- 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
Response to Reviewer PaJe (continued)
- 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).
- 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.
- 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.
- 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.
- 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?
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.
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.
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.
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)