- 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
- 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.
- The manuscript also includes sufficient baseline comparisons and detailed ablation studies.
- The authors have uploaded their code to an anonymous GitHub repository.
Critical weaknesses
- 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.
- 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?
- 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.
- 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
- See the weaknesses.
- 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?
- 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
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.
| Variant | GECCO | MSL | PSM |
|---|---|---|---|
| TEDA (original OT) | 0.6340 ± 0.0052 | 0.3176 ± 0.0005 | 0.5979 ± 0.0033 |
| OT with compatible marginals | 0.6264 ± 0.0054 | 0.3174 ± 0.0004 | 0.5978 ± 0.0035 |
| Softmax assignment | 0.6034 ± 0.0068 | 0.3103 ± 0.0015 | 0.5456 ± 0.0151 |
| Nearest-centroid assignment | 0.6082 ± 0.0076 | 0.3121 ± 0.0013 | 0.5423 ± 0.0118 |
| Reconstruction only | 0.6125 ± 0.0018 | 0.3133 ± 0.0002 | 0.5522 ± 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 differs from the number of centroids . The observed plans have row sums close to and column sums close to 1. We therefore tested explicit targets of and , which give the same total mass on both sides. This control yielded similar mean VUS-PR on all three datasets.
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 representation | GECCO | MSL | PSM |
|---|---|---|---|
| Complex (real and imaginary) | 0.6340 ± 0.0052 | 0.3176 ± 0.0005 | 0.5979 ± 0.0033 |
| Magnitude | 0.5974 ± 0.0053 | 0.3124 ± 0.0012 | 0.5235 ± 0.0041 |
| Log-magnitude | 0.5977 ± 0.0061 | 0.3117 ± 0.0013 | 0.5201 ± 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.
| Dataset | Original | Circular shift by | Hann taper in FFT branch |
|---|---|---|---|
| GECCO | 0.6340 ± 0.0052 | 0.6265 ± 0.0020 | 0.6333 ± 0.0045 |
| MSL | 0.3176 ± 0.0005 | 0.3076 ± 0.0009 | 0.3181 ± 0.0009 |
| PSM | 0.5979 ± 0.0033 | 0.5793 ± 0.0037 | 0.5957 ± 0.0034 |
Both tables report mean ± sample standard deviation over three seeds.
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.
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.
Table R3. Efficiency comparison of different methods on the PSM dataset.
References: [1] Xiao, Chunjing, et al. "Imputation-based time-series anomaly detection with conditional weight-incremental diffusion models." KDD 2023.