acceptodds
Under review as a conference paper at ICLR 2027

TEDA: Time-Frequency Entropy-Aware Distribution Modeling for Time Series Anomaly Detection

Abstract

Time series anomaly detection plays an important role in many real-world systems. Existing advanced unsupervised methods mainly rely on prediction or reconstruction errors to identify abnormal patterns. However, such a paradigm only focuses on sample-wise errors and fails to capture whether a sample conforms to the distribution of whole data. As a result, simple anomalies but with low-probability may be well reconstructed, while complex but normal patterns may be failed. In this paper, we propose TEDA, a time-frequency distribution-aware framework for unsupervised time series anomaly detection. TEDA learns distributional centroids and introduces optimal transport into the reconstruction-based paradigm to model normal patterns in the latent distributional space. Specifically, a patch-based branch captures local temporal distributions, while a frequency-domain branch models long-term periodicity, trends, and global distributions. To improve the stability and discriminative ability of optimal transport, TEDA uses entropy regularization to preserve smooth transport plans during training, and removes it to perform sharp assignment during inference. The final anomaly score combines reconstruction errors and distribution-aware optimal transport costs. Extensive experiments on six real-world datasets show that TEDA achieves better performance compared with classical and recent state-of-the-art methods. The results demonstrate the effectiveness of enhancing time series anomaly detection with the distribution of whole data.

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

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZUCBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TEDA, an unsupervised time-series anomaly detection framework that augments reconstruction-based models with distribution awareness in both time and frequency domains. TEDA learns learnable centroids for local temporal patches and frequency tokens and uses optimal transport (OT) to align sample representations to these centroids, employing entropy-regularized (smooth) transport during training for stability and sharp nearest-centroid assignment at inference for discriminative scoring. The final anomaly score fuses point-wise reconstruction error with distribution-aware transport costs, and experiments on six datasets report consistent improvements over a broad set of baselines.

Critical strengths
  1. This work introduces a dual time–frequency distribution modeling approach with separate centroid sets (temporal patches and FFT tokens), which is overall novel to my konwledge.
  2. The manuscript also includes sufficient baseline comparisons and detailed ablation studies.
  3. The authors have uploaded their code to an anonymous GitHub repository.
Critical weaknesses
  1. First, I believe the paper has serious issues with its research motivation. Specifically, the authors observe that some anomalies are difficult to detect through reconstruction and therefore propose a new distribution-based detection approach. While this reasoning is generally sound, the authors selectively overlook a substantial body of existing work on generative anomaly detection. A brief search was sufficient to identify numerous papers in this area, as listed in [1]–[6]. Therefore, the research problem itself is not novel and may already be a relatively mature research direction.
  2. Second, the method design also raises concerns in light of the aforementioned motivation. The role of OT in TEDA appears to be more akin to that of a non-linear activation function. If it introduces any distributional bias, that bias is imposed on the learned representations, making its effect difficult to interpret. I suggest that the authors visualize how these matrices are learned on real-world datasets and use tools such as t-SNE to directly and concisely demonstrate the effect of distribution shaping. Furthermore, I notice that the two OT modules yield only marginal improvements in the ablation study (Table 3), especially compared with some less central design choices. Are these modules genuinely effective, or do they merely introduce additional parameters?
  3. Channel-independent modeling may miss cross-variable (contextual) anomalies and inter-sensor dependencies critical in multivariate settings (e.g., SWaT/PSM); the approach models each channel separately and aggregates scores afterward. The frequency branch uses raw real/imag FFT features; sensitivity to phase, stationarity assumptions, and windowing choices is not fully explored. Magnitude vs phase contributions and robustness to phase shifts/drifts are unclear.
  4. For evaluations, missing comparisons with several strong recent generative families. No computation/runtime analysis; dual OT modules and centroid tables per window may be heavy during training. Memory/time complexity vs baselines is not reported. Additionally, could the authors report standard deviations alongside the tabulated results to improve the rigor and credibility of their experimental evaluation? This is particularly relevant to Table 5 in the appendix, which has substantial room for increasing its information density.

References:
[1] Wang, Huazhang, et al. "ECGGAN: A framework for effective and interpretable electrocardiogram anomaly detection." KDD 2023.
[2] Xiao, Chunjing, et al. "Imputation-based time-series anomaly detection with conditional weight-incremental diffusion models." KDD 2023.
[3] Hu, Rongyao, et al. "Unsupervised anomaly detection for multivariate time series using diffusion model." ICASSP 2024.
[4] Cai, Saihua, et al. "DDP-DAR: Network intrusion detection based on denoising diffusion probabilistic model and dual-attention residual network." Neural Networks 2025.
[5] Abdulganiyu, Oluwadamilare Harazeem, et al. "XIDINTFL-VAE: XGBoost-based intrusion detection of imbalance network traffic via class-wise focal loss variational autoencoder." The Journal of Supercomputing 2025.
[6] Qiao, Yan, et al. "ImConDM: Accurate MTS Anomaly Detection Integrating Imputation With Conditional Diffusion Models." IEEE TIFS 2026.

Questions for authors
  1. See the weaknesses.
  2. How are the OT marginals (μ, ν) chosen in practice for each branch? Are they uniform over source tokens and centroids, learned, or data-dependent? Did you try unbalanced/partial OT?
  3. How sensitive is TEDA to phase shifts or frequency leakage (e.g., window boundaries)? Did you try magnitude-only spectra, log-magnitude scaling, band selection, or learnable spectral filters?
3 comments
AuthorsOpenReview ↗

Response to W1 & W4 — Motivation and Generative Baselines

Thanks for your valuable suggestion. We agree that generative time-series anomaly detection is well established. TEDA focuses on a specific case: a window may be reconstructed accurately even when its local waveform or spectrum deviates from typical patterns in the normal training data. TEDA therefore uses optimal transport to learn separate sets of normal prototypes in the time and frequency domains and combines nearest-prototype distances with reconstruction error. Deviations from these prototypes provide an additional signal for anomaly scoring even when reconstruction error is small.

We evaluated two additional generative baselines on three datasets: DiffAD[1] and Donut (a VAE adapted for multivariate inputs). The table below reports the mean and sample standard deviation over three random seeds (VUS-PR / AUC-ROC).

Table R1. Generative baselines.

MethodGECCOMSLPSM
TEDA0.6340±0.0052 / 0.9872±0.00400.3176±0.0005 / 0.7839±0.00070.5979±0.0033 / 0.7216±0.0044
DiffAD0.0744±0.0029 / 0.7735±0.00410.2466±0.0027 / 0.7074±0.00250.4394±0.0092 / 0.6467±0.0065
Donut0.0483±0.0230 / 0.7510±0.07490.1783±0.0005 / 0.5149±0.00030.4899±0.0355 / 0.7109±0.0141

TEDA achieved the highest mean results on all three datasets. All methods used the same data splits and were evaluated at the same time points. MSL retained one channel. The input window lengths were 128 for DiffAD and 120 for Donut.


Computational Cost. We conducted additional experiments to analyze computational efficiency on the MSL and PSM datasets.

Table R2. Efficiency comparison of different methods on the MSL dataset.

MethodTraining Time (s)Inference Time (s)Total Params (M)VUS-PR
ModernTCN22.101.310.010.2779
CATCH44.594.5348.540.3132
CrossAD65.9715.951.330.3133
TEDA52.4216.130.940.3176

Table R3. Efficiency comparison of different methods on the PSM dataset.

MethodTraining Time (s)Inference Time (s)Total Params (M)VUS-PR
ModernTCN83.552.670.440.5223
CATCH1066.04189.26210.880.5066
CrossAD1707.9122.581.330.5604
TEDA1845.1522.950.940.5979

References: [1] Xiao, Chunjing, et al. "Imputation-based time-series anomaly detection with conditional weight-incremental diffusion models." KDD 2023.

AuthorsOpenReview ↗

Response to W2 & Q2 — Role of OT and Choice of Marginals

Role of OT. The training transport plan couples the assignments of all tokens in a window. Each token’s assignment depends on both its distances to the centroids and how the other tokens are assigned. To isolate the effect of this matching rule from parameter count, we kept the backbone and both centroid sets fixed and replaced OT with per-token softmax or nearest-centroid assignment. We also report a reconstruction-only reference. The table shows VUS-PR as mean ± sample SD over three seeds.

Table R4.

VariantGECCOMSLPSM
TEDA (original OT)0.6340 ± 0.00520.3176 ± 0.00050.5979 ± 0.0033
OT with compatible marginals0.6264 ± 0.00540.3174 ± 0.00040.5978 ± 0.0035
Softmax assignment0.6034 ± 0.00680.3103 ± 0.00150.5456 ± 0.0151
Nearest-centroid assignment0.6082 ± 0.00760.3121 ± 0.00130.5423 ± 0.0118
Reconstruction only0.6125 ± 0.00180.3133 ± 0.00020.5522 ± 0.0059

As shown in the table, TEDA with the original OT design achieves the highest mean VUS-PR on all three datasets, suggesting that the OT design is both effective and necessary.

We also visualized the soft transport plans and nearest-centroid usage. In one GECCO run, all sampled temporal tokens selected the same nearest centroid out of 32, showing that balanced transport mass during training can coexist with concentrated centroid usage at inference.


Choice of marginals. Both branches use fixed, uniform targets. The submitted implementation sets each row and column target to 1, giving unequal total masses when the number of tokens NN differs from the number of centroids KK. The observed plans have row sums close to K/NK/N and column sums close to 1. We therefore tested explicit targets of μi=K/N\mu_i=K/N and νj=1\nu_j=1, which give the same total mass KK on both sides. This control yielded similar mean VUS-PR on all three datasets.

AuthorsOpenReview ↗

Response to W3 & Q3 — Channel Independence and Spectral Robustness

Channel independence. We use channel-independent modeling because anomalies in a multi-sensor system can manifest as changes in the local waveform or spectrum of an individual sensor. Sharing the encoder and centroids across channels lets TEDA learn recurring patterns across sensors while scoring each sensor’s deviation from those patterns. We then average the channel scores at each time step. Across five random seeds, TEDA achieved AUC-ROCs of 0.9872 ± 0.0039 on GECCO (9 channels) and 0.8484 ± 0.0028 on SWAT (31 channels). These results show that within-channel temporal and spectral patterns provide an effective detection signal on both real-world multi-sensor datasets.


Magnitude and phase. TEDA computes the FFT separately for each input window. The real and imaginary components preserve phase information, while magnitude-only features discard it. We trained models with complex, magnitude-only, and log-magnitude inputs. Complex inputs achieved the highest mean VUS-PR on three datasets.

Table R5. VUS-PR with different spectral representations.

Spectral representationGECCOMSLPSM
Complex (real and imaginary)0.6340 ± 0.00520.3176 ± 0.00050.5979 ± 0.0033
Magnitude0.5974 ± 0.00530.3124 ± 0.00120.5235 ± 0.0041
Log-magnitude0.5977 ± 0.00610.3117 ± 0.00130.5201 ± 0.0055

Phase shifts and window boundaries. At inference, we circularly shifted each input window by one-eighth of its length and aligned the scores back to the original timestamps. This reduced mean VUS-PR on all three datasets. Circular shifts also change the placement of window boundaries. To test sensitivity to tapering, we separately applied a Hann window to the FFT branch at inference. The resulting mean VUS-PR changed by less than 0.0023 on each dataset. Natural phase drift, frequency-band selection, and learnable spectral filters remain untested.

Table R6. VUS-PR under circular shifts and hann tapering.

DatasetOriginalCircular shift by W/8W/8Hann taper in FFT branch
GECCO0.6340 ± 0.00520.6265 ± 0.00200.6333 ± 0.0045
MSL0.3176 ± 0.00050.3076 ± 0.00090.3181 ± 0.0009
PSM0.5979 ± 0.00330.5793 ± 0.00370.5957 ± 0.0034

Both tables report mean ± sample standard deviation over three seeds.

Reviewer_SuKTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes TEDA for unsupervised time series anomaly detection. It argues that reconstruction error cannot capture distributional abnormality, which leads to false negatives on simple anomalies and false positives on complex normal patterns. TEDA has a temporal branch (patch-based Transformer encoder) and a frequency branch (rFFT tokens). Each branch matches its tokens to learnable centroids via optimal transport, and the resulting representations are fed to a reconstruction decoder. Training uses entropy-regularized (Sinkhorn) transport plans. At inference, the nearest-centroid distance is added to the reconstruction error to form the anomaly score.

Critical strengths
  • The evaluation covers six datasets and 18 baselines, including recent methods (CrossAD, CATCH, PaAno, ALoRa). It reports threshold-independent metrics such as VUS-PR and range-based AUC. This makes the comparison comprehensive in coverage, although the significance of the reported gains is questionable.
  • Adding a centroid-distance term to the reconstruction score improves over the reconstruction-only variant in the ablation. This suggests a lightweight and practical modification.
Critical weaknesses
  • Reconstruction-based methods can already detect distributional abnormality. A model trained only on normal data reconstructs well what lies in the normal distribution. It is also the simplest approach, so it is unclear that this is a distinct problem.
  • “Rarely observed in normal data” is essentially the definition of an anomaly. If such patterns appear in normal data, can they be called anomalies?
  • If a model reconstructs a commonly observed complex normal pattern as badly as in Fig. 1, it is simply a poor model. A reasonable model should follow the periodicity of the series. This is not a limitation of the reconstruction paradigm.
  • Why OT is needed is never justified in the problem statement. OT and the centroids have nothing to do with the heterogeneity challenge. Only the dual-branch design addresses it. That design is common in prior work you already cited (e.g., Wu et al., 2022).
  • The method feeds OT-induced representations into the decoder to help reconstruction, so simple anomalies may be reconstructed even better, worsening the FN problem.
  • The additive score Srec+βSOTS_{\text{rec}} + \beta S_{\text{OT}} cannot reduce the large SrecS_{\text{rec}} of complex normal samples, so the FP case is not addressed.
  • The authors state below Eq. (11) that Γsharpb\Gamma^{b}_{\mathrm{sharp}} is not an OT solution, so inference is just the nearest-centroid distance.
  • The OT source is the set of tokens within a single window. OT therefore matches the within-window distribution, not the distribution of normal data. No standard deviations are reported despite five seeds, and most margins are 0.001–0.01.
  • Performance values in Table 1 and Fig. 3 are inconsistent.
  • No ablation replaces OT with simpler assignment methods, and prototype/memory-based baselines (MemAE, Deep SVDD) are missing.
  • Overall, the method reduces to a nearest-centroid distance added to the reconstruction error, and its empirical gains are not established.
Questions for authors

See the weakness.

Reviewer_xpXQOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This study proposes a model for time-series anomaly detection. The main idea is to learn a set of centroid vectors that characterize normal patterns and detect anomalies by measuring deviations from these learned representations. To capture normal behavior comprehensively, the model considers two complementary views of the input time series: the time domain and the frequency domain. In the time domain, historical windows are divided into local patches to capture short-term temporal patterns and dependencies. In the frequency domain, the input windows are transformed using the Fast Fourier Transform (FFT) to capture longer-term periodicities and global temporal patterns. Each branch learns its own set of centroids, representing typical normal behavior in the corresponding domain. During training, these centroids are jointly optimized with the rest of the model using an optimal transport (OT)-based objective, encouraging them to become representative of normal temporal and frequency patterns.

Critical strengths
  • S1. The experimental evaluation is comprehensive, covering multiple datasets and evaluation metrics. The paper also includes ablation studies and sensitivity analyses of important hyperparameters, which provide additional evidence for the effectiveness of the proposed method.
  • S2. The paper is well organized.
Critical weaknesses
  • W1. The use of both patch-based temporal representations and frequency-domain representations has become increasingly common in time-series modeling, particularly in tasks such as anomaly detection and forecasting.

  • W2. The concept of learning trainable centroid vectors to represent normal behavior has been explored in previous studies. For instance, the MEMTO [1] model adopts a similar approach by learning centroid vectors, referred to as memory items, to capture normal patterns for anomaly detection. From a conceptual perspective, the proposed TEDA method shares similarities with MEMTO, as both methods rely on learnable centroid vectors to model normal behavior. However, the key distinction of TEDA lies in its additional use of centroid vectors learned from frequency-domain representations, enabling the modeling of normal patterns in both the time and frequency domains. Another difference is the incorporation of optimal transport for learning the centroids. MEMTO is not discussed in the Related Work section. I suggest that the authors include MEMTO in their discussion of related studies and explicitly clarify how their proposed method differs from it, particularly in terms of methodological novelty and technical contributions.

  • W3. Use of channel-independent modeling: This is a concern, as real-world systems often exhibit complex behaviors that require modeling both temporal and cross-variable dependencies to accurately capture their normal dynamics. Indeed, multivariate modeling has become well established in statistics and mathematics precisely because it can capture relationships that univariate approaches may overlook. I encourage the authors to justify this choice and discuss its potential limitations, particularly for detecting anomalies arising from changes in cross-variable relationships.

  • W4. Using fixed window sizes across all compared methods may not be the most appropriate approach for ensuring a fair evaluation. Each method is designed with its own characteristics and assumptions, and not all methods necessarily require the same window size. For a fair and meaningful comparison, the evaluation should follow the official implementation details and recommended configurations described in the original papers and their publicly available repositories.

  • W5. The authors should report the standard deviation of the results presented in Tables 1 and 2.

  • W6. The computational efficiency analysis in Appendix C.3 (Table 6) is limited to the GECCO dataset. I would like to see this analysis extended to additional datasets with different dimensionalities to better assess the computational cost and scalability of the proposed method.

References

[1] Song et al. (2023). “MEMTO: Memory-guided Transformer for Multivariate Time Series Anomaly Detection.” Advances in Neural Information Processing Systems (NeurIPS 2023).

Questions for authors

Minor comments:

  • I suggest that the authors consider changing the abbreviation “Std-F1” used for standard F1-score. Since “std” is commonly used to denote standard deviation, this notation may be slightly confusing to readers. Using simply “F1” or another less ambiguous abbreviation could improve clarity.
5 comments
AuthorsOpenReview ↗

Response to W1 — Contribution

Thanks for your valuable suggestion. We agree that combining temporal patches with frequency-domain representations is common in time-series modeling. TEDA’s contribution lies in how it uses optimal transport to learn a set of centroids representing normal patterns in each domain. During training, temporal patches and frequency tokens from each window are matched to their respective centroids, which are shared across training windows. At inference, TEDA combines the distance to the nearest centroid with reconstruction error. A pattern can therefore receive an additional anomaly score when it deviates from the learned normal patterns, even if its reconstruction error is small.

AuthorsOpenReview ↗

Response to W2 — Comparison with MEMTO

Thanks for your valuable suggestion. We will cite MEMTO (Song et al., NeurIPS 2023). MEMTO uses a learned gate to update a single memory bank. TEDA uses entropy-regularized optimal transport to jointly match tokens within each window to separate sets of temporal and spectral centroids. At inference, MEMTO uses memory distance to weight reconstruction error, while TEDA adds a distance-based deviation term to the reconstruction error.

We compared both methods under the same data protocol using three seeds. Results are mean ± sample standard deviation.

Table R1. Anomaly Detection Performance of TEDA and MEMTO.

DatasetTEDA VUS-PRTEDA AUC-ROCMEMTO VUS-PRMEMTO AUC-ROC
GECCO0.6340 ± 0.00520.9872 ± 0.00400.0243 ± 0.00300.5152 ± 0.0513
MSL0.3176 ± 0.00050.7839 ± 0.00070.1291 ± 0.00190.4313 ± 0.0489
PSM0.5979 ± 0.00330.7216 ± 0.00440.3058 ± 0.00030.4983 ± 0.0028

TEDA achieved higher mean VUS-PR and AUC-ROC on all three datasets.

AuthorsOpenReview ↗

Response to W3 — About Channel-independent

Thanks for your insightful suggestion. We use channel-independent modeling because anomalies in a multi-sensor system can manifest as changes in the local waveform or spectrum of an individual sensor. Sharing the encoder and centroids across channels lets TEDA learn recurring patterns across sensors while scoring each sensor’s deviation from those patterns. We then average the channel scores at each time step. Across five random seeds, TEDA achieved AUC-ROCs of 0.9872 ± 0.0039 on GECCO (9 channels) and 0.8484 ± 0.0028 on SWAT (31 channels). These results show that within-channel temporal and spectral patterns provide an effective detection signal on both real-world multi-sensor datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.