acceptodds
Under review as a conference paper at ICLR 2027

MNP-GDA: Multimodal Normality Prototypical Graph Domain Adaptation for Multivariate Time Series Anomaly Detection

Abstract

Traditional multivariate time series anomaly detection (MTS-AD) methods are primarily developed under the single-domain assumption that both training and test data originate from the same domain, while this rarely holds in real-world applications. The newly emerging cross-domain MTS-AD problem aims to transfer knowledge from a labeled source domain to assist MTS-AD in an unlabeled target domain. However, how to explicitly model spatio-temporal dependencies and transfer multimodal normalities across heterogeneous MTS domains remains largely under-explored. To address this, we propose a Multimodal Normality Prototypical Graph Domain Adaptation (MNP-GDA) framework that captures latent spatio-temporal dependencies by adaptive graph modeling. Furthermore, MNP-GDA introduces a prototypical domain adaptation module with theoretical justification, which models and transfers diverse normal modes across domains upon multiple normal prototypes, by compacting normal samples around their nearest prototypes while separating anomalies from all prototypes. Finally, a label-conditioned adaptive reliability weighting mechanism supported by theoretical analysis is devised to weight pseudo-normal and -anomalous target samples oppositely, mitigating negative influence of unreliable pseudo-labels during domain adaptation. Extensive experiments on real-world server-monitoring and industrial-control benchmark datasets demonstrate state-of-the-art performance of MNP-GDA on cross-domain MTS-AD tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph domain adaptation, Time series, Anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uwe3OpenReview ↗
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:

  1. How is the F1 threshold selected? Do you use the best F1 over all thresholds, a fixed percentile, or point adjustment?
  2. 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?
  3. What do the unsupervised baselines (for example MTGFlow) and MNP-GDA without the source domain achieve when trained only on target unlabeled data?
  4. 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

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.

TaskMethodMetricS-valT-valS+T-val
PSM→SMDDACADF118.35±2.1617.91±2.4819.14±1.95
AUPR14.82±1.8314.26±2.0515.37±1.64
AUROC57.94±2.6858.63±2.8760.18±2.45
VANDAF121.08±2.6422.45±2.1822.68±5.27
AUPR17.18±2.3518.12±1.9418.64±3.76
AUROC64.45±3.4263.83±1.3866.31±2.86
MNP-GDA (Ours)F124.29±0.9823.40±1.9223.39±0.88
AUPR21.39±1.4721.18±3.2720.40±0.92
AUROC71.95±3.3372.79±2.4472.56±1.90
JD1→SMDDACADF119.38±6.9618.09±5.0518.82±4.67
AUPR15.07±6.9514.05±5.8314.63±3.89
AUROC56.49±4.2563.40±4.9759.34±4.27
VANDAF119.83±1.9823.63±2.0620.00±2.66
AUPR15.45±0.6618.72±1.6015.72±2.54
AUROC61.05±2.5668.82±3.2560.00±9.82
MNP-GDA (Ours)F122.75±3.1225.32±3.5527.84±2.63
AUPR17.85±2.3021.35±3.0223.17±2.28
AUROC61.69±9.9072.05±2.4276.62±2.14
WADI→SWaTDACADF122.32±4.5119.73±3.6821.26±3.92
AUPR14.54±5.7913.34±3.1714.18±3.41
AUROC54.63±7.6057.28±6.4365.37±5.64
VANDAF116.95±5.1819.12±8.2324.88±3.78
AUPR12.21±7.0814.38±9.3215.36±5.51
AUROC59.12±3.1860.45±2.9467.12±5.16
MNP-GDA (Ours)F144.82±2.6346.15±2.3845.38±1.94
AUPR27.46±2.1829.18±2.4530.64±2.05
AUROC77.94±1.8579.26±1.6278.63±1.47

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.

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.

TaskMetricSettingOmniAnomalyTHOCMTAD-GATHSTVAEMTGFlowMNP-GDA (Ours)
SMD→PSMF1Target-only28.37±2.6426.74±3.8234.18±2.1925.62±1.3723.81±1.7635.29±3.41
Transfer22.96±0.1925.19±4.5623.00±0.1923.17±0.1821.01±0.1938.41±2.47
AUPRTarget-only21.49±2.1819.63±2.9125.82±2.0718.74±1.6317.29±3.8227.16±3.28
Transfer12.92±0.4016.78±8.8712.93±0.3113.74±0.2709.21±0.8129.37±3.09
AUROCTarget-only69.82±2.7367.45±3.4676.53±1.8461.85±1.1448.23±1.2458.98±2.92
Transfer56.07±0.9060.57±3.6656.21±0.7757.70±0.7759.22±4.9080.10±2.30
PSM→SMDF1Target-only20.73±3.5618.63±4.2523.48±2.6317.29±1.8419.15±2.1024.85±3.68
Transfer09.99±0.3808.84±0.9110.06±0.4110.31±0.4926.40±2.1735.08±3.21
AUPRTarget-only14.71±1.5922.06±7.2325.48±2.8320.98±1.1424.40±3.1818.47±5.74
Transfer04.96±0.0904.53±1.2304.96±0.1105.46±0.1725.24±2.0127.97±4.93
AUROCTarget-only58.16±5.1971.94±3.8977.82±1.6475.60±0.1378.25±4.8875.24±5.31
Transfer58.94±1.1055.71±1.3958.55±1.0059.27±0.9175.08±2.6384.11±1.09
WADI→SWaTF1Target-only54.70±2.7642.64±8.1868.00±1.4939.77±2.0320.25±3.0035.17±3.88
Transfer22.37±0.0056.82±18.8822.37±0.0022.37±0.0064.53±3.8073.90±1.06
AUPRTarget-only46.54±3.5033.01±4.4163.53±1.2223.00±2.9112.00±0.9520.57±6.35
Transfer09.31±1.4147.09±20.3409.34±1.6509.84±1.5649.03±9.0975.96±3.96
AUROCTarget-only73.14±1.9179.57±5.6384.73±1.6278.25±0.8666.16±5.0177.00±5.17
Transfer80.29±0.7477.31±3.6980.44±0.7180.44±1.0482.68±1.5488.00±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).

TaskVariantF1AUPRAUROC
PSM→SMDMNP-GDA (full)35.08±3.2127.97±4.9384.11±1.09
w/o adaptive graph modeling31.17±3.1025.64±7.0682.82±2.36
w/o spatial dependency33.81±1.1524.79±3.2581.49±2.14
w/o temporal dependency26.28±1.5215.10±1.1678.14±1.20
w/o multiple normal prototypes28.86±3.0821.71±4.3380.48±1.62
w/o normal attraction loss30.38±8.3721.08±8.8679.46±5.83
w/o anomaly repulsion loss28.62±8.9620.56±7.7575.06±8.53
w/o adaptive reliability weighting33.71±1.9024.96±4.5882.34±2.09
JD1→SMDMNP-GDA (full)35.30±1.3527.26±4.9282.35±0.93
w/o adaptive graph modeling28.49±7.5120.74±4.9678.75±1.03
w/o spatial dependency30.50±0.9121.90±1.0978.79±1.37
w/o temporal dependency31.19±6.5122.88±6.2878.02±2.51
w/o multiple normal prototypes32.18±3.4924.34±4.0080.41±2.05
w/o normal attraction loss33.87±2.2226.60±3.4776.89±4.65
w/o anomaly repulsion loss33.06±3.0526.48±3.4476.74±3.34
w/o adaptive reliability weighting33.41±3.8022.53±7.9779.12±1.55
WADI→SWaTMNP-GDA (full)73.90±1.0675.96±3.9688.00±0.85
w/o adaptive graph modeling72.79±0.0971.88±0.6682.65±2.13
w/o spatial dependency72.36±0.1272.36±0.9083.96±1.53
w/o temporal dependency70.33±4.5962.61±11.7086.19±0.99
w/o multiple normal prototypes71.63±5.1769.72±13.0386.75±1.19
w/o normal attraction loss72.90±0.8667.11±9.0287.34±0.60
w/o anomaly repulsion loss71.90±0.9967.45±10.8286.38±2.33
w/o adaptive reliability weighting73.02±0.6172.24±1.2480.94±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.

Reviewer_ze9YOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the problem of cross-domain multivariate time series anomaly detection where the source domain is labeled and the target domain is unlabeled, and the two domains have different variable sets and data distributions. The proposed method is to develop domain-specific adaptive graphs and spatio-temporal encoders, then transfer normal patterns through multiple shared normal prototypes. Further, it uses reconstruction errors, forecasting errors, and prototype distances to select target pseudo-labels and assign label-conditioned reliability weights.

Critical strengths
  1. Detailed mathematical formulation and clear exposition of the different loss components.
  2. The idea of transferring multiple normal prototypes across heterogeneous time series domains is well motivated, and anomaly detection is a good application to test that idea.
  3. The benefit of multiple prototypes in representing separated normal modes is theoretically justified by Theorem 1 under explicit assumptions.
  4. The experiments cover several cross-domain tasks, and the additional ablations on training objectives, anomaly scores, and pseudo-label selection provide useful evidence for the proposed components.
Critical weaknesses
  1. The single-prototype ablation will reduce the normal attraction loss in Eq. (3) to zero, so the performance drop may partly reflect the removal of this training signal rather than establish the benefit of multiple prototypes alone.

  2. The experimental section does not clearly specify the source training, target adaptation, validation, and test splits or explain how task-specific hyperparameters and F1 thresholds are selected without target labels.

  3. The baseline experiments do not adequately explain how source-only single-domain methods learn target input projections, and the absence of target-only training baselines makes it difficult to distinguish the benefits of cross-domain transfer from those of using target data.

  4. The experimental comparison could benefit from including additional representative and recent methods for modeling spatio-temporal dependencies, such as [1] SARAD, [2] TopoGDN, and [3] ScatterAD.

  5. Several presentation and implementation details need clarification, including reporting percentage-point gains as percentage improvements, inconsistent selection-ratio notation between Figure 3 and Table 4, and unspecified handling of prototype re-normalization and empty assignments.

[1] Dai Z, He L, Yang S H, et al. Sarad: Spatial association-aware anomaly detection and diagnosis for multivariate time series[J]. Advances in Neural Information Processing Systems, 2024

[2] Liu Z, Huang X, Zhang J, et al. Multivariate time-series anomaly detection based on enhancing graph attention networks with topological analysis[C]//Proceedings of the 33rd ACM international conference on information and knowledge management. 2024

[3] Yin T, Fu S, Zhang Z, et al. ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection[J]. Advances in Neural Information Processing Systems, 2026

Questions for authors

Q1. I have some fundamental concerns about how prototypes updated only using source-normal samples distinguish target-specific normal modes from anomalies when both may lie far from these prototypes.

7 comments

We deeply appreciate the reviewer for constructive comments and valuable suggestions. We have provided clarification on single-prototype ablation, experimental protocol, baseline settings, prototype updates, and handling of target-specific normal modes. Additionally, we have conducted new experiments on validation strategies, target-only unsupervised methods, and recent methods for modeling spatio-temporal dependencies. Our detailed responses are provided as follows.

Target-Specific Normal Modes (Q1). We agree that target-specific normal modes may lie far from the source-normal prototypes. However, in the proposed MNP-GDA, target anomaly detection does not rely solely on prototype distance, but also considers target-specific representations and complementary anomaly signals. 1) Target-specific representation learning. Although normal prototypes are updated using source-normal samples, the target domain has its own adaptive graph and own domain-specific spatio-temporal encoder, enabling MNP-GDA to effectively capture domain-specific temporal-spatial dependencies, rather than only relying on the source-domain representations. 2) Complementary anomaly signals. In Eq. (7), target pseudo-label selection jointly considers reconstruction error, forecasting error, and distance to the nearest normal prototype. A target-specific normal pattern may be distant from the source-normal prototypes while still exhibiting stable spatio-temporal structure and relatively low reconstruction and forecasting errors. These complementary signals provide additional information to distinguish target-specific normal modes that may lie far from the source-normal prototypes. Additionally, the final anomaly classifier in Eq. (12) further integrates the learned representations and these complementary anomaly signals. This enables MNP-GDA to retain source-normal prototypes as transferable references while learning target-specific representations and incorporating complementary evidence for anomaly detection. We will clarify this mechanism in Section 3.3 in the revised manuscript.

Eq. (3) and Single-Prototype Ablation (K=1) (W1). As the reviewer correctly observes, the contrastive normal-attraction loss in Eq. (3) becomes zero when K=1K=1. We would like to clarify how this case is handled in our ablation implementation. Actually, the multi-prototype normal-contrastive loss in Eq. (3) is used only when K>1K>1. While for the single-prototype variant (K=1K=1) reported in Table 2, we have devised another non-degenerate cosine-distance normal-attraction loss as 1−⟨z,p1⟩1-\langle z,p_1\rangle, instead of Eq. (3). Thus, the single-prototype variant reported in Table 2 indeed retains the training signal via such non-degenerate cosine-distance attraction loss rather than a zero loss. The outperformance of full MNP-GDA over the single-prototype variant (K=1K=1) indeed demonstrates the benefit of multiple prototypes. In the revised manuscript, we will add the non-degenerate cosine-distance normal-attraction loss for K=1 in Section 3.2, and clarify the implementation of single-prototype variant in ablation study.

Hyperparameter configuration (W2).

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.

TaskMethodMetricS-valT-valS+T-val
PSM→SMDDACADF118.35±2.1617.91±2.4819.14±1.95
AUPR14.82±1.8314.26±2.0515.37±1.64
AUROC57.94±2.6858.63±2.8760.18±2.45
VANDAF121.08±2.6422.45±2.1822.68±5.27
AUPR17.18±2.3518.12±1.9418.64±3.76
AUROC64.45±3.4263.83±1.3866.31±2.86
MNP-GDA (Ours)F124.29±0.9823.40±1.9223.39±0.88
AUPR21.39±1.4721.18±3.2720.40±0.92
AUROC71.95±3.3372.79±2.4472.56±1.90
JD1→SMDDACADF119.38±6.9618.09±5.0518.82±4.67
AUPR15.07±6.9514.05±5.8314.63±3.89
AUROC56.49±4.2563.40±4.9759.34±4.27
VANDAF119.83±1.9823.63±2.0620.00±2.66
AUPR15.45±0.6618.72±1.6015.72±2.54
AUROC61.05±2.5668.82±3.2560.00±9.82
MNP-GDA (Ours)F122.75±3.1225.32±3.5527.84±2.63
AUPR17.85±2.3021.35±3.0223.17±2.28
AUROC61.69±9.9072.05±2.4276.62±2.14
WADI→SWaTDACADF122.32±4.5119.73±3.6821.26±3.92
AUPR14.54±5.7913.34±3.1714.18±3.41
AUROC54.63±7.6057.28±6.4365.37±5.64
VANDAF116.95±5.1819.12±8.2324.88±3.78
AUPR12.21±7.0814.38±9.3215.36±5.51
AUROC59.12±3.1860.45±2.9467.12±5.16
MNP-GDA (Ours)F144.82±2.6346.15±2.3845.38±1.94
AUPR27.46±2.1829.18±2.4530.64±2.05
AUROC77.94±1.8579.26±1.6278.63±1.47

4) F1 threshold. 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.

Target input projections for source-only single-domain methods (W3). For the single-domain methods, i.e., OmniAnomaly, THOC, MTAD-GAT, HSTVAE, and MTGFlow, Table 1 reports the results under a source-direct-transfer setting to examine how conventional single-domain anomaly detectors perform when directly transferred to heterogeneous target domains. Since the source and target domains have different numbers of variables, we introduce separate linear input projections to map their inputs into a common feature dimension. The source-side projection is jointly trained with the detection model using source-domain data, while the target-side projection is independently initialized and used during target-domain inference.

Additional comparison with Target-only unsupervised methods (W3). We agree that comparison with target-only unsupervised methods can distinguish the benefits of cross-domain transfer from those of using target data alone. We have conducted new experiments to evaluate the performance of five target-only unsupervised baselines, i.e., OmniAnomaly, THOC, MTAD-GAT, HSTVAE, and MTGFlow, using the target-domain samples alone. 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 target-only comparisons in the revised manuscript.

Comparison between target-only learning and cross-domain transfer learning on three representative tasks.

TaskMetricSettingOmniAnomalyTHOCMTAD-GATHSTVAEMTGFlowMNP-GDA (Ours)
SMD→PSMF1Target-only28.37±2.6426.74±3.8234.18±2.1925.62±1.3723.81±1.7635.29±3.41
Transfer22.96±0.1925.19±4.5623.00±0.1923.17±0.1821.01±0.1938.41±2.47
AUPRTarget-only21.49±2.1819.63±2.9125.82±2.0718.74±1.6317.29±3.8227.16±3.28
Transfer12.92±0.4016.78±8.8712.93±0.3113.74±0.2709.21±0.8129.37±3.09
AUROCTarget-only69.82±2.7367.45±3.4676.53±1.8461.85±1.1448.23±1.2458.98±2.92
Transfer56.07±0.9060.57±3.6656.21±0.7757.70±0.7759.22±4.9080.10±2.30
PSM→SMDF1Target-only20.73±3.5618.63±4.2523.48±2.6317.29±1.8419.15±2.1024.85±3.68
Transfer09.99±0.3808.84±0.9110.06±0.4110.31±0.4926.40±2.1735.08±3.21
AUPRTarget-only14.71±1.5922.06±7.2325.48±2.8320.98±1.1424.40±3.1818.47±5.74
Transfer04.96±0.0904.53±1.2304.96±0.1105.46±0.1725.24±2.0127.97±4.93
AUROCTarget-only58.16±5.1971.94±3.8977.82±1.6475.60±0.1378.25±4.8875.24±5.31
Transfer58.94±1.1055.71±1.3958.55±1.0059.27±0.9175.08±2.6384.11±1.09
WADI→SWaTF1Target-only54.70±2.7642.64±8.1868.00±1.4939.77±2.0320.25±3.0035.17±3.88
Transfer22.37±0.0056.82±18.8822.37±0.0022.37±0.0064.53±3.8073.90±1.06
AUPRTarget-only46.54±3.5033.01±4.4163.53±1.2223.00±2.9112.00±0.9520.57±6.35
Transfer09.31±1.4147.09±20.3409.34±1.6509.84±1.5649.03±9.0975.96±3.96
AUROCTarget-only73.14±1.9179.57±5.6384.73±1.6278.25±0.8666.16±5.0177.00±5.17
Transfer80.29±0.7477.31±3.6980.44±0.7180.44±1.0482.68±1.5488.00±0.85
Reviewer_u1DHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies transfer from a fully labeled source multivariate time series to an unlabeled target domain with different variables. MNP-GDA combines domain-specific adaptive graphs and spatio-temporal encoders, a shared projection and bank of source-normal prototypes, normal attraction and anomaly repulsion, and label-conditioned weighting of selected target pseudo-labels. Reconstruction, forecasting and prototype distance also feed a binary anomaly classifier. Two conditional results motivate multiple prototypes and reliability weighting. Across eight transfers using five datasets and five seeds, the results are promising, but evaluation details and controls for the central mechanisms remain insufficient.

Critical strengths
  1. A useful formulation for heterogeneous domains. Separate graphs and encoders accommodate unequal variable sets, while shared normal prototypes provide interpretable anchors for transfer. Prototype-wise pseudo-normal selection addresses the risk that global selection overlooks less frequent normal patterns. This is a coherent design for an important deployment setting.

  2. Substantial and reasonably broad empirical evidence. Table 1 includes server and industrial datasets, eight transfer directions, threshold-free AUPR/AUROC, and variation over five seeds. MNP-GDA has the highest AUPR and AUROC on every reported task, although DACAD has slightly higher F1 on WADI→SWaT. Tables 2 and 5 support contributions from several components. Figure 5 also shows that weighting improves pseudo-anomaly precision on the two analyzed tasks.

Critical weaknesses
  1. The evaluation protocol and comparative evidence are not specified well enough to establish the claimed advantage. Appendix D.1 describes windowing and aggregate counts, but does not identify the exact source/target training, validation and test segments, machine/service subsets, or how temporal overlap is excluded across held-out splits. The use of labeled source anomalies needs to be reconciled with the underlying dataset partitions. Table 4 uses many task-specific hyperparameters, including markedly different loss weights and pseudo-label ratios, without an explanation of how they were selected when target labels are unavailable. The rule for selecting the F1 threshold is also absent. These are uncertainties, not evidence that label leakage occurred, but they directly affect the meaning of the reported gains. The same concern applies to baseline access and adaptation. In Appendix D.3 the single-domain baselines are trained only on source windows after separate input projections; how the target projections are obtained is unclear. There is no target-only unsupervised comparison using the same target observations, so the benefit of source transfer over learning from the target itself is not established. VANDA uses mean-pooled timestamp scores against any-anomaly window labels (Appendices D.1/D.3), which needs a justified, consistently validated aggregation rule. PUAD (Li et al., ICML 2023) is discussed but not compared. It already includes transferable prototypes and adaptation to new series; Section 2 should qualify its description of earlier prototype methods as unable to transfer knowledge. These gaps limit fairness and attribution of novelty.

  2. The evidence does not cleanly isolate the proposed mechanisms. In Eq. (3), K=1 gives a numerator identical to the denominator, so the normal-attraction loss is exactly zero and supplies no normal-attraction gradient. Consequently, the single-prototype ablation in Table 2 changes both the number of modes and whether this training signal exists. Eq. (3) encourages relative prototype preference, whereas Theorem 1 concerns optimal absolute nearest-prototype distance for a fixed distribution. The theorem therefore does not establish that the learned encoders and momentum prototypes attain its improvement, or that the improvement transfers to a different target distribution. A single-center control with a nonzero attraction loss is needed to support the main interpretation. The reliability analysis also has limited scope. Theorem 2 is conditional on nonpositive covariance between weights and pseudo-label error and concerns the prototypical surrogate, rather than a generalization guarantee. Figure 5 is helpful empirical support, but its selected pseudo-anomalies still have weighted precision below 50% on both shown tasks. Meanwhile, Eq. (23) uses those pseudo-labels in an unweighted target BCE term. Thus noisy labels retain a direct training path that the proposed reliability theorem does not cover. Without controls for target BCE and evidence across tasks or training stages, the robustness claim remains only partially supported.

Questions for authors

Q1. Protocol clarification. Please provide the exact files/entities and chronological segments used for labeled source training, unlabeled target adaptation, validation and testing, including source anomaly availability, standardization statistics and overlap handling. Is adaptation transductive on the evaluated target windows, or performed on a disjoint target segment? How were Table 4 hyperparameters, checkpoints and F1 thresholds chosen without target test labels? These answers could resolve the principal evaluation concern. Q2. Comparisons needed to assess transfer. Under the same data access and selection budget, how do target-only unsupervised models and a prototype-based comparator such as PUAD perform? How were the target-side input projections for source-only baselines learned, and is VANDA’s score aggregation robust to an alternative aggregation selected using the same validation policy? These checks would determine whether the gains reflect source transfer and a fair comparison. Q3. Mechanism checks needed for the main claims. Does MNP-GDA outperform a K=1 model with a non-degenerate cosine-distance attraction loss, with other losses and the encoder fixed? For reliability weighting, please compare the present unweighted target BCE with its removal or weighting, and report pseudo-label precision before and after weighting across tasks and epochs. Please also specify prototype re-normalization after momentum updates and handling of empty prototype assignments. These results would distinguish multimodal transfer and noise mitigation from changes in optimization. They are requests relevant to the current claims, rather than suggestions for future work.

9 comments

We deeply appreciate the reviewer for constructive comments and valuable suggestions. We have provided clarification on data segments, transductive adaptation protocol, hyperparameter configuration, F1 threshold, baseline settings, theoretical and implementation details. To address experimental concerns, we have conducted new experiments to evaluate different validation strategies, target-only unsupervised baselines, VANDA's alternative score aggregations, a new baseline PUAD, a target BCE ablation and pseudo-label precision across tasks and epochs. Our detailed responses are provided as follows.

Exact data segments for training and test (W1+Q1). For SMD and PSM datasets \citep{liu2024tsbad}, we adopted the multivariate versions provided by TSB-AD-M (https://thedatumorg.github.io/TSB-AD/), where SMD includes eight entity sequences (IDs 1–8, corresponding to files 057–064), and PSM uses the complete sequence. For JD1, we adopted Dataset2 from the publicly available JumpStarter repository (https://github.com/NetManAIOps/JumpStarter), covering 30 services. SWaT and WADI (https://www.sutd.edu.sg/itrust/request-for-datasets/) are obtained from their respective industrial control system attack datasets, with downsampling and feature preprocessing as described in Appendix D.1. All these five datasets provide ground-truth labels indicating each sample as normal or anomalous. In cross-domain MTS-AD tasks, we follow the common unsupervised domain adaptation setting [1–3], i.e., during training, all samples in the source domain have accessible labels, while all samples in the target domain are without observed labels. For example, in experiments, to simulate the cross-domain MTS-AD task from SMD to PSM, we would assign each sample with its ground-truth label in SMD, while hiding the labels of all samples in PSM.

Transductive adaptation (W1+Q1) In the common setting of unsupervised domain adaptation problem [1–3], the adaptation is transductive on the unlabeled target samples, i.e., the unlabeled target samples can be seen during training, but their ground-truth labels would be accessible only during test. We follow such setting, that is, the adaptation is transductive on the evaluated unlabeled target windows.

[1] Domain adaptation for time series under feature and label shifts. Proc. ICML. 2023.

[2] CauDiTS: Causal disentangled domain adaptation of multivariate time series. Proc. ICML. 2024.

[3] DACAD: Domain adaptation contrastive learning for anomaly detection in multivariate time series. IEEE TKDE, 2025.

Temporal overlap (W1+Q1) In the cross-domain MTS-AD tasks, the labeled source windows and unlabeled target windows originate from different domains, thus, the labeled source windows and unlabeled target windows do not have temporal overlap.

Standardization (W1+Q1) The source and target domains are standardized independently using per-variable Z-score normalization. The mean and standard deviation are computed from the respective domain's input features. For the target domain, these statistics are computed from the unlabeled target sequence and consistently applied during training and test.

Checkpoint (W1+Q1). For all eight cross-domain MTS-AD tasks, MNP-GDA is trained for a fixed 20 epochs, and the final-epoch checkpoint is used for evaluation.

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.

Hyperparameter configuration (W1+Q1).

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.

TaskMethodMetricS-valT-valS+T-val
PSM→SMDDACADF118.35 ± 2.1617.91 ± 2.4819.14 ± 1.95
AUPR14.82 ± 1.8314.26 ± 2.0515.37 ± 1.64
AUROC57.94 ± 2.6858.63 ± 2.8760.18 ± 2.45
VANDAF121.08 ± 2.6422.45 ± 2.1822.68 ± 5.27
AUPR17.18 ± 2.3518.12 ± 1.9418.64 ± 3.76
AUROC64.45 ± 3.4263.83 ± 1.3866.31 ± 2.86
MNP-GDA (Ours)F124.29 ± 0.9823.40 ± 1.9223.39 ± 0.88
AUPR21.39 ± 1.4721.18 ± 3.2720.40 ± 0.92
AUROC71.95 ± 3.3372.79 ± 2.4472.56 ± 1.90
JD1→SMDDACADF119.38 ± 6.9618.09 ± 5.0518.82 ± 4.67
AUPR15.07 ± 6.9514.05 ± 5.8314.63 ± 3.89
AUROC56.49 ± 4.2563.40 ± 4.9759.34 ± 4.27
VANDAF119.83 ± 1.9823.63 ± 2.0620.00 ± 2.66
AUPR15.45 ± 0.6618.72 ± 1.6015.72 ± 2.54
AUROC61.05 ± 2.5668.82 ± 3.2560.00 ± 9.82
MNP-GDA (Ours)F122.75 ± 3.1225.32 ± 3.5527.84 ± 2.63
AUPR17.85 ± 2.3021.35 ± 3.0223.17 ± 2.28
AUROC61.69 ± 9.9072.05 ± 2.4276.62 ± 2.14
WADI→SWaTDACADF122.32 ± 4.5119.73 ± 3.6821.26 ± 3.92
AUPR14.54 ± 5.7913.34 ± 3.1714.18 ± 3.41
AUROC54.63 ± 7.6057.28 ± 6.4365.37 ± 5.64
VANDAF116.95 ± 5.1819.12 ± 8.2324.88 ± 3.78
AUPR12.21 ± 7.0814.38 ± 9.3215.36 ± 5.51
AUROC59.12 ± 3.1860.45 ± 2.9467.12 ± 5.16
MNP-GDA (Ours)F144.82 ± 2.6346.15 ± 2.3845.38 ± 1.94
AUPR27.46 ± 2.1829.18 ± 2.4530.64 ± 2.05
AUROC77.94 ± 1.8579.26 ± 1.6278.63 ± 1.47

Additional comparison with Target-only unsupervised methods (W1+Q2). We agree that comparison with target-only unsupervised methods can determine whether our gains reflect source transfer. We have conducted new experiments to evaluate the performance of five target-only unsupervised baselines, i.e., OmniAnomaly, THOC, MTAD-GAT, HSTVAE, and MTGFlow, using the target-domain samples alone. 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 their 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.

TaskMetricSettingOmniAnomalyTHOCMTAD-GATHSTVAEMTGFlowMNP-GDA (Ours)
SMD→PSMF1Target-only28.37±2.6426.74±3.8234.18±2.1925.62±1.3723.81±1.7635.29±3.41
Transfer22.96±0.1925.19±4.5623.00±0.1923.17±0.1821.01±0.1938.41±2.47
AUPRTarget-only21.49±2.1819.63±2.9125.82±2.0718.74±1.6317.29±3.8227.16±3.28
Transfer12.92±0.4016.78±8.8712.93±0.3113.74±0.2709.21±0.8129.37±3.09
AUROCTarget-only69.82±2.7367.45±3.4676.53±1.8461.85±1.1448.23±1.2458.98±2.92
Transfer56.07±0.9060.57±3.6656.21±0.7757.70±0.7759.22±4.9080.10±2.30
PSM→SMDF1Target-only20.73±3.5618.63±4.2523.48±2.6317.29±1.8419.15±2.1024.85±3.68
Transfer09.99±0.3808.84±0.9110.06±0.4110.31±0.4926.40±2.1735.08±3.21
AUPRTarget-only14.71±1.5922.06±7.2325.48±2.8320.98±1.1424.40±3.1818.47±5.74
Transfer04.96±0.0904.53±1.2304.96±0.1105.46±0.1725.24±2.0127.97±4.93
AUROCTarget-only58.16±5.1971.94±3.8977.82±1.6475.60±0.1378.25±4.8875.24±5.31
Transfer58.94±1.1055.71±1.3958.55±1.0059.27±0.9175.08±2.6384.11±1.09
WADI→SWaTF1Target-only54.70±2.7642.64±8.1868.00±1.4939.77±2.0320.25±3.0035.17±3.88
Transfer22.37±0.0056.82±18.8822.37±0.0022.37±0.0064.53±3.8073.90±1.06
AUPRTarget-only46.54±3.5033.01±4.4163.53±1.2223.00±2.9112.00±0.9520.57±6.35
Transfer09.31±1.4147.09±20.3409.34±1.6509.84±1.5649.03±9.0975.96±3.96
AUROCTarget-only73.14±1.9179.57±5.6384.73±1.6278.25±0.8666.16±5.0177.00±5.17
Transfer80.29±0.7477.31±3.6980.44±0.7180.44±1.0482.68±1.5488.00±0.85

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair and Reviewers,

Thank you for your time and effort in reviewing our paper and your thoughtful and constructive feedback.

Contributions and Strengths

We sincerely appreciate the reviewers' recognition of the following strengths of our work:

  1. A meaningful formulation of heterogeneous cross-domain MTS-AD problem. Our framework addresses domain shifts and unshared variable sets through domain-specific adaptive graphs and spatio-temporal encoders, and shared normal prototypes provide interpretable anchors for knowledge transfer (Reviewers u1DH and ze9Y).

  2. Well-motivated multimodal normality modeling with theoretical support. Multiple transferable normal prototypes provides a principled way to transfer diverse normal modes across domains, with theoretical analysis characterizing the benefits under explicit assumptions. Prototype-wise pseudo-normal selection helps account for less frequent normal patterns that global selection may overlook (Reviewer u1DH and ze9Y).

  3. Strong and consistent empirical performance. For eight tasks across five server and industrial benchmark datasets, MNP-GDA consistently achieves the best overall performance, this is a clear empirical improvement in a setting that has few methods so far. (Reviewers u1DH, ze9Y and uwe3).

  4. Ablation and reliability weight analysis supporting key components. Extensive ablation studies in Table 3 and Table 9 support the proposed key components, including adaptive graph modeling, spatial and temporal dependencies, multiple normal prototypes, normal attraction and anomaly repulsion, reliability weighting, and complementary training objectives. Moreover, higher reliability-weighted selection precision for pseudo-anomalies in Figure 5 provides a useful empirical check of the condition behind Theorem 2 (Reviewers u1DH, ze9Y, and uwe3).

Summary of Revision

We have carefully addressed reviewers' concerns and uploaded a revised manuscript incorporating methodological clarifications and additional experiments. In the updated manuscript, we highlighted our revision in blue. The main revision are summarized as follows:

1. Evaluation Protocol Clarification.

  • We have clarified hyperparameter configurations (Section 4.1 and Appendix D.2), data segments and normalization (Appendix D.1), transductive adaptation (Section 4.1), checkpoint and best-F1 (Section 4.1), and target-side projections for baselines (Appendix D.3). We have added three validation strategies (S-val, T-val, and S+T-val) to validate the robustness of MNP-GDA to different hyperparameter selection strategies (Appendix D.2 and Table 6).

2. Methodological and Theoretical Clarification.

  • We have clarified the non-degenerate cosine-distance attraction loss used in single-prototype ablation (K=1) (Section 4.3), prototype initialization, normalized momentum updates, and empty-assignment handling (Section 3.2, Appendix A.3, and Algorithm 1). We have explained how to distinguish target-specific normal modes (Sections 3.1 and 3.3), and refined theoretical discussion (Remarks of Theorem 1 and Theorem 2).

3. Additional Experimental Results.

  • We have additionally compared with five single-domain unsupervised baselines (OmniAnomaly, THOC, MTAD-GAT, HSTVAE, MTGFlow) and MNP-GDA under target-only setting, providing direct evidence of the benefit of source transfer (Section 4.2, Table 2).

  • We have additionally compared with more baselines, including prototype-based PUAD and recent spatio-temporal MTS-AD methods, ScatterAD, SARAD, and TopoGDN (Appendix F, Table 7). Our MNP-GDA consistently outperforms these single-domain baselines, highlighting the benefit of transferring labeled source-domain knowledge through effective domain adaptation.

  • We have conducted new experiments for target-BCE ablations, comparing its removal and reliability-weighted variants with the original unweighted BCE (Appendix G, Table 10), and additional pseudo-label precision analyses across tasks and epochs (Appendix G, Table 11) to provide further evidence for the proposed reliability weighting mechanism. We have provided standard deviations of ablation results (Table 3 and Table 9).

  • We have additionally compared alternative score aggregations for VANDA, including mean, max, and sum (Table 8).

We hope the clarifications and additional experimental results help Area Chair and reviewers better assess the methodological contributions and empirical evidence of MNP-GDA.

Thank you again for your valuable time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.