acceptodds
Under review as a conference paper at ICLR 2027

IDEAS: Universal Framework for Multivariate Time Series Anomaly Detection

Abstract

Anomaly detection in multivariate time series is typically approached with specialized models designed to capture cross-channel dependencies, overlooking algorithms considered strictly univariate. In this work, we revisit this assumption and demonstrate that univariate detectors, when properly adapted, serve as strong and efficient alternatives to natively multivariate models. We present IDEAS (Independent Dimension Ensemble for Anomaly Scoring), a universal framework that enables the application of any anomaly detector to multivariate time series. Specifically, a detector is applied independently to each channel, and the resulting channel-wise scores are combined through a selected strategy to yield an aggregate anomaly score. To capture anomalies expressed through cross-channel relationships, the original channels are augmented with pairwise differences between highly correlated dimensions. Our experiments show that unsupervised univariate detectors adapted with IDEAS consistently outperform strong multivariate baselines spanning classical machine learning, semi-supervised deep learning, and large language models, achieving up to a 3.8 improvement in VUS-PR over the strongest baselines. Furthermore, natively multivariate detectors also improve when deployed within the IDEAS framework. Importantly, IDEAS enables detection of not only the temporal location of anomalies but also the specific channel(s) in which they occur. Finally, IDEAS can significantly accelerate computation by parallelizing a highly efficient univariate detector. These findings establish IDEAS as a highly competitive and scalable alternative to multivariate methods.

Primary area:
learning on time series and dynamical systems
Keywords:
Anomaly Detection, Time-Series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MeoqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

IDEAS applies a base detector independently to each channel and to pairwise differences of highly correlated channels, centers the resulting scores, and aggregates them with a selected operator. The largest improvements appear on six synthetic benchmarks. On 20 real-world datasets, the method is competitive with existing detectors rather than clearly superior.

Critical strengths
  • The procedure is fully specified: a correlation threshold, independent scoring, mean centering, and aggregation by minimum, maximum, median, or a tail-based weight. The dimension score is the correlation between a channel score and the aggregated score. Section 3 is sufficient for reimplementation.
  • When anomalies occupy few channels, the gain in VUS-PR is large. The reported factor of 3.8 corresponds to dataset B with 20 dimensions, and the one- and two-channel groups of dataset A show a comparable margin.
  • Dataset F isolates cross-channel anomalies. Lowering the correlation threshold from 1.0 to 0.9 introduces difference channels and produces a clear increase in VUS-PR, which supports their contribution in this setting.
  • The real-world results are reported in full, and the runtime comparison separates efficient univariate detectors from joint multivariate models.
Critical weaknesses
  • In datasets A–E, the anomalous channels are designated in advance, and the anomaly class is sampled independently for each channel. This protocol favors detectors that score channels separately. The abstract claim of consistent improvement is therefore largely a consequence of the data-generating process.
  • The default aggregation operator is chosen by the mean ratio-to-winner on datasets A–F, which are also the reported evaluation sets. The correlation threshold is selected in the same setting. Selection on the evaluation collections is not separated from the final comparison.
  • On the 20 real-world datasets, EGI-IDEAS ties PCA and OmniAnomaly under the reported mean ratio-to-winner. The authors attribute this outcome to label noise. The conclusion is more qualified than the abstract, while the abstract states the stronger claim.
  • Anomalous-dimension identification is substantially weaker than temporal detection. Top-1 accuracy on dataset B is 0.376, compared with 0.076 for random selection. Dimension-level F1 on dataset A is 0.534, compared with 0.352 for a random predictor.
Questions for authors
  • If the aggregation operator and the correlation threshold are fixed using data disjoint from datasets A–F, do EGI-IDEAS and PCA-IDEAS remain the leading methods?
  • On dataset F with λ = 0.8 and d = 20, how many difference channels are included, and how often do they receive high scores when the per-channel marginals are left unchanged?
  • For the semi-supervised CNN experiments, which real-world datasets provide an anomaly-free training series, and were EGI and MP evaluated on that series or only on the test series?
6 comments

We thank the reviewer for recognizing the clear methodology, strong performance on sparse-channel anomalies, the evaluation of cross-channel anomalies, and the comprehensive real-world and runtime comparisons.

W1. Synthetic data generation and channel-wise scoring.

Datasets A–E are designed to evaluate anomalies affecting different numbers of channels. We acknowledge that independently generated channel anomalies may favor channel-wise methods. However, this setting is practically relevant, as Pinet et al. (2026) show that many anomalies in existing multivariate benchmarks exhibit strong univariate evidence. Importantly, IDEAS also performs well beyond sparse-channel anomalies:

  • Dataset F incorporates cross-channel anomalies while largely preserving individual channel behavior. PCA-IDEAS achieves VUS-PR of 0.90–0.99 across the five groups, compared with 0.70–0.81 for the strongest multivariate baseline in each group (OFA or CBLOF at d=100d=100).

  • Dataset A (danom=20d_{\mathrm{anom}}=20) evaluates anomalies affecting all channels. EGI-IDEAS achieves VUS-PR of 0.95, compared with 0.74 for CNN and 0.72 for OmniAnomaly.

  • The 20 real-world datasets provide evaluation on naturally occurring anomalies beyond our synthetic generation process. EGI-IDEAS nearly matches PCA (mean RTW: 0.59 vs. 0.60), equals OmniAnomaly (0.59), and outperforms the other eight baselines by 0.05–0.26. It also achieves the highest VUS-PR on 3 of the 20 datasets, more than any other detector (Figure 4).

We revised the abstract and Section 1 to distinguish the consistent improvements on synthetic benchmarks from the competitive performance against the strongest baselines on real-world data.

W2 / Q1. Hyperparameter selection on held-out data.

We repeated the selection procedure in Appendix A.7 using independently generated validation sets with different random seeds, disjoint from the test series in datasets A–F (20 series per group, 600 total). The held-out validation selects the same default aggregation operators for all four detectors as those originally used in the paper. The table reports mean RTW for the best and second-best operators:

DetectorBest operator (mean RTW)Runner-up (mean RTW)Paper default
EGImax (0.94)weighted (0.64)max
PCAmax (0.87)weighted (0.80)max
MPweighted (0.86)max (0.78)weighted
CNNweighted (0.92)max (0.83)weighted

The original operator choices are robust: each default outperforms the runner-up by at least 0.07 in mean RTW and is the best operator in 23 of 24 detector–dataset combinations, with MP on dataset F as the only exception. Thus, independent validation reproduces the reported configurations without changing the test results.

Role of λ\lambda. The correlation threshold λ\lambda controls which channel pairs receive pairwise-difference channels. Lowering λ\lambda includes more pairs, while λ=1\lambda=1 excludes all pairwise-difference channels.

Dataset F. The held-out validation selects λ=0.8\lambda=0.8 for all four IDEAS detectors, consistent with Figure A.3. The table reports mean VUS-PR:

Detectorλ=0.8\lambda=0.8λ=0.9\lambda=0.9λ=1.0\lambda=1.0
EGI0.860.800.39
MP0.730.600.25
CNN0.910.900.12
PCA0.970.960.15

Dataset F introduces cross-channel anomalies while largely preserving individual-channel behavior. Without difference channels (λ=1\lambda=1), these anomalies are difficult to detect, yielding VUS-PR of only 0.12–0.39. Difference channels expose disrupted relationships, substantially improving performance at λ=0.9\lambda=0.9. Lowering λ\lambda to 0.8 increases the proportion of selected pairs from approximately 55% to 80%, providing further gains, particularly for EGI-IDEAS (0.80 to 0.86) and MP-IDEAS (0.60 to 0.73).

Datasets A–E. These datasets are not designed to introduce cross-channel anomalies. On the held-out validation sets, only 0.8% of channel pairs exceed a correlation of 0.9, compared with 0.9% at a threshold of 0.8. Consequently, lowering λ\lambda adds fewer than one difference channel per series on average, and mean VUS-PR differs by at most 0.018 across all detectors and datasets. We therefore retain λ=0.9\lambda=0.9 as the default for A–E.

The held-out validation reproduces the original configurations, leaving the reported test results unchanged, with EGI-IDEAS and PCA-IDEAS remaining the leading methods. We added the validation results to the Appendix A.7 and clarified the selection protocol in Section 4.2.

W3. Real-world results and the abstract.

On the 20 real-world datasets, EGI-IDEAS performs on par with the strongest baselines (mean RTW: 0.59 vs. 0.60 for PCA and 0.59 for OmniAnomaly) and outperforms the other eight baselines, while requiring no training data, unlike OmniAnomaly, and being much more computationally efficient than PCA.To investigate the performance difference, we analyzed the 498 real-world test series by the number of labeled anomalous segments:

  • 1–3 segments (248 series), as in our synthetic datasets: EGI-IDEAS ranks first among all 14 detectors (mean RTW: 0.57 vs. 0.50 for PCA and OmniAnomaly).
  • 4 or more segments (250 series): EGI-IDEAS ranks fifth, with PCA and OmniAnomaly performing significantly better (paired Wilcoxon, p<0.001p < 0.001).

This trend also holds within datasets. Regression with dataset fixed effects shows that EGI-IDEAS's VUS-PR decreases by 0.074 relative to PCA and OmniAnomaly per tenfold increase in anomalous segments (p<0.001p < 0.001), while the fraction of anomalous time points has no significant association (p=0.71p=0.71). This is consistent with EGI's reliance on rare subsequence patterns, which may become less distinguishable when anomalous patterns recur frequently.

Real-world benchmarks also present possible annotation and evaluation challenges (Appendix A.10 and Figure A17): 30.5% of series have more than 10% anomalous time points (up to 98.5%); in 20.3%, a channel's largest deviation occurs outside labeled anomalies and exceeds its largest deviation within them by at least threefold; and in 11.8%, none of the 14 detectors achieves VUS-PR above 0.2.

We revised the abstract, Section 1, and the conclusion to distinguish the consistent improvements on the synthetic data from the competitive real-world results, and added this analysis to Appendix A.10.

W4. Anomalous-dimension localization.

Anomalous-dimension localization is substantially better than random selection. For the revision, we re-evaluated localization with a single procedure for IDEAS and the multivariate baselines below: the dimension anomaly scores DjD_j of Eq. (7) are computed only for the original dimensions, from the same runs as the reported detection results. The submitted evaluation incorrectly treated the constructed pairwise-difference channels as additional dimensions. EGI-IDEAS achieves a dimension-level F1 of 0.546 versus 0.347 for random selection on dataset A, and a top-1 accuracy (HR@1) of 0.410 versus 0.076 on dataset B, more than five times higher. It also exceeds random selection on datasets C–E, and in dimension-level AUC-PR on all six datasets (Table 2).

We further compared EGI-IDEAS with three natively multivariate localization baselines: ALoRa-Loc (ICLR 2026) and the per-channel reconstruction errors of TranAD and OmniAnomaly. EGI-IDEAS outperforms all three on datasets A–E.

Mean dimension-level F1 over the groups of each dataset:

MethodABCDEF
EGI-IDEAS0.5460.2740.6730.2370.6690.729
ALoRa-Loc0.2980.1320.3990.1040.3940.693
TranAD0.2830.1270.3600.0980.3560.523
OmniAnomaly0.3480.1340.4310.1060.4310.713
Random0.3470.1160.4510.0900.4610.550

Mean HR@1 over the groups of each dataset:

MethodABCDEF
EGI-IDEAS0.6780.4100.8820.4340.9120.883
ALoRa-Loc0.2680.0680.4640.0820.5970.677
TranAD0.3750.1020.6340.0860.6480.664
OmniAnomaly0.3850.0940.6060.0760.6200.586
Random0.2250.0760.4690.0500.4910.661

Table 2 in the revised paper reports dimension-level AUC-PR. EGI-IDEAS exceeds the strongest localization baseline by up to 2.2× in dimension-level F1 and 5.0× in HR@1, whereas all three baselines fall below random selection in F1 on C and E and are often close to random. Although localization remains more challenging than temporal detection, these results demonstrate the effectiveness of IDEAS in identifying anomalous dimensions.

We also evaluated localization on series with strong temporal detection (series-level AUC-PR > 0.5):

Series with AUC-PR > 0.5EGI-IDEASMP-IDEASCNN-IDEASPCA-IDEAS
B: HR@10.770.900.950.26
D: HR@10.830.860.990.41
A: F10.670.600.560.53
C: F10.720.620.550.55
E: F10.720.550.540.58

Localization is more accurate on series with strong temporal detection, particularly HR@1 on datasets B and D. We included the baseline comparison in Section 5.3 (Table 2) and Appendix A.13, which reports dimension-level F1, HR@1, and AUC-PR for every group.

Q2. Difference channels on dataset F (λ=0.8\lambda=0.8, d=20d=20).

On average, 150.8 of 190 possible channel pairs satisfy ∣ρ∣>0.8|\rho| > 0.8 (range: 100–190), resulting in approximately 171 channels per series, including the 20 original channels.

We further analyzed channel scores across 195 labeled anomaly intervals, where cross-channel dependencies are disrupted while individual-channel distributions are largely preserved:

MeasurementEGI-IDEASMP-IDEASCNN-IDEASPCA-IDEASReference
Intervals where a difference channel has the highest peak score96.4%99.0%99.5%99.5%88%
Flagged points that fall within anomalies (difference channels)31.8%18.4%24.3%27.3%5.8%
Flagged points that fall within anomalies (original channels)14.4%11.2%8.6%7.2%5.8%

For the first row, we compare each channel's peak score within an anomaly interval. For the remaining rows, each channel flags its top 7.5% highest-scoring time points. The reference values are the proportion of difference channels (88%) and the anomaly rate (5.8%).

Paired difference channels produce the highest peak scores in 96.4–99.5% of labeled anomaly intervals, and their flagged points are 1.6–3.8× as likely to fall within anomalies as those flagged by original channels. This is consistent with Figure A3, where adding paired difference channels raises VUS-PR on the validation set of dataset F from 0.09–0.47 (λ=1\lambda=1) to 0.50–1.00 (λ=0.8\lambda=0.8).

Q3. Training series on real-world datasets.

Every real-world test series is paired with a training series:

  • For the 16 datasets from mTSBench, we use the official training files, which contain no labeled anomalies.
  • For SWaT A2/A5 and WADI A2, we use normal-operation recordings from the testbeds.
  • For TSB-AD-M, we use the benchmark's original split, with the first part of each series used for training and the remainder for testing.

Only semi-supervised detectors use these training series: CNN-IDEAS, CNN, LSTMAD, OmniAnomaly, TranAD, ALLM4TS, and OFA. EGI-IDEAS, MP-IDEAS, and PCA-IDEAS, along with the unsupervised baselines, operate exclusively on the test series without using training data. We clarified this protocol in Appendix A.3. We appreciate the reviewer’s constructive suggestions, which have helped us strengthen both the evaluation and the scope of our claims. Independent validation reproduces the original configuration choices, while the new localization comparisons demonstrate gains over established alternatives. The additional real-world analysis clarifies the conditions under which IDEAS performs well and where challenges remain. We hope these results address the concerns about evaluation rigor and empirical scope and support a more favorable assessment of our work.

Reviewer_adg3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a framework for multivariate time series anomaly detection. The proposed scheme is detector agnostic. It first construct pair-wise difference channels for input channels that exhibit strong correlation and next run uni-variate anomaly detection at each channel (origina + constructed) independently. The individual anomaly scores are combined to detect anomalies. The proposed scheme is evaluated on six synthetic and twenty real-world datasets and posts competetive performance. The proposed scheme also exhibits large runtime advantages when paired with efficient univariate anomaly detectors.

Critical strengths
  • The proposed method is simple and easy to understand. It lends itself well in deployment in industry settings.
  • Experimental setup is comprehensive.
  • The proposed scheme can be easily parallelized, which translates into runtime performance gains.
  • The proposed scheme makes it relatively straightforward to attribute an anomaly to a particular dimension (channel) in the input signal. This is useful for real-world applications.
Critical weaknesses
  • The novelty is somewhat modest. See (i) (Tafazoli and Keogh, 2023) and (ii) (Pinet et al., 2026).
    • Still the idea is well-executed in this work.
  • There seems to be a performance gap between results obtained on synthetic data and on real-world benchmarks. This gap should be acknowledged.
  • \lambda is set to 0.9 for most experiments. The paper can benefit from a sensitivity analysis on \lambda.
  • As acknowledged in the work, aggregation strategy is highly dataset dependent.
  • It is not immediately clear to me the role of centering since no ablation/experiments study this explicitly.
  • I suspect that the paper presupposes access to powerful univariate anomaly detectors. This fact is somewhat undersold in this work.
Questions for authors
  • Sensitivity analysis around \lambda
  • How can one explain the difference between results obtained on synthetic and real-world data
  • Ablation to centering idea from Sec. 3.2.
  • How would one deal with quadratic increase in pair-wise difference channels for large input dimensions?
3 comments

We thank the reviewer for recognizing the simplicity and practical applicability of IDEAS, the comprehensive experimental evaluation, and the benefits of parallelization and anomalous-dimension localization.

W1. Methodological novelty.

Both cited works are already discussed in Section 2. Tafazoli and Keogh (2023) propose an approach for identifying anomalous subsets of dimensions, demonstrated using Matrix Profile. In contrast, IDEAS provides a unified framework for applying both univariate and multivariate detectors through independent channel-wise scoring, aggregation, and pairwise-difference channels. Pinet et al. (2026) analyze anomaly characteristics in existing benchmarks and introduce a synthetic cross-channel anomaly setting, rather than a general-purpose anomaly detection method. Our contribution lies in the generality and demonstrated effectiveness of the framework. The proposed framework fills an important gap in the existing literature by enabling the consistent application of highly efficient univariate anomaly detectors to multivariate time series data, with demonstrated state-of-the-art performance.

W2 / Q2. Synthetic vs. real-world performance.

We agree and have revised the abstract and Section 1 accordingly. On synthetic datasets, the best IDEAS detector outperforms all baselines in every group. On the 20 real-world datasets, EGI-IDEAS performs on par with the two strongest baselines (mean RTW: 0.59 vs. 0.60 for PCA and 0.59 for OmniAnomaly) and outperforms the other eight, while requiring no training data, unlike OmniAnomaly, and being much more computationally efficient than PCA.

To investigate this difference, we compared the anomaly characteristics of synthetic and real-world series (details of this analysis are included in Appendix A.10). Each synthetic series contains 1–3 anomalous segments, covering 5–6% of time points (Table A1). In contrast, half of the 498 real-world series evaluated in Figure 4 contain at least four labeled anomalous segments (up to 678), and 30.5% have more than 10% anomalous time points. We therefore grouped the real-world series by the number of labeled anomalous segments:

Real-world seriesSeriesEGI-IDEASPCAOmniAnomalyEGI-IDEAS rank (of 14)
1–3 anomalous segments2480.570.500.501
4 or more anomalous segments2500.610.690.685

On series with 1–3 anomalous segments, EGI-IDEAS achieves the highest mean RTW among all 14 detectors, with no significant difference from PCA or OmniAnomaly (paired Wilcoxon test, p>0.4p > 0.4). On series with four or more segments, both baselines significantly outperform EGI-IDEAS (p<0.001p < 0.001). This is consistent with how EGI detects anomalies: it identifies rare subsequence patterns, which may become less distinguishable when anomalous patterns recur frequently. Our synthetic datasets do not evaluate such frequent anomalies. Label inconsistencies may also contribute to performance variability on real-world benchmarks, as discussed in Appendix A.10.

W3 / Q1. Sensitivity to λ\lambda.

Figure A3 evaluates λ∈0.8,0.9,1.0\lambda \in \\{0.8, 0.9, 1.0\\} on the validation set of dataset F, where anomalies disrupt cross-channel dependencies. Without paired difference channels (λ=1\lambda=1), VUS-PR ranges from 0.09 to 0.47. At λ=0.9\lambda=0.9, performance improves substantially across all four IDEAS detectors. Lowering λ\lambda to 0.8 yields further gains, particularly for EGI-IDEAS (0.55–0.99 to 0.79–1.00) and MP-IDEAS (0.34–0.74 to 0.50–0.87).

For the validation sets of datasets A–E, which contain only a small fraction of cross-channel anomalies, λ\lambda has a limited effect. Our additional experiments show that λ=0.8\lambda=0.8 and 0.90.9 differ by at most 0.018 in mean VUS-PR across all detectors and datasets. We included a discussion of these results in Appendix A.6.

W4. Dependence of aggregation strategy on the dataset.

The aggregation strategy depends primarily on the base detector rather than the dataset. We repeated the selection on independent held-out validation sets and obtained the same default aggregation operator for each of the four detectors. Results on the validation sets reported in Table A4 show that the selected default performs best in 23 of the 24 dataset-detector combinations: max performs best for EGI and PCA across all six datasets, while weighted performs best for CNN across all six datasets and for MP on five. Thus, the aggregation strategy can be selected once per base detector rather than tuned for each dataset. For new detectors, if no validation set is available, we recommend max as a strong default, consistent with Appendix A.7.

W5 / Q3. Role of centering.

We conducted a centering ablation for EGI-IDEAS. The results show substantial improvements due to centering in mean VUS-PR:

Dataset A: Varying anomalous dimensions

danomd_{\mathrm{anom}}1251020Mean
With centering0.390.520.810.920.950.72
Without centering0.200.250.270.280.230.24

Dataset B: Varying total dimensions

dd5102050100Mean
With centering0.460.440.380.300.350.38
Without centering0.390.320.200.110.090.22

In the ablation runs, centering improves VUS-PR on 428 of 500 series in dataset A and 321 of 500 in dataset B. By subtracting each channel's mean score, centering prevents channels with consistently high background scores from dominating the aggregated score, allowing contributions from prominent anomaly-score peaks to stand out. On dataset A, VUS-PR increases from 0.39 to 0.95 as anomalous dimensions increase, whereas without centering it remains between 0.20 and 0.28. On dataset B, the advantage of centering grows with dimensionality, reaching 0.35 versus 0.09 at d=100d=100.

We clarified the role of centering and added this ablation study in Appendix A.8.

W6. Dependence on the base detector.

We agree that IDEAS inherits the detection ability of its base detector, and we stated this explicitly as a limitation in the paper. However, its gains do not require a strong base detector. Using the same base detector, the IDEAS adaptation outperforms the corresponding multivariate version in all 30 groups of datasets A–F (VUS-PR):

  • CNN-IDEAS vs. CNN: 30/30 groups (median 2.2× higher);
  • PCA-IDEAS vs. PCA: 30/30 (median 1.5×);
  • MP-IDEAS vs. a previously proposed multidimensional extension of Matrix Profile (MP-XXVIII): 30/30 (median 2.3×).

CNN and PCA are generic models rather than specialized univariate detectors, suggesting that these gains arise from the IDEAS framework itself: independent channel-wise scoring, centering, aggregation, and difference channels. Because IDEAS is detector-agnostic, improvements in anomaly detection can also be incorporated directly into the framework.

Q4. Quadratic growth of pairwise-difference channels.

Although the worst-case number of difference channels is O(d2)O(d^2), the actual number depends on how many channel pairs satisfy the correlation threshold. On our real-world test series at λ=0.9\lambda=0.9, only 3% of possible pairs are selected on average, increasing the total number of channels by a factor of 1.28 (median 1.0, 95th percentile 2.5). Even for d=248d=248 in OPPORTUNITY, fewer than 50 pairs were selected in 7 of the 8 series. For higher-dimensional data, the number of paired difference channels can be capped by retaining only the kk most correlated partners per channel. We clarified this practical scaling behavior in Section 3.2.

We thank the reviewer for recognizing IDEAS’s simplicity, practical applicability, and comprehensive evaluation. The new centering ablation, correlation-threshold analysis, and independent validation of aggregation choices strengthen our understanding of which components drive its performance. The additional real-world analysis also helps explain the gap between synthetic and real-world results and clarifies the framework’s limitations. We believe these additions materially strengthen the paper and reinforce the reviewer’s positive assessment of its practical value.

Thank you again for your constructive review and for recognizing IDEAS's strengths. We have addressed your main concerns in our rebuttal and revised manuscript:

Sensitivity and centering: On datasets A–E, λ=0.8\lambda=0.8 and 0.9 differ by at most 0.018 in mean VUS-PR, while difference channels substantially improve cross-channel anomaly detection on dataset F. Our centering ablation shows improvements on 428/500 series in A and 321/500 in B.

Aggregation and base detector: Independent validation confirms the aggregation choices for all four detectors, with the selected operators performing best in 23/24 detector–dataset combinations. CNN-IDEAS, PCA-IDEAS, and MP-IDEAS also outperform their corresponding multivariate baselines across all 30 synthetic groups.

Real-world performance and scalability: We clarified the distinction between synthetic and real-world results and analyzed the effect of anomalous-segment frequency. In real-world data, only 3% of possible channel pairs are selected on average, increasing the channel count by just 1.28×. Pair selection can also be capped.

We would appreciate knowing whether these additions address your concerns. Please let us know if any specific issue remains unresolved.

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

The study proposes IDEAS, a universal framework that enables the application of existing univariate anomaly detection methods to multivariate time series. The framework applies a univariate anomaly detector independently to each time series and aggregates the resulting anomaly scores using different aggregation functions to obtain an overall anomaly score for each multivariate observation. Additionally, to capture anomalies arising from changes in the linear relationships between time series, the framework introduces auxiliary channels constructed from the differences between highly correlated pairs of time series. These auxiliary channels are also processed by the univariate anomaly detector, allowing the framework to identify anomalies that may not be detectable by examining individual time series independently.

Critical strengths
  • (1) The idea of constructing auxiliary channels from the differences between highly correlated pairs of time series to detect anomalies arising from changes in their linear relationships is interesting.

  • (2) The authors provide an extensive experimental evaluation across multiple datasets and a wide range of evaluation metrics. Additionally, they employ appropriate evaluation metrics that align with recent recommendations in the time-series anomaly detection literature.

  • (3) The authors consider the anomaly localization task in addition to anomaly detection, recognizing the importance of identifying which time-series dimensions are responsible for anomalous behavior.

Critical weaknesses
  • (1) The proposed IDEAS framework relies primarily on applying existing anomaly detectors independently to each time-series channel and aggregating the resulting anomaly scores using standard aggregation functions. This strategy is conceptually similar to the channel-wise aggregation of reconstruction errors commonly employed in multivariate time-series anomaly detection, where operators such as the L1 norm, L2 norm, or maximum are used to obtain an observation-level anomaly score. Although IDEAS additionally introduces pairwise-difference channels to capture anomalies in cross-channel relationships, I believe there is room for a stronger methodological contribution beyond the combination of these relatively straightforward components.

  • (2) The authors fix the detection threshold at τ = 0.925 across all experiments. Although Appendix clarifies that the same thresholding procedure is applied to the comparison baselines, the rationale behind selecting this particular value is not sufficiently discussed. A more comprehensive evaluation could report the best achievable performance of each method across different detection thresholds, using a metric • The authors fix the detection threshold at τ = 0.925 across all experiments. Although Appendix clarifies that the same thresholding procedure is applied to the comparison baselines, the rationale behind selecting this particular value is not sufficiently discussed. A more comprehensive evaluation could report the best achievable performance of each method across different detection thresholds, using a metric.

  • (3) The experimental evaluation lacks comparisons with recent SOTA methods for MTS anomaly detection. In particular, the most recent baseline included in the study dates from 2024, and none of the selected baselines originate from ICLR. I encourage the authors to extend their experimental comparisons to include representative methods from ICLR 2026, NeurIPS 2025, and ICLR 2026.

  • (4) The experimental results are reported without standard deviations, and the number of independent runs used to obtain the reported mean values is not specified.

Questions for authors
  • (1) Based on the formulation in Equation (7), my understanding is that the proposed localization procedure computes the correlation between each channel's anomaly-score vector and the aggregated anomaly-score vector over the entire time series. This suggests that anomalous dimensions are identified at the time-series level rather than independently at each time step. I encourage the authors to clarify whether this interpretation is correct and explain how the proposed approach could identify the anomalous dimensions associated with individual time steps.

  • (2) The idea of introducing auxiliary channels based on the differences between highly correlated time series is interesting. However, relying solely on Pearson correlation to select channel pairs primarily captures linear relationships, potentially overlooking nonlinear dependencies. Moreover, high Pearson correlation does not necessarily indicate a meaningful relationship, as it may arise from spurious correlations. Could the authors consider a more systematic approach to selecting channel pairs and constructing auxiliary channels, enabling the framework to detect anomalies arising from changes in meaningful cross-channel dependencies under more realistic conditions?

7 comments

We thank the reviewer for recognizing the value of pairwise-difference channels, the comprehensive evaluation, and anomalous-dimension identification.

W1. Relation to channel-wise aggregation of reconstruction errors.

Unlike multivariate approaches that aggregate channel-wise errors from a jointly trained model, IDEAS applies detectors independently to each channel and centers their scores before aggregation. This preserves sparse anomaly evidence and enables strictly univariate detectors such as EGI and MP to operate on multivariate data.

Our key contribution lies in demonstrating that simple, independent channel-wise scoring can outperform specialized joint multivariate models. Of particular practical importance is the superior capability of the IDEAS detectors to find anomalies in difficult cases where they are contained in only one or few channels among many.

This contribution is supported by our controlled comparison: CNN-IDEAS consistently outperforms jointly trained CNN across all 30 synthetic groups (datasets A–F) (in the revision, we extended dataset F from three to five groups by adding d=50d=50 and d=100d=100), including sparse anomalies (VUS-PR: 0.29 vs. 0.11, dataset A, danom=1d_{anom} = 1) and cross-channel anomalies (0.92 vs. 0.16, dataset F, d=20d = 20). Combined with pairwise-difference channels, IDEAS provides a detector-agnostic framework for multivariate detection and anomalous-dimension localization.

W2. Threshold selection (τ=0.925\tau = 0.925).

Out of 12 metrics that we report, 9 are threshold-independent. Our primary metric, VUS-PR, is threshold-independent, as are AUC-PR, AUC−PTRT\mathrm{AUC-P}_{T}\mathrm{R}_{T}, AUC-ROC, and VUS-ROC. Moreover, Max-F1, Event-based F1, R-based F1, and Affiliation-F1 already optimize over detection thresholds. For example, under Max-F1, the best IDEAS detector outperforms all baselines across all dataset A–F groups (see Figures A4–A9). For 3 threshold-dependent metrics (Precision, Recall, and F1), we use the same τ=0.925\tau = 0.925 across all detectors and datasets, flagging the top 7.5% of time points as anomalous. Using a fixed τ\tau value for threshold-dependent metrics is important for cross-detector and cross-dataset comparability. This threshold was fixed without using test labels and is slightly above datasets anomaly ratio (5–6%, Table A1). We clarified this rationale in Appendix A.5 of the revised manuscript.

W3. Recent baseline comparisons.

We extended our evaluation with two recent baselines: ALoRa (ICLR 2026) and CATCH (ICLR 2025). EGI-IDEAS outperforms both these baselines across all 30 groups of datasets A–F, achieving 1.3–7.5× higher VUS-PR (median 3.2×) than the strongest new baseline (see Figures 2 and 3 in the revised manuscript for more details). The table reports mean VUS-PR values across groups within each dataset. We included these baselines in the revised experimental comparisons and report results across all 12 evaluation metrics in Appendix A.9.

DatasetEGI-IDEASALoRaCATCH
A0.720.330.20
B0.380.090.09
C0.850.400.27
D0.340.100.08
E0.820.450.27
F0.890.170.12

W4. Standard deviations and number of runs.

Each detector is evaluated once per time series, and results are averaged over 100 time series per group (or 500 series per dataset). We capture variability by evaluating time series with different signal patterns, anomaly types, locations, and dimensions. The table reports mean ± standard deviation of VUS-PR and Wilcoxon p across series for one group per dataset:

Dataset (group)Best IDEAS detectorStrongest baselineWilcoxon p
A (danom=5d_{\mathrm{anom}}=5)EGI-IDEAS 0.81 ± 0.29CNN 0.28 ± 0.26<0.001
B (d=20d=20)EGI-IDEAS 0.38 ± 0.34MP-XXVIII 0.11 ± 0.15<0.001
C (d=20d=20)EGI-IDEAS 0.84 ± 0.28OmniAnomaly 0.39 ± 0.38<0.001
D (n=2×104n=2\times10^4)EGI-IDEAS 0.34 ± 0.32OmniAnomaly 0.14 ± 0.16<0.001
E (n=2×104n=2\times10^4)EGI-IDEAS 0.87 ± 0.25ALoRa 0.52 ± 0.33<0.001
F (d=20d=20)PCA-IDEAS 0.98 ± 0.05OFA 0.79 ± 0.22<0.001

The standard deviations are large for many detectors because the time series within each group differ in their signals and anomalies. When compared on the same series, the IDEAS improvements are statistically significant across all 30 groups (paired Wilcoxon signed-rank test, p<0.001p < 0.001). We clarified that averaging is over time series in the caption of Figure 2 and reported standard deviations (Tables A6-A11) and statistical tests (Tables A13 and A14) for all groups and metrics in Appendix A.9 and Appendix A.11. We also included box-plots that visualize distributions of VUS-PR values (Figure 3 and Figures A10-A15).

Q1. Granularity of anomalous-dimension identification.

The reviewer's interpretation is correct. Equation (7) identifies anomalous dimensions at the time-series level by measuring the correlation between each dimension's anomaly scores and the aggregated anomaly scores. IDEAS therefore separately identifies when anomalies occur and which dimensions contain them. This is consistent with datasets A–E, where all anomalous dimensions share the same anomaly intervals by construction. Thus, the identified dimensions can be associated with each detected anomalous time step.

IDEAS can also support dimension identification at individual time steps without requiring an additional model. Since IDEAS computes anomaly scores for every channel at every time step, we can directly examine which channels contribute most to each detected anomaly. Specifically, with max aggregation, the channel with the highest centered anomaly score at time step tt determines the aggregated score at that step. This provides a direct way to attribute individual anomalous time steps using the existing channel-wise scores. We included a more detailed discussion of our anomaly localization process in Section 3.2 and Appendix A.5 and detailed analysis of anomaly localization results for all groups in datasets A–F using three different dimension-level metrics (Tables A16-A18) in Appendix A.13.

Q2. Pearson correlation for pair selection.

We use Pearson correlation as a simple and efficient default for selecting channel pairs. However, IDEAS does not require Pearson correlation. Practitioners can choose alternative dependence measures based on their data characteristics. To demonstrate this flexibility, we evaluated Spearman rank correlation and mutual information (rescaled to the correlation scale for consistent thresholding). The following table reports EGI-IDEAS results on dataset F (d=20d=20, 100 series, threshold 0.8):

Pair-selection criterionPairs per seriesVUS-PR
None (λ=1\lambda=1)00.39
Pearson, test series (paper)150.80.91
Spearman, test series151.10.91
Mutual information, test series167.90.92

Both alternatives achieve comparable performance to the original Pearson-based selection. Since dataset F contains approximately linear dependencies, Spearman and mutual information select largely the same pairs as Pearson (overlap: 0.99 and 0.90), with no statistically significant performance differences.

These results demonstrate that the choice of dependence measure can be adapted to the application without changing IDEAS's independent scoring and aggregation framework. We included additional analysis to show that Pearson correlation is a default choice, not a requirement, and discussed alternative pair-selection criteria in Appendix A.12.

We appreciate the reviewer’s constructive feedback and recognition of our comprehensive evaluation and pairwise-difference channels. The new comparisons with recent baselines and paired statistical analyses materially strengthen the evidence for IDEAS, demonstrating substantial gains on the controlled benchmarks. Together with the clarifications on thresholding and anomalous-dimension localization, these additions better establish the framework’s practical contribution and empirical reliability. We hope these findings address the reviewer’s concerns and support a more favorable assessment of our work.

Thank you again for your detailed review. We have addressed your main concerns in our rebuttal and revised manuscript:

Methodological contribution: IDEAS independently scores and centers channels, enabling strictly univariate detectors to operate on multivariate data. Using the same base architecture, CNN-IDEAS outperforms jointly trained CNN across all 30 synthetic groups.

Recent baselines and statistical support: We added ALoRa (ICLR 2026) and CATCH (ICLR 2025). EGI-IDEAS outperforms both across all 30 groups, achieving 1.3–7.5× higher VUS-PR than the stronger baseline (median 3.2×). Paired Wilcoxon tests confirm significant improvements across all groups (p<0.001)(p<0.001), supported by per-series variability analyses.

Thresholding and localization: Nine of 12 metrics are threshold-independent or optimize over thresholds. Only Precision, Recall, and F1 use a fixed threshold, applied uniformly without test labels. We also clarified time-step-level dimension attribution and showed comparable pair-selection results using Pearson, Spearman, and mutual information.

We would appreciate knowing whether these additions address your concerns. Please let us know if any specific issue remains unresolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.