acceptodds
Under review as a conference paper at ICLR 2027

CARMA: Correlation-Aware Reconstruction for Multivariate Time Series Anomaly Detection

Abstract

Reconstruction-based detectors for multivariate time series anomaly detection typically learn a single entangled representation of normal behavior and flag anomalies based on reconstruction error. However, real-world anomalies are not only isolated spikes in a single channel but also disruptions in the relationships between channels, where each sensor individually appears normal, but their joint behavior deviates from the norm. Because these models compress both temporal evolution and inter-channel structure into this single space, they lack an explicit mechanism to isolate such inter-channel dependencies from normal temporal dependencies, leading to lower detection rates for cross-channel anomalies. We propose CARMA (Correlation-Aware Reconstruction for Multi- variate Time Series Anomaly Detection), which is a reconstruction-based variational-autoencoder with two complementary latent spaces: a temporal latent that captures sequential dynamics, and a correlation latent that is trained to re- construct the window’s pairwise Pearson correlations. Both representations are fused in the signal decoder, and anomalies are detected from the reconstruction error. We evaluate CARMA on the TSB-AD-M benchmark suite (17 datasets, 180 series) against 29 baseline methods, and show that it significantly outperforms all of them across multiple threshold-free and range-aware metrics, achieving a VUS-PR of 0.4929. Ablations show that removing the correlation branch degrades performance, that adding a second latent without supervision recovers only part of the gain, and that supervising the correlation latent with an explicit correlation target accounts for the rest and produces a sharper error response at anomalies. URL: anonymous.4open.science/r/CARMA-AE8F

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

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_9cjJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CARMA, a variational autoencoder for multivariate time-series anomaly detection that combines timestep-level temporal latents with a window-level correlation latent. The latter is supervised to reconstruct the input window’s pairwise Pearson correlations using a Huber-plus-cosine objective and is broadcast into the signal decoder. At inference, anomalies are scored exclusively through signal reconstruction error, followed by temporal smoothing.

CARMA is coupling an explicitly supervised correlation representation to signal reconstruction. Evaluation on TSB-AD-M comprises 180 series from 17 datasets, 29 baselines, and three training seeds. CARMA achieves the highest reported mean VUS-PR, 0.4929, compared with 0.4409 for the strongest baseline. Additional evidence includes synthetic correlation-breaking anomalies, architectural and objective ablations, and paired statistical comparisons.

Critical strengths

Broad empirical improvement. Results cover 180 series and improve several complementary metrics, with repeated runs and statistical comparisons.

Useful supervision ablation. Adding correlation supervision improves VUS-PR from 0.4719 to 0.4929 within the same dual-latent architecture, supporting its value beyond additional latent capacity.

Critical weaknesses

Nonlinear dependency modeling is insufficiently validated. The explicit correlation target captures only linear dependence. The paper does not establish whether the correlation branch improves detection when nonlinear dependencies change while Pearson correlations remain unchanged, leaving its broader dependency-modeling claims insufficiently supported.

Computational efficiency is not evaluated. The paper provides no complexity analysis, FLOPs, inference latency, or throughput measurements. Consequently, the additional computational cost, accuracy–efficiency trade-off, and scalability to larger channel counts or windows remain unclear.

Questions for authors

Does correlation supervision improve detection of nonlinear dependency anomalies that preserve Pearson correlations? A comparison with the variant without correlation supervision would help establish the method’s scope.

What are CARMA’s computational complexity, FLOPs, inference latency, and throughput relative to the single-latent variant under identical hardware and input settings? Please distinguish training-only correlation computations from inference costs and clarify scaling with channel count and window length.

3 comments

S1 and S2: Broad empirical improvement. Results cover 180 series and improve several complementary metrics, with repeated runs and statistical comparisons. Useful supervision ablation. Adding correlation supervision improves VUS-PR from 0.4719 to 0.4929 within the same dual-latent architecture, supporting its value beyond additional latent capacity.

We thank you for finding the work useful and especially for observing the improvement coming from the dual latent architecture.

W1 and Q1: Does correlation supervision improve detection of nonlinear dependency anomalies that preserve Pearson correlations? A comparison with the variant without correlation supervision would help establish the method’s scope.

The question is very interesting to see if there's any improvement in the detection of nonlinear dependency anomalies. We have applied Pinet's et al. analysis to the TSB-AD-M dataset to isolate series that break only nonlinear relationships, while keeping the Pearson correlations unchanged [1]. But we didn't find any series that break only nonlinear relationships. All the segments that break nonlinear relationships are also either breaking Pearson correlations or are univariate type of anomalies. A summary table of anomaly types is uploaded (https://anonymous.4open.science/r/CARMA-AE8F/rebuttal/summary_tsbadm.xlsx). As a result we couldn't isolate any series that break only nonlinear relationships.

[1] M. Pinet, J. Cumin, S. Berlemont, and D. Vaufreydaz, “Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate,” Jun. 15, 2026, arXiv: arXiv:2606.02670. doi: 10.48550/arXiv.2606.02670.

W2 and Q2: What are CARMA’s computational complexity, FLOPs, inference latency, and throughput relative to the single-latent variant under identical hardware and input settings? Please distinguish training-only correlation computations from inference costs and clarify scaling with channel count and window length.

We provide a scaling cost analysis in the table below. The evaluated channel counts are 20, 50, 100 and the window lengths are 100, 150, and 200 respectively. CARMA’s FLOPs stay within a few percent of the single-latent model (no zcz_c) on every row. Computation increases with the length of the window. All evaluations are performed on the same hardware NVIDIA L40S 48GB VRAM, AMD EPYC 9454 48-Core Processor CPU, 128GB RAM. The training-only correlation computation time ms/window is 0.0018 (C=50, T=100)

Table. Scaling cost. Here SL = Single latent (no zcz_c) version, C=channel count, T=window length.

SettingCARMA params (×106\times 10^{6})SL params (×106\times 10^{6})CARMA FLOPs (×106\times 10^{6})SL FLOPs (×106\times 10^{6})CARMA latency (ms)SL latency (ms)CARMA throughput (×103\times 10^{3}/s)SL throughput (×103\times 10^{3}/s)
C=20C=20, T=100T=1007.612.475084920.01690.012559.080.1
C=20C=20, T=150T=15010.172.477627390.02290.019743.750.8
C=20C=20, T=200T=20012.732.4710169850.03180.027631.436.3
C=50C=50, T=100T=1007.782.505155000.01730.012757.779.0
C=50C=50, T=150T=15010.342.507737500.02320.019943.150.2
C=50C=50, T=200T=20012.902.5010309990.03220.027931.035.8
C=100C=100, T=100T=1009.762.565315120.01760.012856.878.3
C=100C=100, T=150T=15012.322.567957680.02420.020341.449.2
C=100C=100, T=200T=20014.882.56105810230.03350.028429.835.2

W2 and Q2: What are CARMA’s computational complexity, FLOPs, inference latency, and throughput relative to the single-latent variant under identical hardware and input settings? Please distinguish training-only correlation computations from inference costs and clarify scaling with channel count and window length.

(Continue...)

Additionally, we compare the mean time (s) per series on the TSB-AD-M dataset for CARMA with baselines in the table below. The second-best baseline is HYDRA with a VUS-PR of 0.4409 and a mean time per series of 61.29 seconds. The third baseline, StreamVAE, takes longer than CARMA and achieves a VUS-PR of 0.4277. A figure comparing the mean VUS-PR vs mean Time (s) per series on TSB-AD-M is uploaded (https://anonymous.4open.science/r/CARMA-AE8F/rebuttal/VUS-PR_vs_Time_compare_carma_and_baseline.png).

Table. Mean VUS-PR and mean Time (s) per series on TSB-AD-M. The best result in each column is in bold.

MethodVUS-PRTime (s)
CARMA0.4929111.59
CARMA w/o zcz_c0.462083.02
HYDRA0.440961.29
StreamVAE0.4277170.88
PaAno_PAI0.40494.59
USAD0.40266.29
TimesNet0.391022.67
PatchTST0.390221.74
xLSTMAD0.3895163.42
FITS0.364250.31
MSCRED0.362210.56
OFA0.3557132.24
MMPAD0.3532333.98
MTAD_GAT0.350625.15
GDN0.331810.67
OmniAnomaly0.3153112.30
CNN0.29605.76
CST_GL0.293735.01
KMeansAD0.286736.36
AutoEncoder0.283513.44
PCA0.269191.86
TranAD0.265113.12
OCSVM0.262818.52
IForest0.248760.07
RobustPCA0.237976.54
EIF0.205285.76
COPOD0.20150.42
HBOS0.18560.23
KNN0.169330.74
LOF0.131030.68
AnomalyTransformer0.098930.19

Thank you for your comments. We believe the complexity analysis provides additional support for CARMA. We will include it in the appendix of the manuscript.

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

The paper introduces a multivariate reconstruction-based anomaly detector with auxiliary correlation supervision. The correlation latent is trained to reconstruct pairwise Pearson correlations and is used to condition signal reconstruction. Anomalies are detected using reconstruction error alone. The authors provide an ablation study showing promising performance and improvement compared to adding an additional latent.

Critical strengths

The paper provides interesting perspective on whether explicit supervision of a VAE latent with pearson correlations provides a useful inductive bias that improves anomaly detection beyond other methods of learning multivariate normality. The paper is generally well written and offers an evaluation against 29 baselines on 17 datasets which seems to be a substantial effort.Critical Weaknesses*

Critical weaknesses

One major issue in the approach' motivation is that the authors conflate element-wise reconstruction losses with independent per-channel modelling, i.e., a conventional multivaritate AE already learns representations of the joint input and can capture inter-channel dependencies. Therefore, it is unclear what fundamental limitation CARMA addresses, beyond adding an additional correlation-based inductive bias. But on the other hand, CARMA provides no theoretical or controlled empirical argument showing that its Pearson-correlation supervision is preferable to learning the normal joint distribution directly. For general multivariate distribution, pairwise correlations cannot fully characterize the distribution and these correlations contain only 2nd order dependency information, while in principle, a joint density estimator, capture much richer relationships. There is probably an advantage in using these statistics, but more focused experimentation and analysis needs to be provided.

The experiment the authors are providing by introducing a second latent is interesting but seems problematic. Introducing a 2nd latent increases the bottleneck capacity and provides an additional information pathway to the decoder. Consequently, improvements may partly result from relaxed compression constraints rather than meaningful separation of temporal and correlation information. Although the unsupervised second-latent ablation partially addresses this concern, the paper lacks a capacity-matched single-latent baseline with the same auxiliary correlation supervision.

Questions for authors

Why is explicit correlation supervision needed when conventional multivariate autoencoders already learn joint representations and inter-channel dependencies, despite using elementwise reconstruction losses? What specific limitation does CARMA address?

What motivates separating temporal and correlation latents? Is there evidence that these representations are disentangled or complementary? Could the authors compare against a capacity-matched single-latent VAE with the same auxiliary correlation objective and comparable KL regularization?

2 comments

S: The paper provides interesting perspective on whether explicit supervision of a VAE latent with pearson correlations provides a useful inductive bias that improves anomaly detection beyond other methods of learning multivariate normality. The paper is generally well written and offers an evaluation against 29 baselines on 17 datasets which seems to be a substantial effort.

Thank you for acknowledging our efforts, and for finding the perspective of explicit Pearson supervision interesting.

W1 and Q1: Why is explicit correlation supervision needed when conventional multivariate autoencoders already learn joint representations and inter-channel dependencies, despite using elementwise reconstruction losses? What specific limitation does CARMA address?

Conventional multivariate autoencoders can learn joint representations, but that reconstruction does not explicitly constrain that representation onto the manifold of normal windows. So a broken channel relation can still be reconstructed with low series error. CARMA addresses this gap by training the correlation latent zcz_c to reconstruct Pearson correlations on normal windows. This forces zcz_c to the manifold of normal dependence structures. We already evaluate the conventional joint representation learning setting considered in the question as a single-latent VAE with elementwise reconstruction only (CARMA without zcz_c). Without any explicit learning of the normal-window correlation, that joint representation (CARMA without zcz_c) reaches only a VUS-PR of 0.4620 on TSB-AD-M, which empirically confirms the gap. Adding the correlation latent zcz_c and supervising it to learn that manifold increases VUS-PR to 0.4929.

Q2: What motivates separating temporal and correlation latents? Is there evidence that these representations are disentangled or complementary?

Separating temporal and correlation latents is already present in the literature, such as InterFusion [1]. However, InterFusion uses only ELBO to train its inter-metric latent, it does not train any explicit correlation structure. Literature shows that reconstruction loss only measures per-channel error, so broken correlation dependencies can still lead to low error scores. CARMA maintains a similar separation, but we condition the correlation latent zcz_c on the pearson of normal windows so that this latent learns the manifold of normal-window dependence. And then the series decoder reconstructs every timestep as X^=g(zt,zc)\hat{X}=g(z_t,z_c) conditioned on both signal and correlaiton latent. This design helps the model to detect anomalies better, as shown in the Table. Adding an unsupervised second latent slightly improves the detection to 0.4719, but supervising it with the pearson correlation improves the detection to 0.4929. We do not claim that the representations are disentangled, but we argue that the latents are complementary in the sense that they are guided by different objectives. Figure 6 in the appendix illustrates this representation in a case study. However, to address your question regarding disentangled representations, we conducted an experiment by introducing disentanglement in the encoder using depthwise convolutions. The depthwise encoder forces each channel through its own filters. Both depthwise variants fall well below the original encoder. So, guiding the latent ztz_t and zcz_c with different objectives improves detection more than imposing disentanglement in the encoder.

Table. TSB-AD-M ablation. The best result in each column is in bold.

MethodAUC-PRAUC-ROCVUS-PRVUS-ROCStandard-F1
CARMA w/o zcz_c (depthwise)0.3740 ± 0.00100.7470 ± 0.00200.4147 ± 0.00240.7755 ± 0.00200.4423 ± 0.0006
CARMA (Ours) (depthwise)0.3780 ± 0.00190.7508 ± 0.00180.4186 ± 0.00130.7792 ± 0.00190.4481 ± 0.0020
CARMA w/o zcz_c0.4180 ± 0.00470.8041 ± 0.00340.4620 ± 0.00320.8287 ± 0.00360.4936 ± 0.0053
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}0.4243 ± 0.00600.8157 ± 0.00320.4719 ± 0.00550.8424 ± 0.00310.5033 ± 0.0084
CARMA (Ours)0.4425 ± 0.00090.8330 ± 0.00170.4929 ± 0.00270.8598 ± 0.00100.5301 ± 0.0032

[1] Z. Li et al., “Multivariate Time Series Anomaly Detection and Interpretation using Hierarchical Inter-Metric and Temporal Embedding,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, in KDD ’21. New York, NY, USA: Association for Computing Machinery, Aug. 2021, pp. 3220–3230. doi: 10.1145/3447548.3467075.

W2 and Q2: Could the authors compare against a capacity-matched single-latent VAE with the same auxiliary correlation objective and comparable KL regularization?

We have compared CARMA with a capacity-matched single-latent VAE with the same auxiliary correlation objective and comparable KL regularization in the table below. The capacity-matched model achieves a VUS-PR of 0.4615. This result is close to the single-latent model (CARMA w/o zcz_c) with a VUS-PR of 0.4620 but still lower than CARMA (Ours). The results for the capacity-matched single-latent version suggest that it is unlikely that most of the improvements stem from relaxed compression constraints. The CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}} version also has a relaxed compression constraints and achieves a VUS-PR of 0.4719, still behind the CARMA supervised by the auxiliary correlation objective.

Table. TSB-AD-M ablation. CARMA (Capacity-matched) is a single-latent VAE with the same auxiliary correlation objective and comparable KL regularization. Mean ± std over three seeds. The best result in each column is in bold.

MethodAUC-PRVUS-PRVUS-ROCStandard-F1
CARMA + Pearson lag=50.3468 ± 0.00690.4042 ± 0.00570.8376 ± 0.00180.4315 ± 0.0067
CARMA (Capacity-matched)0.4135 ± 0.00080.4615 ± 0.00060.8163 ± 0.00170.4865 ± 0.0018
CARMA w/o zcz_c0.4180 ± 0.00470.4620 ± 0.00320.8287 ± 0.00360.4936 ± 0.0053
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}0.4243 ± 0.00600.4719 ± 0.00550.8424 ± 0.00310.5033 ± 0.0084
CARMA + Spearman corr0.4302 ± 0.00330.4780 ± 0.00350.8488 ± 0.00060.5102 ± 0.0030
CARMA σt,0=1\sigma_{t,0}=10.4384 ± 0.00210.4881 ± 0.00330.8575 ± 0.00660.5271 ± 0.0038
CARMA (Ours)0.4425 ± 0.00090.4929 ± 0.00270.8598 ± 0.00100.5301 ± 0.0032

Thank you for your review. We hope the replies and extra experiments address your questions. We also found this useful for understanding CARMA more clearly.

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

This paper presents CARMA, a new method for finding anomalies in time series data with multiple channels. Some anomalies break the normal relationships between channels, and older reconstruction-based methods often miss them. CARMA uses a neural network with two parts: one learns how values change over time, and the other learns how channels relate to each other. It uses a simple correlation measure (Pearson correlation) to help guide the reconstruction. Extensive experiments demonstrate its effectiveness.

Critical strengths

S1: The paper considers a clearly motivated problem. It points out an important weakness of reconstruction-based multivariate time-series anomaly detection. When an anomaly breaks the links between channels, it may not cause large reconstruction errors if each channel still stays within its normal range. Directly asking the model to represent cross-channel links is a reasonable way to deal with this problem, especially in industrial monitoring and other multivariate sensing applications.

S2: The empirical evaluation is fairly broad and includes different kinds of evidence.

Critical weaknesses

W1: The method's novelty compared with existing dependency-aware anomaly detectors is not clearly shown. The main parts of CARMA have important precedents. InterFusion already separates temporal and inter-metric representations within a hierarchical VAE, while MSCRED directly models and reconstructs inter-channel correlation structures. CARMA's more specific contribution seems to be the use of an explicitly supervised Pearson-correlation latent to condition signal reconstruction. Although this is a possibly useful design choice, the paper does not do enough to show why this particular formulation is a substantial methodological advance over these closely related approaches. In particular, InterFusion is discussed as related work but not included in the baseline comparison.

W2: The link between the correlation supervision objective and the final anomaly score is still not fully explained. CARMA directly reconstructs Pearson correlations during training, but its anomaly score at inference time is based only on signal reconstruction error. Since the correlation latent is inferred from the input window itself, it is not clear why the model cannot encode abnormal dependencies in this latent and successfully reconstruct the corresponding anomalous signal. The ablation shows that correlation supervision improves overall performance, but not fully explain how this auxiliary objective reliably turns dependency violations into larger signal reconstruction errors. This is especially important because the proposed mechanism, rather tan the general use of a second latent variable, is the paper's main methodological contribution.

Questions for authors

Q1: Why does the correlation-supervised latent make dependency-breaking anomalies harder to reconstruct? At inference time, CARMA scores anomalies using only signal reconstruction error, not correlation reconstruction error. Since the correlation latent is inferred from the same input window, what stops it from encoding abnormal correlations and letting the signal decoder reconstruct them? Can the authors give a more direct analysis of correlation reconstruction, signal reconstruction, and anomaly scores on controlled dependency-breaking events, including a comparison with the single-latent and unsupervised-correlation variants?

Q2: How much of the reported improvement comes specifically from detecting cross-channel anomalies? Can the authors report results separately for univariate, cross-channel, and mixed anomalies, or otherwise give a controlled evaluation that separates these anomaly types? On NoiseFlip and NoiseRoll, can the authors compare the full model against the correlation-branch ablations and relevant dependency-aware methods such as MSCRED, MTAD-GAT, and GDN under the same evaluation protocol?

3 comments

S1 and S2: The paper considers a clearly motivated problem. It points out an important weakness of reconstruction-based multivariate time-series anomaly detection. When an anomaly breaks the links between channels, it may not cause large reconstruction errors if each channel still stays within its normal range. Directly asking the model to represent cross-channel links is a reasonable way to deal with this problem, especially in industrial monitoring and other multivariate sensing applications. The empirical evaluation is fairly broad and includes different kinds of evidence.

We really appreciate your understanding of the limitations and addressing the usefulness of cross-channel linking modeling. Thank you for acknowledging the broad evaluation and evidence we provided.

W1 and Q1: Why does the correlation-supervised latent make dependency-breaking anomalies harder to reconstruct? At inference time, CARMA scores anomalies using only signal reconstruction error, not correlation reconstruction error. Since the correlation latent is inferred from the same input window, what stops it from encoding abnormal correlations and letting the signal decoder reconstruct them?

The correlation-supervised latent makes a dependence break harder to reconstruct because it conditions the correlation latent zcz_c to learn the normal pattern of the correlation structure, and the series decoder gg is then conditioned on that same zcz_c. The reconstructed series depends on both the signal latent and the correlation latent X^=g(zt,zc)\hat{X}=g(z_t,z_c). When channel relations break, zcz_c is still a representation shaped on normal windows, so the decoder tends to create a series whose correlations do not match the broken window. That mismatch is measured as signal reconstruction error, which is the anomaly score. We score from the signal reconstruction, because we already condition the series decoder gg on both ztz_t and zcz_c. LcorrL_{\mathrm{corr}} stays a training objective that explicitly shapes zcz_c toward the window’s Pearson correlation distribution.

The correlation latent zcz_c is inferred from the same input window, but seeing the window does not always mean the abnormal dependence is encoded in a form the series decoder can reconstruct with low error. The model is trained only on normal data, limiting its ability to encode these specific patterns effectively. The correlation reconstruction loss LcorrL_{\mathrm{corr}} and the Kullback-Leibler (KL) regularizer keep the correlation latent zcz_c on a compact set of normal Pearson structure. The series decoder gg, trained only on these normal representations, produces X^=g(zt,zc)\hat{X}=g(z_t,z_c) conditioned on both latents. If zcz_c deviates from the training manifold, the decoder has never learned to map that out-of-distribution representation to a valid series, thus producing an error.

Additionally, we evaluated InterFusion on the TSBAD-M dataset. InterFusion achieves a VUS-PR of 0.297 and a Standard-F1 of 0.363, compared with CARMA's 0.4929 and 0.5301, respectively. We have included it in the revised manuscript.

Q1: Can the authors give a more direct analysis of correlation reconstruction, signal reconstruction, and anomaly scores on controlled dependency-breaking events, including a comparison with the single-latent and unsupervised-correlation variants?

We report a direct comparison of signal reconstruction, correlation reconstruction, and anomaly scores on controlled dependency-breaking events. We include CARMA, the single-latent variant (CARMA w/o zcz_c), and the unsupervised-correlation variant (CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}). At inference we score anomalies using the signal reconstruction error. All three variants show a large gap between normal and anomalous signal error, and CARMA is best on the detection metrics.

Table. Signal reconstruction error.

VariantNormalAnomalousAnomalous / Normal
CARMA0.0009470.668702706.36
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}0.0012000.705607588.24
CARMA w/o zcz_c0.0013290.751194565.04

Table. Correlation (Lcorr\mathcal{L}_{\mathrm{corr}}). "--" means the variant has no correlation head.

VariantNormalAnomalousAnomalous / Normal
CARMA0.0008460.210746248.97
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}------
CARMA w/o zcz_c------

Table. Anomaly detection scores.

VariantAUC-PRAUC-ROCVUS-PRVUS-ROCStandard-F1
CARMA0.97700.99670.98750.99840.9165
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}0.97620.99660.98700.99830.9159
CARMA w/o zcz_c0.97560.99650.98650.99820.9146

W2 and Q2: How much of the reported improvement comes specifically from detecting cross-channel anomalies? Can the authors report results separately for univariate, cross-channel, and mixed anomalies, or otherwise give a controlled evaluation that separates these anomaly types? On NoiseFlip and NoiseRoll, can the authors compare the full model against the correlation-branch ablations and relevant dependency-aware methods such as MSCRED, MTAD-GAT, and GDN under the same evaluation protocol?

To separate improvements coming from different anomaly types, we applied anomaly category analysis developed by Pinet et al. to the TSB-AD-M dataset [1]. The analysis shows each series contains a mix of univariate and cross-channel anomalies. The Pinets analysis marks which segments are univariate or mixed-type anomalies. We then label a series as univariate or mixed-type when most of the anomalous length belongs to that type. The results show that CARMA outperforms the baselines on both univariate and mixed anomaly types. To show the performance gains from cross-channel-only anomalies, we tested the methods on the NoiseFlip and NoiseRoll datasets, where again CARMA outperforms the baselines including MSCRED, MTAD-GAT, and GDN.

Table. TSB-AD-M series with univariate-only anomalies. Mean ± std over three seeds. The best result in each column is in bold.

MethodAUC-PRAUC-ROCVUS-PRVUS-ROCStandard-F1
CARMA0.3889 ± 0.00470.7994 ± 0.00660.3886 ± 0.00730.8210 ± 0.00530.4776 ± 0.0062
HYDRA0.3537 ± 0.00060.7044 ± 0.00070.3439 ± 0.00080.7348 ± 0.00050.3870 ± 0.0016
StreamVAE0.3075 ± 0.00600.7226 ± 0.00070.2990 ± 0.00640.7408 ± 0.00150.3640 ± 0.0033
MSCRED0.1737 ± 0.00160.6274 ± 0.00150.1605 ± 0.00390.6572 ± 0.00130.2179 ± 0.0029
GDN0.1512 ± 0.00660.6101 ± 0.00230.1419 ± 0.00470.6510 ± 0.00270.1953 ± 0.0086
MTAD_GAT0.1321 ± 0.00320.6147 ± 0.00280.1243 ± 0.00280.6558 ± 0.00320.1800 ± 0.0047

Table. TSB-AD-M series with mixed univariate and cross-channel anomalies. Mean ± std over three seeds. The best result in each column is in bold.

MethodAUC-PRAUC-ROCVUS-PRVUS-ROCStandard-F1
CARMA0.5378 ± 0.00260.8903 ± 0.00160.5561 ± 0.00080.8911 ± 0.00190.6196 ± 0.0014
StreamVAE0.5045 ± 0.00240.8394 ± 0.00400.5119 ± 0.00420.8195 ± 0.00330.5793 ± 0.0033
MSCRED0.5359 ± 0.00070.8100 ± 0.00050.4948 ± 0.00040.7998 ± 0.00080.5843 ± 0.0002
HYDRA0.4791 ± 0.01080.8428 ± 0.00640.4937 ± 0.00980.8544 ± 0.00640.5480 ± 0.0047
MTAD_GAT0.4855 ± 0.00110.8081 ± 0.00190.4618 ± 0.00260.8217 ± 0.00160.5357 ± 0.0017
GDN0.4439 ± 0.01050.7981 ± 0.00850.4277 ± 0.00970.8174 ± 0.00680.5078 ± 0.0085

Table. Mean results on NoiseFlip and NoiseRoll. The best result in each column is in bold.

MethodAUC-PRAUC-ROCVUS-PRVUS-ROCStandard-F1
CARMA0.9768 ± 0.00040.9968 ± 0.00010.9888 ± 0.00030.9986 ± 0.00000.9259 ± 0.0008
CARMA w/o Lcorr\mathcal{L}_{\mathrm{corr}}0.9765 ± 0.00030.9968 ± 0.00000.9885 ± 0.00020.9985 ± 0.00000.9256 ± 0.0007
CARMA w/o zcz_c0.9762 ± 0.00070.9967 ± 0.00010.9883 ± 0.00040.9985 ± 0.00010.9255 ± 0.0011
StreamVAE0.9716 ± 0.00040.9962 ± 0.00000.9858 ± 0.00020.9982 ± 0.00000.9245 ± 0.0005
MSCRED0.8904 ± 0.04780.9595 ± 0.01620.9037 ± 0.04940.9696 ± 0.01550.8406 ± 0.0505
HYDRA0.8677 ± 0.00200.9379 ± 0.00140.8675 ± 0.00250.9496 ± 0.00150.8176 ± 0.0016
GDN0.4874 ± 0.13160.7136 ± 0.04080.5050 ± 0.12250.7383 ± 0.03400.5425 ± 0.1309
MTAD_GAT0.3044 ± 0.00010.6761 ± 0.00010.3215 ± 0.00020.7028 ± 0.00010.3969 ± 0.0003

[1] M. Pinet, J. Cumin, S. Berlemont, and D. Vaufreydaz, “Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate,” Jun. 15, 2026, arXiv: arXiv:2606.02670. doi: 10.48550/arXiv.2606.02670.

Thank you for reviewing our paper. Your comments were useful in understanding the method better. We hope the replies and additional experiments address your questions.

Comments on the paper

2 comments

Updated Manuscript: Summary of Revisions and Additional Experiments

Dear Area Chair and Reviewers,

We sincerely appreciate the time and effort you have dedicated to reviewing our submission and providing such valuable feedback. We have carefully revised the manuscript in response to your comments and have uploaded the updated PDF. Below is a summary of the changes we made to address the reviewers’ comments.

Reviewer DiqP We included InterFusion in the baseline comparison, increasing the total from 29 to 30 baselines. We updated Tables 1, 4, and 5 and the critical difference diagram. We also updated the baseline count in the abstract (ln. 028), contributions (ln. 120), results (ln. 370), and conclusion (ln. 536). CARMA remains well ahead of InterFusion (VUS-PR 0.4929 vs. 0.297), showing the benefit of explicitly supervising the correlation latent and feeding it into the signal decoder.

Reviewer Vaga We added two additional variants (depthwise encoder and capacity-matched single latent) to Table 3. We also updated the Ablation Variants (ln. 473--477) and Component Contribution (ln. 484--514) sections accordingly. Both variants fall below CARMA, showing that a supervised correlation latent improves detection compared to a disentangled encoder or a capacity-matched single latent.

Reviewer 9cjJ We added a complexity comparison across all 30 baselines in Appendix Figure 7, and we added the scaling-cost results in Table 6. CARMA is computationally competitive with the second-best baseline (HYDRA) while remaining stronger in detection, and it is both more accurate and faster than the third-best baseline (StreamVAE).

Additional We moved some text from the synthetic benchmark (ln. 331--338) section to the appendix (ln. 917--922) to stay within the 10-page limit.

We hope you will have time to read our responses and the revised manuscript. We have put substantial time and care into this work, and we would be grateful for any further comments or suggestions.

Thank you again for your time and effort.

Best wishes, The Authors

Dear Area Chair and Reviewers,

We would really appreciate your feedback. We have addressed all of your comments and would be happy to address any further suggestions you may have.

Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.