ground truth is used, and faithfulness in the standard sense (deletion/insertion tests, stability under perturbation) is not evaluated. The paper is careful to label Table A4 "illustrative, not deployment evidence" — but then the explainability claim rests entirely on consistency with injected types.
  6. The main results lack variance and significance, and the internal check only partially covers this. Appendix B reports a stability check on "representative subsets" with std < 0.01, but no full-benchmark error bars or significance tests against baselines. The paper's own disclosure that TSPulse reports a higher VUS-PR under a different pipeline further weakens the "superior to all evaluated baselines" framing — TSPulse was excluded, and the comparison the paper offers instead is a footnote.
  7. The deployment claim is unverifiable as written. "Deployed at scale… monitoring millions of series daily… consistently achieving high recall and precision" — no recall/precision numbers, no false-alarm rate, no duration, no A/B comparison. Reviewers cannot check any of it, and as it stands the sentence functions as marketing rather than evidence.
  8. Cost is acknowledged but not quantified against baselines. The paper concedes DeDe is more expensive than single-branch detectors and gives absolute inference times (4.52 ms/sample at batch 1 on an RTX 4090), but no training/inference time or memory comparison against the 17 baselines under the same budget.
  9. The disclosed failure cases sit inside the method's strongest regime. Under strong periodicity and repeated wave structures — precisely where an STL route should have the advantage — DeDe and the baselines all struggle. The paper is honest about this; the honesty does not resolve the internal tension.
Questions for authors
  1. Can you provide multi-seed mean±std and significance tests for the headline comparison on the full 350-series benchmark? Without them, are the reported improvements robust?
  2. How does DeDe behave under multi-period, time-varying, or non-sinusoidal seasonality? Can the seasonal module be extended (learned basis, harmonic mixture), and what would the controlled STL comparison look like in that regime?
  3. How does the method handle multi-modal normal behavior (multiple operating regimes) given that SLMS is a per-component Gaussian (Mahalanobis) reference? Is the term "manifold" justified?
  4. Is there any real-data validation of attribution (expert-verified root causes), or only controlled injection? Can you add a faithfulness test (deletion/insertion, stability under perturbation)?
  5. What are the concrete deployment numbers (recall, precision, false-alarm rate, monitoring duration) supporting the "millions of series daily" claim? If these cannot be shared, I would prefer the claim removed.
  6. Since TSPulse reports a higher VUS-PR (0.52) under a different pipeline, what prevents including a tuned TSPulse in the main table for a like-for-like comparison?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8EPT: Additional Experiments and Clarifications

Thank you for your time and constructive feedback, and for recognizing the motivation behind component-wise modeling and attribution. In response to your concerns, we have added a simple STL-Residual baseline, controlled component-recovery experiments, full-benchmark multi-seed results, a PaAno comparison, and computational-cost measurements. We have also clarified the latent-reference terminology and attribution rule, expanded the failure-case analysis, and removed the unsupported claim of consistently high recall and precision from the abstract. We have uploaded a revised manuscript incorporating the changes described below. Additions and revisions are highlighted in blue to help you locate the changes. Below, we address your concerns point by point and clarify the remaining limitations.

  1. Contribution and comparison with a simpler detector (W1)

DeDe retains separate normal references for trend, seasonal, and residual embeddings and uses the resulting distances for both detection and component attribution. We make decomposition learnable so that component extraction is shaped by the same forecasting and CL objectives as the encoders. Appendix C.1 / Table A3 holds the encoders, CL objective, and Dual scoring fixed: learned decomposition gives VUS-PR 0.495, compared with 0.459 for classical STL.

MSHTrans fuses seasonal and trend features to reconstruct the input and uses reconstruction error as its anomaly score (Eq. 26). DeDe computes a Mahalanobis distance for each trend, seasonal, and residual embedding using that component’s training mean and covariance.

OracleAD combines prediction error with deviations of pairwise distances between variable embeddings from a normal reference matrix (Eqs. 11–17). Its diagnosis identifies variables involved in changed relationships. DeDe combines prediction error with each temporal component’s deviation from its own normal reference, showing whether the trend, seasonal, or residual representation departs from normal behavior within a signal.

We also evaluated a standalone STL-Residual detector, selecting H = 200 from {50, 100, 200} on Tuning. Across all 350 Eval series, it achieved AUC-PR/VUS-PR of 0.224/0.278, compared with DeDe’s 0.461/0.495. This directly tests the simpler decomposition-and-residual-scoring pipeline and complements the controlled decomposition replacement in Appendix C.1.

  1. Seasonal representation and periodic failure cases (W2, W9 / Q2)

The submitted model uses a fixed-phase sine at the FFT-selected dominant frequency, followed by learned convolutional refinement. We compared single-sine, single-cosine, and sine+cosine bases in the complete model on the same controlled sequence, using identical data splits and training settings without component supervision. Across all 951 test windows, their seasonal RMSEs were 0.115, 0.114, and 0.079, respectively, versus 0.132 for classical STL supplied with the true period. All three variants also reduced trend and residual errors, supporting effective decomposition across these basis choices in this setting. Appendix C.1 reports the setup, full component metrics, and illustrative decompositions (Figure A2).

Thank you for flagging this issue. Our original WSD illustration showed a clear response to the labelled anomaly; selecting it as a failure case was a presentation error. We have replaced Figure A4 in Appendix E with a missed-detection example. Here, forecasting responds to the drop, but structural scoring responds more strongly to nearby normal variation and peaks two steps after the drop. Their sum ranks the normal fluctuation above the anomaly (35.99 versus 31.99), illustrating how differences in response timing and magnitude can weaken anomaly discrimination under additive fusion.

In Figure A5 (Appendix E), a nearby normal pulse scores 7.050, above the anomaly maximum of 6.243. With W = 50, the last 39 labelled anomaly points and the following normal segment have identical all-zero histories and targets, so they receive the same score. A threshold cannot separate these points. A practical next step is to retain longer activity history or model operating states to distinguish changes in activity from sustained inactivity.

In Figure A6 (Appendix E), forecasting contributes 183.99 of the total score of 202.93 at t = 48400, showing that sharp normal beat transitions can receive high anomaly scores. DeDe uses a fixed window of W = 50 across all TSB-AD-U series, while beats in this segment recur every 275–288 samples. This short history does not cover a full heartbeat cycle, which may make normal beat transitions harder to anticipate. A practical next step is to adapt the window length and seasonal period to the temporal structure of the application.

(To be continued)

AuthorsOpenReview ↗

Response to Reviewer 8EPT (continued)

  1. Normal regimes and the latent reference (W3 / Q3)

We have adopted the name “Component-wise Latent Reference.” Each reference summarizes a component’s nonlinear embeddings with a mean and regularized covariance, forming a compact ellipsoidal model of normal temporal patterns. Regime-specific and non-parametric references provide extensions for multiple normal regimes.

  1. Univariate scope and industrial examples (W4, W7 / Q5)

Our focus on univariate detection reflects our internal financial monitoring tasks, where channels often operate independently even within the same business scenario. We therefore prioritize detecting and explaining anomalies within each channel. Joint modeling is valuable when cross-channel dependencies provide diagnostic information, as with related metrics in cloud operations. We will consider extending DeDe to such settings.

DeDe runs on NVIDIA A10 hardware as an anomaly-detection service in our internal financial forecasting workflows. It handles thousands of requests per minute and typically responds within seconds, including network overhead. Before deployment, our expert team reviewed DeDe’s detection and classification results and confirmed that the system met internal operational acceptance criteria. Experts continue to review results during daily use and guide improvements. We cannot disclose detailed business data or evaluation results for confidentiality reasons. Accordingly, we have removed the unsupported claim of consistently high recall and precision from the abstract.

Component scores support initial assessment of the three broad cause categories used internally and feed into subsequent agentic root-cause analysis.

  1. Attribution evidence and validation (W5 / Q4)

Domain experts also review detection and classification outputs in real monitoring workflows (see W7/Q5 above). For controlled single-component injections, Table A4 reports 418/450 correct source predictions (92.89%).

We have aligned Eq. (19) with Appendix C.2: standardize each component distance using normal training statistics, then select the component with the largest mean score over the evaluation interval.

  1. Statistical reliability (W6a / Q1)

Appendix B reports full DeDe results over all 350 Eval series across three random seeds: AUC-PR 0.461 ± 0.008 and VUS-PR 0.495 ± 0.006 (mean ± SD). Appendix Table A6 reports all nine metrics as mean ± SD for the 23 TSB-AD-U subsets, covering all 350 Eval series.

  1. Comparison with PaAno and TSPulse (W6b / Q6)

We recognize TSPulse-FT’s strong reported VUS-PR of 0.52. We have not included it in the current comparison because it uses substantial external pretraining data, whereas DeDe uses only the prescribed TSB-AD-U training prefixes. We will consider including TSPulse-FT in the final version, with the difference in training resources clearly stated.

We completed the PaAno reproduction on all 350 TSB-AD-U Eval series using the authors’ configuration: patch size 96 and 100 optimizer updates. PaAno obtained AUC-PR 0.4690 and VUS-PR 0.5290, closely matching the published means of 0.4682 and 0.5296. DeDe’s submitted values are 0.4610 and 0.4950. Table 1 now reports all nine metrics for PaAno, and the performance discussion has been updated accordingly.

PaAno achieves strong benchmark results, but its scoring pipeline uses future observations (Park and Kang, 2026, Algorithm 2). These observations are unavailable when an industrial monitoring system must issue an immediate alert, so its benchmark performance may not carry over to real-time use. DeDe scores each observation upon arrival using historical context, provides component-level attribution, and achieves competitive detection performance.

  1. Computational cost (W8)

Under the training settings in Appendix B, the total training time for DeDe across all 350 Eval series is approximately 3.5 hours, excluding test inference and metric evaluation. Across all 350 series, this involves approximately 82 million training windows of length 50 accumulated over epochs, counting windows before contrastive augmentation.

We measured DeDe and seven baselines on the same NVIDIA A100 80 GB using an 8,192-point synthetic univariate series and inference batch sizes of 16, 32, 64, 256, and 512. The complete scoring procedure is timed, excluding fitting and warm-up. The figure reports runtime mean and standard deviation, together with average and peak allocated GPU memory. DeDe’s scoring time decreases with larger batches. We have added these curves as Figure A1 in Appendix B.

Thank you for your detailed reply, which has answered my questions. I choose to keep my rating.

AuthorsOpenReview ↗

Response to Reviewer 8EPT (round 2)

Thank you for your feedback and for confirming that our response has answered your questions. We ask that you consider raising your rating in light of the revised manuscript and the additional experimental evidence.

In response to feedback from all five reviewers, including yours, we conducted additional experiments and substantially revised the manuscript. The new results strengthen the empirical evaluation. The revisions also clarify the paper’s contribution, explain how the attribution mechanism works, and make the limitations more explicit.

If you decide to maintain your rating, could you clarify which concerns still carry the most weight in your assessment? In particular, it would help us to know whether any of the points you originally raised still affect your recommendation and, if so, where you find our explanation or evidence insufficient.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.