- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies cross-domain multivariate time series anomaly detection. The source domain is labeled, the target domain is unlabeled, and the two domains have different variable sets. MNP-GDA learns a Gumbel-Softmax adjacency per domain, encodes windows with a graph convolution branch and a block-causal attention branch, and projects both domains into a shared space. A bank of K normal prototypes, initialized by K-means on source normals, is used to pull normal samples toward their nearest prototype and push anomalies away from all prototypes. Target pseudo-labels come from a score combining reconstruction error, forecasting error and prototype distance, and each pseudo-labeled sample is weighted by its normalized score. Two theorems state that more prototypes cannot increase the nearest-prototype distance and that reliability weighting tightens a pseudo-label error bound under a covariance condition. On 8 transfer tasks over SMD, PSM, JD1, SWaT and WADI, the method improves average F1, AUPR and AUROC over VANDA by 6.60, 9.48 and 5.93 points.
Critical strengths
Strength 1: The gains in Table 1 are large and consistent. MNP-GDA has the best AUROC on all 8 tasks (80.16 average against 74.23 for VANDA), with 5 seeds and standard deviations reported. If the protocol concerns below are resolved, this is a clear empirical improvement in a setting that has few methods so far.
Strength 2: The design is reasonable and the ablations support most components. Removing multiple prototypes, either prototype loss, or the reliability weights lowers performance on the three ablation tasks (Table 2). Figure 5 also checks that the reliability-weighted selection precision exceeds the unweighted one for pseudo-anomalies. This is a useful empirical check of the condition behind Theorem 2.
Critical weaknesses
Weakness 1: The evaluation protocol is not fully specified, and it may use target labels. The paper does not say how the F1 threshold is chosen. Table 4 lists different hyperparameters for each task (learning rate, dimensions, K, temperature, selection ratios, loss weights), but there is no description of how they were chosen without target labels. The sensitivity study in Appendix E reports target-domain F1, AUPR and AUROC for each setting, and the chosen values match the best target scores (for example, tau = 0.1 gives the best target AUROC and AUPR). In unsupervised domain adaptation, tuning on target test labels can explain a large part of the gap, so this needs to be clarified before the numbers can be trusted.
Weakness 2: The baseline setup favours the proposed method. The single-domain detectors (OmniAnomaly, THOC, MTAD-GAT, HSTVAE, MTGFlow) are unsupervised, but here they are trained on the source domain and tested on the target through linear projections. This gives near-random AUROC (about 55 to 61 on most tasks). The natural baseline is to train them on the unlabeled target training data directly, which also tests whether transfer helps at all. It is also not stated whether the baselines received the same per-task tuning as MNP-GDA. Finally, the theory adds little. Theorem 1 is the standard result that a K-means-type objective does not increase with K, and Theorem 2 assumes the key condition (that weights are negatively correlated with pseudo-label error) instead of deriving it.
Questions for authors
Clarification requests:
- How is the F1 threshold selected? Do you use the best F1 over all thresholds, a fixed percentile, or point adjustment?
- How were the per-task hyperparameters in Table 4 selected? If target labels were used, what are the results with one fixed configuration for all 8 tasks?
- What do the unsupervised baselines (for example MTGFlow) and MNP-GDA without the source domain achieve when trained only on target unlabeled data?
- Table 2 covers only 3 of the 8 tasks and reports single values. Are these averages over seeds? Some differences, for example the 0.88-point F1 drop on WADI to SWaT without reliability weighting, may be within seed variance.
Suggestions for future work:
Please write the gains as percentage points instead of percent, and note that DACAD has a higher F1 on WADI to SWaT (74.90 vs 73.90). A comparison against recent unsupervised MTS-AD methods trained on the target would make the case for transfer much stronger.
5 comments
Additional comparison with Target-only unsupervised methods (W2+Q3). We agree that comparison with the unsupervised methods directly trained on the unlabeled target data can test whether transfer helps at all. We have conducted new experiments to evaluate the performance of the five unsupervised baselines (i.e., OmniAnomaly, THOC, MTAD-GAT, HSTVAE, and MTGFlow), and also our MNP-GDA, trained only on the target unlabeled data. From the results, we have two key findings:
1) For unsupervised baselines, target-only learning (i.e. training on unlabeled target-domain data alone) generally outperforms direct source transfer (i.e. directly applying the model trained on the labeled source domain to unlabeled target domain). This may be attributed to the fact that these single-domain unsupervised methods lack domain adaptation mechanism to explicitly mitigate domain discrepancy, thus easily resulting in negative transfer.
2) In contrast, the full MNP-GDA consistently outperforms both its target-only variant and all the five target-only unsupervised baselines. This highlights the benefit of transferring labeled source-domain knowledge through effective domain adaptation, beyond learning from unlabeled target data alone. We will include these additional comparisons with target-only unsupervised baselines in the revised manuscript.
Comparison between target-only learning and cross-domain transfer learning on three representative tasks.
| Task | Metric | Setting | OmniAnomaly | THOC | MTAD-GAT | HSTVAE | MTGFlow | MNP-GDA (Ours) |
|---|---|---|---|---|---|---|---|---|
| SMD→PSM | F1 | Target-only | 28.37±2.64 | 26.74±3.82 | 34.18±2.19 | 25.62±1.37 | 23.81±1.76 | 35.29±3.41 |
| Transfer | 22.96±0.19 | 25.19±4.56 | 23.00±0.19 | 23.17±0.18 | 21.01±0.19 | 38.41±2.47 | ||
| AUPR | Target-only | 21.49±2.18 | 19.63±2.91 | 25.82±2.07 | 18.74±1.63 | 17.29±3.82 | 27.16±3.28 | |
| Transfer | 12.92±0.40 | 16.78±8.87 | 12.93±0.31 | 13.74±0.27 | 09.21±0.81 | 29.37±3.09 | ||
| AUROC | Target-only | 69.82±2.73 | 67.45±3.46 | 76.53±1.84 | 61.85±1.14 | 48.23±1.24 | 58.98±2.92 | |
| Transfer | 56.07±0.90 | 60.57±3.66 | 56.21±0.77 | 57.70±0.77 | 59.22±4.90 | 80.10±2.30 | ||
| PSM→SMD | F1 | Target-only | 20.73±3.56 | 18.63±4.25 | 23.48±2.63 | 17.29±1.84 | 19.15±2.10 | 24.85±3.68 |
| Transfer | 09.99±0.38 | 08.84±0.91 | 10.06±0.41 | 10.31±0.49 | 26.40±2.17 | 35.08±3.21 | ||
| AUPR | Target-only | 14.71±1.59 | 22.06±7.23 | 25.48±2.83 | 20.98±1.14 | 24.40±3.18 | 18.47±5.74 | |
| Transfer | 04.96±0.09 | 04.53±1.23 | 04.96±0.11 | 05.46±0.17 | 25.24±2.01 | 27.97±4.93 | ||
| AUROC | Target-only | 58.16±5.19 | 71.94±3.89 | 77.82±1.64 | 75.60±0.13 | 78.25±4.88 | 75.24±5.31 | |
| Transfer | 58.94±1.10 | 55.71±1.39 | 58.55±1.00 | 59.27±0.91 | 75.08±2.63 | 84.11±1.09 | ||
| WADI→SWaT | F1 | Target-only | 54.70±2.76 | 42.64±8.18 | 68.00±1.49 | 39.77±2.03 | 20.25±3.00 | 35.17±3.88 |
| Transfer | 22.37±0.00 | 56.82±18.88 | 22.37±0.00 | 22.37±0.00 | 64.53±3.80 | 73.90±1.06 | ||
| AUPR | Target-only | 46.54±3.50 | 33.01±4.41 | 63.53±1.22 | 23.00±2.91 | 12.00±0.95 | 20.57±6.35 | |
| Transfer | 09.31±1.41 | 47.09±20.34 | 09.34±1.65 | 09.84±1.56 | 49.03±9.09 | 75.96±3.96 | ||
| AUROC | Target-only | 73.14±1.91 | 79.57±5.63 | 84.73±1.62 | 78.25±0.86 | 66.16±5.01 | 77.00±5.17 | |
| Transfer | 80.29±0.74 | 77.31±3.69 | 80.44±0.71 | 80.44±1.04 | 82.68±1.54 | 88.00±0.85 |
Theorem 1 (W2). We appreciate the reviewer's observation regarding the monotonicity of the optimal prototype distance. Beyond this property, Theorem 1 also establishes a strictly positive lower bound on the gap between the optimal single-prototype and multi-prototype distances under the multimodal separation assumptions (Eq. (6)).
Importantly, this bound explicitly relates the benefit of multiple prototypes to the separation and mixture weights of different normal modes. It therefore provides a quantitative theoretical motivation for modeling distinct normal patterns by multiple normal prototypes in MNP-GDA.
Theorem 2 (W2). We agree that Theorem 2 assumes nonpositive covariance between reliability weights and pseudo-label errors. Under this condition, the theorem establishes that reliability weighting does not increase the class-wise pseudo-label error rate or the corresponding surrogate error bound.
In MNP-GDA, the class-conditional weighting design reflects the distinct reliability characteristics of the two pseudo-label classes: lower anomaly scores indicate more reliable pseudo-normal samples, whereas higher scores indicate more reliable pseudo-anomalous samples (Eqs. (8)–(9)). Figure 5 validates that reliability weighting effectively improves pseudo-anomaly precision over the unweighted counterpart, providing empirical support for this design.
Ablation Results and Variance (Q4). The results in Table 2 are actually averaged over five random seeds. We agree that reporting only mean values does not fully reflect the variability across different seeds. In the revised manuscript, we will include the corresponding standard deviations (mean ± std) for all ablation results in Table 2.
Table 2 ablation on three tasks (mean ± std).
| Task | Variant | F1 | AUPR | AUROC |
|---|---|---|---|---|
| PSM→SMD | MNP-GDA (full) | 35.08±3.21 | 27.97±4.93 | 84.11±1.09 |
| w/o adaptive graph modeling | 31.17±3.10 | 25.64±7.06 | 82.82±2.36 | |
| w/o spatial dependency | 33.81±1.15 | 24.79±3.25 | 81.49±2.14 | |
| w/o temporal dependency | 26.28±1.52 | 15.10±1.16 | 78.14±1.20 | |
| w/o multiple normal prototypes | 28.86±3.08 | 21.71±4.33 | 80.48±1.62 | |
| w/o normal attraction loss | 30.38±8.37 | 21.08±8.86 | 79.46±5.83 | |
| w/o anomaly repulsion loss | 28.62±8.96 | 20.56±7.75 | 75.06±8.53 | |
| w/o adaptive reliability weighting | 33.71±1.90 | 24.96±4.58 | 82.34±2.09 | |
| JD1→SMD | MNP-GDA (full) | 35.30±1.35 | 27.26±4.92 | 82.35±0.93 |
| w/o adaptive graph modeling | 28.49±7.51 | 20.74±4.96 | 78.75±1.03 | |
| w/o spatial dependency | 30.50±0.91 | 21.90±1.09 | 78.79±1.37 | |
| w/o temporal dependency | 31.19±6.51 | 22.88±6.28 | 78.02±2.51 | |
| w/o multiple normal prototypes | 32.18±3.49 | 24.34±4.00 | 80.41±2.05 | |
| w/o normal attraction loss | 33.87±2.22 | 26.60±3.47 | 76.89±4.65 | |
| w/o anomaly repulsion loss | 33.06±3.05 | 26.48±3.44 | 76.74±3.34 | |
| w/o adaptive reliability weighting | 33.41±3.80 | 22.53±7.97 | 79.12±1.55 | |
| WADI→SWaT | MNP-GDA (full) | 73.90±1.06 | 75.96±3.96 | 88.00±0.85 |
| w/o adaptive graph modeling | 72.79±0.09 | 71.88±0.66 | 82.65±2.13 | |
| w/o spatial dependency | 72.36±0.12 | 72.36±0.90 | 83.96±1.53 | |
| w/o temporal dependency | 70.33±4.59 | 62.61±11.70 | 86.19±0.99 | |
| w/o multiple normal prototypes | 71.63±5.17 | 69.72±13.03 | 86.75±1.19 | |
| w/o normal attraction loss | 72.90±0.86 | 67.11±9.02 | 87.34±0.60 | |
| w/o anomaly repulsion loss | 71.90±0.99 | 67.45±10.82 | 86.38±2.33 | |
| w/o adaptive reliability weighting | 73.02±0.61 | 72.24±1.24 | 80.94±1.94 |
Reporting percentage improvements. We will revise the descriptions to report percentage improvements as percentage-point gains. As shown in Table 1, compared with VANDA, MNP-GDA achieves average improvements of 6.60, 9.48, and 5.93 percentage points in terms of F1, AUPR, and AUROC, respectively.
We deeply appreciate the reviewer for constructive comments and valuable suggestions. We have provided clarification on hyperparameter configuration, F1 threshold, and the assumptions and implications of Theorems 1 and 2. We have also conducted additional experiments on validation strategies, and compared with target-only unsupervised baselines to validate the benefit of cross-domain transfer. Our detailed responses are provided as follows.
1) Common experimental protocol of unsupervised domain adaptation. Conventional unsupervised domain adaptation literature typically follows the common experimental protocol [1-3], where the source domain is fully labeled while the target domain is completely unlabeled, thus labeled target-domain validation data are unavailable, making conventional supervised model selection infeasible. Our experiments actually follow such common protocol of unsupervised domain adaptation, i.e., did not use a validation set for hyperparameter tuning.
[1] Time Series Domain Adaptation via Sparse Associative Structure Alignment. Proc. AAAI, 2021.
[2] Boosting Transferability and Discriminability for Time Series Domain Adaptation. Proc. NeurIPS, 2024.
[3] DACAD: Domain Adaptation Contrastive Learning for Anomaly Detection in Multivariate Time Series. IEEE TKDE, 2025.
2) Hyperparameter setting. Actually, our task-specific hyperparameter configuration just uses a small set of commonly adopted default values based on empirical experience. As shown in Table 4, across the eight tasks covering different server-monitoring and industrial control datasets with rather distinct characteristics, most hyperparameters only take two default values, while the key training settings, including batch size, training epochs, encoder layers, and attention heads, always keep the same across all the eight tasks. We did not conduct exhaustive task-specific hyperparameter tuning to optimize test performance.
3) Additional experiments for hyperparameter selection using validation set. To further address this concern, we have conducted additional experiments to select hyperparameters upon a held-out validation set, using three strategies, i.e., S-val used the source validation data, T-val used the target validation data, and S+T-val used both. For all three strategies, source-domain data were split into training and validation at a 9:1 ratio, while target-domain data were split into training, validation, and test at an 8:1:1 ratio. That is, the training and test sets are kept identical across three strategies, while only validation data used for hyperparameter selection differ. We can see that across all three validation strategies, MNP-GDA consistently outperforms state-of-the-art cross-domain MTS-AD baselines, i.e., DACAD and VANDA, in terms of F1, AUPR, and AUROC. Such results together with the original results without validation set demonstrate that the proposed MNP-GDA always maintains a strong competitive performance no matter with or without validation-based hyperparameter selection, highlighting its robustness to different hyperparameter selection strategies.
Comparison of different validation strategies on three cross-domain MTS-AD tasks. Bold indicates the highest mean under the same task, metric, and validation strategy.
F1 threshold (W1+Q1). We follow existing literature of domain adaptation for MTS-AD (i.e. VANDA and DACAD) to report the best-F1, i.e., enumerating all possible thresholds to search for the best F1.