acceptodds
Under review as a conference paper at ICLR 2027

Learning Dependencies and Localizing Anomalies with Neural Rosenblatt Transform

Abstract

Multivariate time series anomaly localization remains relatively understudied, despite its importance for diagnosing anomalies in many real-world complex systems. Existing models may result in ambiguous per-series anomaly scores, mainly due to inaccurate modeling of cross-series dependencies. Furthermore, methods focused on dependency learning typically do not address anomaly localization and are mostly designed for continuous-valued time series, limiting their applicability to mixed continuous–discrete settings common in real-world systems. To address these limitations, **NeuralRos** is proposed to jointly learn cross-series dependencies and localize anomalies in mixed-type multivariate time series. NeuralRos reformulates the Rosenblatt transformation from a diagnostic tool into a learning mechanism, combining a learnable variable ordering with neural conditional distribution models to capture cross-series dependencies. Two complementary localization scores are derived: a likelihood-based score from the learned normal conditional distributions and a Rosenblatt-based score from deviations of the transformed variables from their theoretical behavior. Experiments on synthetic and real-world datasets demonstrate the effectiveness of NeuralRos across both tasks.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time series, cross-correlation, anomaly localization, mixed data.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LhwhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NeuralRos, a framework that repurposes the Rosenblatt transformation from a diagnostic tool into a learning mechanism for multivariate time series. The key idea is to jointly learn cross-series dependencies and localize anomalies by modeling the conditional distributions induced by a learned Rosenblatt ordering. The method is evaluated on eight synthetic datasets for dependency identification and three real-world datasets (SMD, MSDS, SWaT) for anomaly localization.

Critical strengths
  1. Repurposing the Rosenblatt transformation as a learning objective is novel.
  2. Most existing methods assume purely continuous variables. The randomized quantile extension is a practical way to handle discrete series.
  3. Extensive experiments on 11 datasets indicate the effectiveness.
Critical weaknesses
  1. The ordering score is derived from attention asymmetry (Aj,iA_{j, i} - Ai,jA_{i, j}), but it is unclear why attention asymmetry could indicate directional ordering in a Rosenblatt factorization.
  2. The LRO uses a straight-through estimator with hard sorting in the forward pass. It is unclear whether the learned ordering converges consistently across runs.
  3. This paper states that NeuralRos "also has generative capabilities, which are important for system diagnosis," but no generation experiments or qualitative samples are provided.
  4. The largest real-world dataset has only 51 variables (SWaT), which doubts whether NeuralRos can scale to hundreds or thousands of channels.
Questions for authors
  1. Does the learned Rosenblatt ordering converge to a consistent permutation across different random seeds? Can you provide quantitative analysis?
  2. Why does attention asymmetry provide directional ordering evidence? Is there a formal link to conditional independence or causal ordering? How do you handle numerical issues from polynomial transformations (e.g., negative values, large magnitudes)?
  3. How does NeuralRos scale to high-dimensional MTS?
  4. Due to the claim that NeuralRos has generative capabilities, can you demonstrate this by generating realistic MTS samples and showing their utility for diagnosis or anomaly simulation?
  5. In the ablation "without Rosenblatt-based learning process", what exactly is removed? Is it the uniformity and independence losses, the entire Rosenblatt transform, or both? Please clarify the ablated variant.
4 comments

Comment 1:

We thank the Reviewer for this insightful comment. The motivation behind using attention asymmetry is to identify a suitable ordering for the Rosenblatt transformation based on cross-series dependencies. As explained in Section 4.2, the attention mechanism compares the original time series with linear and nonlinear transformations of the other series, allowing us to assess their relative dependencies. Specifically, if modeling the conditional distribution of series ii relies more strongly on the representations of series jj than the reverse, we favor placing jj before ii in the Rosenblatt ordering. This allows the current observation of series jj to be used when estimating the conditional distribution of series ii, thereby facilitating dependency learning. The attention asymmetry therefore provides a data-driven preference for the relative ordering of each pair of time series, which is subsequently aggregated to obtain the final ordering. We have further clarified the above points in the revised manuscript (see Section 4.2, pp. 4–5, lines 180-195).

Furthermore, in response to a related comment from another Reviewer, we have extended our ablation study by comparing the proposed LRO module with both a fixed natural ordering and a marginal variance-based ordering. The additional results, reported in Table 7 of Appendix B (page 26), show that LRO consistently achieves substantially lower HHD across all four datasets, providing further empirical support for the effectiveness of the proposed ordering mechanism.

 

Comment 2

We thank the Reviewer for this insightful comment. The Rosenblatt transformation is theoretically valid for any variable ordering. The proposed LRO module allows the model to learn a suitable ordering in a data-driven manner, adapting to the underlying dependencies of each dataset. However, this does not necessarily imply that the learned ordering must be identical across independent training runs. We acknowledge this potential variability, which was an important consideration in our experimental design. Specifically, we evaluate NeuralRos over 30 independent runs with different random seeds, which may result in different learned orderings, and report the mean performance together with the corresponding standard deviations. As shown in Table 5 of Appendix B, the standard deviations across the cross-series dependency identification metrics are consistently very small, indicating that NeuralRos achieves stable dependency identification performance despite potential variations in the learned ordering. These results provide empirical evidence of the robustness of the proposed approach across independent training runs.

To further address the Reviewer's comment, we conducted an additional quantitative analysis of the learned permutations across the 30 independent runs. The results show only minor variations, with approximately 2–3 changes in variable positions for SMD, 1–2 for MSDS, and 3-4 for SWaT, on average. These findings provide further evidence of the consistency of the learned Rosenblatt ordering across different random seeds. We have included this additional analysis in Appendix B of the revised manuscript, under “Supplementary Ablation Studies” (page 27, lines:1250-1260).

 

Comment 3

Many thanks for your comment.

NeuralRos has generative capabilities, since the neural conditioners learn the conditional distributions of the individual time series, and the Rosenblatt factorization provides a probabilistic representation of their joint distribution, from which new observations can be generated. We would like point out that the intention was not to overclaim this capability, as it follows from the probabilistic formulation of the proposed model. Nevertheless, we agree with the Reviewer that these capabilities were not explicitly evaluated in our experimental study. Therefore, we have removed this statement from the list of contributions in the revised manuscript. We nevertheless consider generative capabilities an important advantage of the proposed probabilistic framework compared with commonly used reconstruction-based approaches for anomaly detection, which do not explicitly model the joint probability distribution of the observations.

 

Comment 4

We have included the available real-world datasets suitable for evaluating the anomaly localization task, as datasets with reliable feature-level ground-truth labels remain limited. Nevertheless, the computational cost analysis presented in Table 3 of the revised manuscript (page 10) provides insights into how model complexity and inference time scale with increasing numbers of time series. The results show that NeuralRos remains highly computationally efficient as the number of variables increases. Specifically, on SWaT, the largest real-world dataset considered (d=51d=51), NeuralRos requires only approximately 2.8 million learnable parameters, compared with approximately 28.9 million for AERCA, one of the strongest baselines for cross-series dependency identification. Similarly, NeuralRos achieves an inference time of only 0.11 ms per sample, compared with 45 ms for AERCA. These results demonstrate the computational efficiency of the proposed approach, even as the number of time series increases.

Q1

To quantitatively assess the consistency of the learned Rosenblatt ordering, we examined variations in the learned permutations across 30 independent runs with different random seeds. We observed that the learned orderings exhibit only minor variations, with approximately 2–3 changes in variable positions for SMD, 1–2 for MSDS, and 3-4 for SWaT, on average. These results indicate that the LRO module learns relatively consistent orderings across independent runs, with small but expected variations. We emphasize that the Rosenblatt transformation is theoretically valid for any variable ordering, and the purpose of the LRO module is to identify a suitable ordering that facilitates dependency learning in each independent run, rather than enforcing an identical permutation across all runs. Furthermore, the robustness of the proposed approach is supported by the consistently strong mean performance and low standard deviations reported across the evaluation metrics (see Tables 5 and 6 of Appendix B, page 25-26). These findings suggest that small variations in the learned ordering do not substantially affect the overall performance of NeuralRos. To address the Reviewer's comment, we have included this additional analysis in Appendix B of the revised manuscript, under “Supplementary Ablation Studies” (page 27, lines:1250-1260).

 

Q2

Regarding the motivation for using attention asymmetry to determine the Rosenblatt ordering, please see the reply to Comment 1.

Regarding the numerical stability of the polynomial transformations, we clarify that all datasets undergo preprocessing, including normalization, before being provided as input to NeuralRos. Consequently, the Query and Key representations are constructed from normalized time-series data, reducing the risk of excessively large values during polynomial transformations. Furthermore, as explicitly described in Section 4.2, the resulting Query and Key representations are also row-normalized before computing the attention matrices, ensuring that the attention scores remain bounded between −1 and 1. Regarding negative input values, the polynomial transformations use integer powers, which are mathematically well-defined for both positive and negative values. Odd powers preserve the sign of negative inputs, while even powers produce non-negative values. Therefore, negative values do not introduce numerical difficulties in the polynomial transformations. These normalization steps help maintain numerical stability throughout the proposed LRO mechanism. We appreciate the Reviewer's comment and welcome any further suggestions for improving numerical stability. Nevertheless, we did not encounter any numerical stability issues during our experiments across the evaluated datasets.

 

Q3

Please see our reply to your Comment 4

 

Q4

Please see our reply to your Comment 3

 

Q5

We thank the Reviewer for this comment. We clarify that the ablated variant removes the entire Rosenblatt-based learning mechanism, including the Rosenblatt transformation and its associated uniformity and independence losses. Specifically, the ablation compares two models with the same neural conditioner architecture: one trained using the proposed Rosenblatt-based learning procedure and the other trained without it. This allows us to isolate the contribution of our core methodological novelty, namely, using the Rosenblatt transformation as a learning mechanism rather than solely relying on fitting the conditional distributions through likelihood optimization. The results (Main Manuscript: Fig. 1; Appendix: Table 6 and Fig. 6, pp. 26–27) demonstrate that the performance improvements arise from the proposed Rosenblatt-based learning mechanism rather than simply from the expressive capacity of the neural conditioners. To avoid any ambiguity, we have explicitly clarified this distinction in the ablation study discussion on page 9 of the revised manuscript.

 

Finally, we would like to note that we have included three additional baseline methods for comparison in the revised manuscript to further strengthen our experimental evaluation.

Thank you for your rebuttal. Most of my concerns have been resolved and I will increase my rating to 3.

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

The paper proposes a method that localizes anomalies in multivariate time series data and also learns cross-channel dependencies. The paper uses the Rosenblatt transformation to learn these dependencies. The paper sets up a Learnable Rosenblatt Ordering (LRO) module that relies upon polynomial-key multi-head attention followed by Sinkhorn-normalization to produce differentiable permutations over variables. Furthermore, conditionals are set up using Kolmogorov-Arnold Networks (KAN). The approach can deal with continuous variables, counts, and integer values. The model is trained in three stages: 1) minimize the negative log-likelihood, 2) force the U values to be uniform on [0,1] and enforce mutual independence of U components, and 3) jointly optimize (1) and (2). Once the model is trained, a high value of the negative log-likelihood of a new reading given its conditioners suggests that this reading is anomalous. The model also outputs Rosenblatt tail distance score. Scalars inside the KAN conditioners provide the dependency matrix that tells us how much one variable depends upon another variable. The proposed model is evaluated on eight synthetic datasets and three real-world datasets against ten dependency baselines and ten localization baselines.

Critical strengths
  • It is interesting to see how Rosenblatt transformation can be used as a training objective.
  • The three-stage loss (NLL, uniformity, independence) is grounded in the theory of Rosenblatt transformations.
  • The proposed approach is able to deal with mixed channel types, i.e., some channels are continuous, others contain counts, and yet others contain integer values.
  • The LRO module is new and offers end-to-end differentiability of ordering.
  • The proposed method ties or leads on almost all benchmarks.
Critical weaknesses
  • The three-stage training seems cumbersome. The paper should include more details about how the training takes place in practice (transition between stages, early stopping, learning rate schedule). I suspect that in its present form the results reported in this paper are not reproducible.
  • Is there a theoretical connection between attention asymmetry and partial-order discovery? The method as proposed seems to work in practice as shown in Figure 1, but to me the theoretical connection isn't clear.
  • The dependency matrix seems to come from the conditioners and not from Rosenblatt scores. The LRO module, for example, constrains the distribution properties, but it doesn't directly control the dependency matrix. Is this understanding correct?
  • It is not clear how hyperparameters were tuned for the baselines. Recall that for some baselines, their performance is closely tied to the choice of hyperparameters.
  • Please discuss in detail the computational cost of training this model. It is briefly mentioned in Sec. 5, but that claim should be expanded and supported by ablation (in the main paper).
Questions for authors
  • The dependency matrix depends upon the conditioners. Is it invariant w.r.t. ordering (relabeling of channels)? I.e., if we train two instances with different learned orderings, would we get the same dependencies?
  • Is it possible to compare LRO with a simpler ordering baseline, like ordering by marginal variance? This would showcase the need for using LRO and justify its added complexity.
  • Scores on SWaT are low across the board. Is there a reason for that?
2 comments
AuthorsOpenReview ↗

Responses to Comments 1–5.

Comment 1:

We thank the Reviewer for this valuable comment, which has helped us improve the clarity of the training procedure and the reproducibility of our methodology. To address this concern, we have added a paragraph in Appendix A.2 providing further implementation details on the three-stage training procedure, including the transition between stages, the allocation of training epochs, the learning-rate scheduling strategy, and the early stopping criterion based on validation performance. We have also added an explicit reference to these details at the end of Section 4.4 of the main manuscript.

 

Comment 2

We thank the Reviewer for this insightful comment. The motivation behind using attention asymmetry is to identify a suitable ordering for the Rosenblatt transformation based on cross-series dependencies. As explained in Section 4.2, the attention mechanism compares the original time series with linear and nonlinear transformations of the other series, allowing us to assess their relative dependencies. Specifically, if modeling the conditional distribution of series ii relies more strongly on the representations of series jj than the reverse, we favor placing jj before ii in the Rosenblatt ordering. This allows the current observation of series jj to be used when estimating the conditional distribution of series ii, thereby facilitating dependency learning. The attention asymmetry therefore provides a data-driven preference for the relative ordering of each pair of time series, which is subsequently aggregated to obtain the final ordering. We clarify that this mechanism is motivated by the conditional factorization of the Rosenblatt transformation rather than by a theoretical guarantee of partial-order recovery. We have further clarified the above points in the revised manuscript (see Section 4.2, pp. 4–5, lines 180-195).

 

Comment 3

We thank the Reviewer for this comment, which has helped us highlight better our methodology. The reviewer's understanding is correct. The LRO module determines the ordering of the time series in the Rosenblatt transformation, while the cross-series dependencies are explicitly represented by the learnable weights γi,j\gamma_{i,j} of the neural conditioners, defined in Eq. (9). We have clarified this distinction in Section 4.5 (page 7, lines: 325-328 ) of the revised manuscript, where the cross-series dependency matrix is introduced. We also note that the attention weights learned by the LRO module can provide additional insights into cross-series relationships; however, the final dependency matrix is defined and learned through the weights γi,j\gamma_{i,j} of the neural conditioners.

 

Comment 4

We reproduced the results of all baseline methods using their official or publicly available GitHub implementations, following the hyperparameter configurations recommended in their respective papers to ensure a fair and consistent comparison. Furthermore, we evaluated each baseline over 30 independent runs, reporting the mean performance and standard deviation. This allows us to assess not only the performance of each method but also its variability across different runs, which is particularly important for drawing reliable conclusions about the comparative effectiveness and robustness of the evaluated methods. To address the reviewer's comment, we have added a new subsection, "Baseline", in Appendix A.4 (page 21) , where we clarify the hyperparameter selection protocol and provide the links to the publicly available source code for each baseline evaluated in this study.

 

Comment 5

We thank the Reviewer for this valuable comment. We would like to clarify that the original submission already included a detailed computational cost analysis, along with the corresponding experimental results. This study was included in the Appendix due to space limitations. In the revised manuscript, taking advantage of the additional page allowance, we have moved this analysis and the corresponding results to the main manuscript (see page 10). The results demonstrate that the improved performance of the proposed method in both cross-series dependency identification and anomaly localization is accompanied by strong computational efficiency compared to other competitive baselines, further highlighting the effectiveness and applicabilityof our approach.

AuthorsOpenReview ↗

Responses to Questions

Q1

We thank the Reviewer for this insightful comment. We acknowledge that the learned dependencies may exhibit some variability across different learned orderings. This is one of the motivations for introducing the Learnable Rosenblatt Ordering (LRO) module, which allows the model to adaptively learn a suitable ordering for each dataset rather than relying on an arbitrary fixed ordering. Although the Rosenblatt transformation is theoretically valid for any variable ordering, learning the ordering enhances the model's ability to capture the underlying cross-series dependencies. To account for this variability, we evaluate NeuralRos over 30 independent runs with different random seeds, which may result in different learned orderings. We report the mean performance across these runs, together with the corresponding standard deviations in Table 5 of Appendix B (page 25). The consistently small standard deviations across the evaluation metrics indicate that the proposed approach achieves stable results despite potential variations in the learned ordering. These results demonstrate the robustness of NeuralRos across independent training runs.

To further assess the consistency of the learned Rosenblatt orderings, we have conducted an additional analysis across 30 independent runs with different random seeds. The results show only minor variations in the learned permutations, with approximately 2–3 changes in variable positions for SMD, 1–2 for MSDS, and 3–4 for SWaT, on average, indicating that the LRO module learns relatively consistent orderings across independent runs. We have included this additional analysis in Appendix B of the revised manuscript, under “Supplementary Ablation Studies” (page 27, lines 1250–1260), further demonstrating that small variations in the learned ordering do not affect the performance of NeuralRos.

 

Q2

We thank the Reviewer for this valuable suggestion.

In the original submission, we already included an ablation study comparing the proposed Learnable Rosenblatt Ordering (LRO) module with a fixed natural ordering, demonstrating the importance of learning the variable ordering. To address the reviewer's suggestion, we have extended this analysis by including an additional baseline based on marginal variance, where the time series are ordered according to their individual variances computed from the normal training data. The new results, presented in Table 7 of Appendix B (page 26), demonstrate that while variance-based ordering improves upon the ordinary fixed ordering, the proposed LRO module consistently achieves substantially lower HHD across all four datasets. Furthermore, as demonstrated by the computational cost analysis in the main manuscript (Table 3, page 10 of the revised manuscript), NeuralRos remains highly computationally efficient in terms of model size, training time, and inference time compared with competitive baselines. Therefore, we believe that the substantial performance improvements, together with the computational efficiency of the proposed method, justify the adoption of the LRO module.

 

Q3

We thank the Reviewer for this comment. As correctly observed, all methods achieve lower localization performance on the SWaT dataset. This may be partly attributable to the restrictive ground-truth labels, which predominantly identify only the primary sensors involved in each anomaly and do not account for other components that may be affected and exhibit anomalous responses.

  Finally, we would like to note that we have included three additional baseline methods for comparison in the revised manuscript to further strengthen our experimental evaluation.

Reviewer_p98WOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes NeuralRos, which jointly learns cross-series dependencies and localizes anomalies in multivariate time series containing both continuous and discrete variables. The authors argue that existing models can produce ambiguous per-series anomaly scores because they learn cross-series dependencies inaccurately, and they propose to use the Rosenblatt transformation as a learning mechanism rather than as a diagnostic tool.

  • Learnable Rosenblatt Ordering (LRO): Pairwise precedence scores between variables are computed from attention asymmetry, and the variable ordering of the Rosenblatt factorization is learned via Sinkhorn normalization and a straight-through estimator.
  • Neural conditioner:
    • The conditional distribution of each variable is modeled given its history and the current values of the variables preceding it in the ordering. A Gaussian mixture is used for continuous variables, and a Negative Binomial or discretized Gaussian for discrete variables.
    • The contribution of each source series is represented by a KAN-style function and a learnable weight γ\gamma, and these weights are collected into a dependency matrix Γ\Gamma.
  • Training: In addition to the conditional NLL, a uniformity loss and a mutual-independence loss (dHSIC) are applied to the Rosenblatt-transformed values. Discrete variables are handled through a randomized PIT.
  • Localization scores: NeuralRos-L is based on the conditional NLL, and NeuralRos-R is based on how close the transformed value is to the boundaries of [0,1][0,1].

Experiments cover two tasks: dependency identification on eight synthetic datasets (four continuous and four mixed-type), and anomaly localization on three real-world datasets (SMD, MSDS, SWaT). NeuralRos reports better results than the compared methods on both tasks.

Critical strengths

S1 (Quality): The paper handles mixed-type time series within a consistent framework.

  • A randomized PIT is used so that the transformed values of discrete variables also follow a uniform distribution, and this is proven in Appendix C.

S2 (Quality): The paper proposes a localization approach with practical value.

  • If well constructed, it provides a localization approach that can offer practically meaningful value in time-series anomaly detection.

S3 (Reproducibility): The variability of the results is reported through repeated experiments.

  • Means and standard deviations over 30 runs are reported for both dependency identification and anomaly localization.
Critical weaknesses

W1 (Significance): The problem statements are abstract, and the key claims lack supporting evidence.

  • The abstract asserts that ambiguous per-series anomaly scores arise "mainly" from inaccurate modeling of cross-series dependencies. However, ambiguity can arise even in well-trained models from anomaly propagation, variable ordering, the attribution method, or model capacity.
  • The following claims are not supported by evidence or references:
    • "As these systems become increasingly complex, they also become more vulnerable to faults (anomalies)"
    • "several methods cannot explicitly infer the dependencies learned by the model, limiting their practical value for system diagnosis"
  • The claim that directly using per-series anomaly scores leads to ambiguous localization is also not supported by concrete examples or experiments. Most models already produce per-time, per-variable scores, and the paper does not show why these are insufficient.
  • The paper does not connect specific failures of existing methods to specific design choices of the proposed method.

W2 (Soundness): Per-series attribution depends on the variable ordering.

  • The same joint distribution can be expressed as different products of conditional distributions depending on the variable ordering. Even if the overall likelihood is similar, which variable's conditional score an anomaly is attributed to can therefore change with the ordering.
  • LRO only selects a single global ordering that facilitates dependency learning on normal data, so it does not resolve the dependence of attribution on the ordering.

W3 (Soundness): The justification for interpreting γ\gamma as dependency is insufficient.

  • A larger γπi,j\gamma_{\pi_i,j} does not necessarily imply a larger actual influence on the observations or a larger causal effect.

W4 (Soundness): The properties of the localization scores for discrete and mixed-type variables are not analyzed.

  • For discrete variables, the NeuralRos-R score depends on an auxiliary random variable, so the same observation can receive a different score each time.
  • For highly probable values, the jump of the CDF is wide, so a large part of the score is determined by randomness. This is not analyzed.
  • NeuralRos-L uses −log⁡-\log density for continuous variables and −log⁡-\log pmf for discrete variables. These quantities are on different scales, so ranking them across variables is not a fair comparison. Even among continuous variables, the density depends on the units of each variable.

W5 (Originality): Differences from the closest prior work are not established, and the Related Work section is a list.

  • The Related Work section merely lists prior studies and does not clarify how the proposed method differs from them.
  • In particular, MTGFlow (AAAI 2023) is not cited. MTGFlow learns an inter-entity dependency graph via self-attention, estimates entity-wise densities, and localizes anomalies through entity-level anomaly scores.
  • There are also many studies on channel dependency in anomaly detection (e.g., CATCH), but the paper does not explain how its cross-series dependency learning differs from them.

W6 (Soundness): The design of LRO is insufficiently justified.

  • No justification is given that applying the head-specific transformation Φh\Phi_h only to the keys allows a meaningful comparison between each raw time series and the transformed representations of the other series.
  • Attention only indicates how relevant a key is to a query, so there is insufficient basis for interpreting its asymmetry as the "direction" of a dependency.
  • pt(i≺j)p_t(i\prec j) is obtained by applying a sigmoid to a directional score and is not a calibrated probability, so it should be interpreted as a probability with caution.

W7 (Quality): Justifications for the main design choices are not provided.

  • KAN: No reason is given for using KAN-style functions, and there is no comparison with alternatives such as MLPs.
  • Negative Binomial: The rationale for modeling discrete variables with a Negative Binomial distribution is not provided, and it is not specified which variables are treated as discrete in the real-world datasets.
  • dHSIC: There is no justification for choosing dHSIC to enforce mutual independence, nor evidence that it remains reliable when the number of variables is large.

W8 (Quality): The set of baselines is limited.

  • Recent general multivariate time-series anomaly detection methods (e.g., CATCH, D3R) are not compared.
  • Anomaly detection methods with built-in modules for learning inter-variable dependencies are also not compared. It is therefore difficult to judge whether the advantage of the proposed method stems from its dependency-learning approach itself.

W9 (Quality): Actual anomaly detection performance is not presented.

  • Even when localization is the goal, anomalies must first be detected accurately in practice.
  • However, the localization metrics are computed at the ground-truth anomalous time points, and the anomaly detection performance of NeuralRos is not reported.

W10 (Clarity): The readability and completeness of the paper are insufficient.

  • There are no figures illustrating the problem setting or the structure of the method, which makes the paper difficult to understand. The only figure in the main text shows ablation results.
  • The abstract only states that the experiments "demonstrate the effectiveness" without giving a representative result for each task or the baselines it improves upon. It also does not clearly state the scope of the contribution (enabling per-series diagnosis in mixed-type data).
  • The paper ends without a Discussion or Conclusion, so it does not summarize what has and has not been demonstrated.
Questions for authors

Q1 (W1): What is the evidence for the problem statements?

  • Is there experimental or other evidence showing that the ambiguity of per-series scores is "mainly" due to inaccurate dependency modeling?
  • Can you provide concrete examples in which the per-series scores of existing methods fail at localization?

Q2 (W2): How much does attribution depend on the variable ordering?

  • How much do the localization results change when a different ordering is used?

Q3 (W3): What does γ\gamma represent?

  • Can you show how γ\gamma corresponds to the magnitude of the actual influence on the observations?

Q4 (W6): What is the rationale for the design of LRO?

  • What justifies applying Φh\Phi_h only to the keys?
  • What is the basis for interpreting attention asymmetry as the direction of a dependency?

Q5 (W5): How does NeuralRos differ from the closest prior work?

  • What are the methodological differences and advantages of NeuralRos compared with MTGFlow, which learns a dependency graph and localizes anomalies through entity-wise densities?
  • How does cross-series dependency learning differ from channel-dependency studies in anomaly detection (e.g., CATCH)?
10 comments
AuthorsOpenReview ↗

Responses to W1, Points 1-3

W1:

  • Bullet point 1: The points raised by the reviewer largely reflect the motivation of our work and have already been discussed in the original manuscript. We are indeed considering well-trained models. However, a well-trained model does not necessarily guarantee an appropriate architecture with sufficient capacity to accurately capture the temporal and cross-series dependencies that characterize multivariate time series. We agree that anomaly propagation is a fundamental challenge. However, what makes per-series anomaly scores ambiguous is that, when cross-series dependencies are not estimated accurately , anomalies can propagate not only to dependent series but also to independent series. Therefore, the challenge we address is precisely the combination of anomaly propagation and inaccurately learned cross-series dependencies, which motivates our approach to explicitly controlling anomaly propagation through accurate modeling of these dependencies.

To address the Reviewer's comment and avoid potential misunderstanding, we have revised the Abstract and Introduction to explicitly clarify that the ambiguity in per-series anomaly scores arises from anomaly propagation combined with inaccurate modeling of cross-series dependencies, making it difficult to distinguish the actual anomaly source from other affected series.

  • Bullet point 2: Please note that both statements reflect well-established challenges in the literature [1,2,3]. The first concerns the increasing complexity of real-world systems and the associated challenges in fault diagnosis. To further address the reviewer's comment, we have included two references supporting this statement (see revised manuscript: line 30-32)

Regarding the second statement, widely used architectures, including Transformers, RNNs, and CNNs, implicitly learn cross-series dependencies but do not inherently provide explicit, interpretable representations of these dependencies. This is a well-recognized limitation of deep learning architectures that has motivated extensive research on interpretability [3]. Therefore, we consider both statements well grounded and do not believe they raise substantive concerns regarding our work. To further address the reviewer's comment, we have included additional references in the revised manuscript to explicitly support this statement (see revised manuscript: line 40-46).

References

[1] Soleimani et al., “Diagnostics and prognostics for complex systems: A review of methods and challenges,” Quality and Reliability Engineering International , 2021.

[2] Aldrini et al., “Fault diagnosis and self-healing for smart manufacturing: A review,” Journal of Intelligent Manufacturing, 2024.

[3] Barredo Arrieta et al., “Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI,” Information Fusion, 2020.

  • Bullet point 3: The original submission contained comparisons with SOTA methods that use per-series anomaly scores for localization, such as MEMTO, OmniAnomaly, DAEMON, and SARAD. We additionally compare NeuralRos with methods that employ techniques to refine anomaly scores, such as STAS-SFAS, and ALoRa-Loc, as well as methods specifically designed to better capture cross-series dependencies, such as AERCA. Our results demonstrate that NeuralRos consistently outperforms these methods.

To address the Reviewer's comment, we have included additional state-of-the-art methods, namely MTAD-GAT, a graph neural network-based approach designed to capture cross-series dependencies, ICDiffAD (ICLR 2026), and MTGFlow (AAAI 2023) which also produces per-series anomaly scores. The additional results further demonstrate the superior localization performance of NeuralRos. Furthermore, we have expanded the localization results discussion (pages 9-10) to explicitly compare NeuralRos with methods that rely on per-series anomaly scores for localization.

AuthorsOpenReview ↗

Responses to W1-point 4, W2, W3

  • W1, Bullet point 4: The Abstract and Introduction already state that a key limitation of existing methods in anomaly localization is their insufficient ability to accurately capture cross-series dependencies. As discussed above, inaccurately learned dependencies can lead to anomaly propagation across unrelated series, producing ambiguous localization scores. This directly motivates our approach, which transforms the Rosenblatt transformation from a statistical diagnostic tool into a learning mechanism for accurately modeling cross-series dependencies in mixed-type multivariate time-series data. Our experimental evaluation demonstrates that NeuralRos outperforms strong state-of-the-art methods in cross-series dependency identification (Table 1 and Figure 1 of the main manuscript, and Table 4 and Figure 6 of Appendix B) and achieves superior anomaly localization performance (Table 2 of the main manuscript), directly supporting the connection between our methodological design and the identified limitations of existing approaches.

 

W2

  • We thank the reviewer for this comment. However, we emphasize that this is precisely the motivation behind our Learnable Rosenblatt Ordering (LRO) module, which is specifically designed to address the dependence on variable ordering by learning an appropriate ordering directly from the data. This allows the framework to adapt the ordering to each dataset rather than relying on a fixed, predefined ordering across all datasets. Furthermore, as with any machine learning model, the learned parameters may vary across independent runs. We emphasize that we report the mean and standard deviation over 30 independent runs, whereas most existing anomaly diagnosis methods typically report results using at most five runs. Our results demonstrate the robustness of the proposed approach while also providing a comprehensive assessment of its performance variability across independent runs.

To further address the reviewer's concern, we have introduced a window-adaptive variant of LRO, in which ordering scores are computed separately for each input window rather than using a single global ordering. We evaluate this variant through an additional ablation study, with the results reported in Appendix B of the revised manuscript under “Supplementary Ablation Studies.” The results show improved or comparable anomaly localization performance across all evaluation metrics and datasets. Furthermore, for cross-series dependency identification, both ordering strategies achieve comparable results, indicating that window-adaptive ordering can improve localization without compromising dependency identification performance.

 

W3 :

By design, the learnable weight γπi,j\gamma_{\pi_i,j} controls the contribution of time series jj to the conditional modeling of time series πi\pi_i. Since the transformations ϕπi,j\phi_{\pi_i,j} are normalized to ensure comparable scales across series, the magnitude of γπi,j\gamma_{\pi_i,j} reflects the relative strength of the learned contribution of series jj to series πi\pi_i. Importantly, we empirically evaluate the effectiveness of these learned weights in identifying cross-series dependencies across multiple datasets. Our extensive comparisons with state-of-the-art methods, together with additional quantitative and qualitative analyses, demonstrate the effectiveness of our approach in cross-series dependency identification. This is quantitatively supported by the results in Table 1 (page 9) and the consistently low standard deviations across all evaluation metrics and datasets reported in Table 5 of Appendix B (page 25). Furthermore, the qualitative results presented in Figure 1 of the main manuscript and Figure 6 of Appendix B (page 27) provide additional evidence of the effectiveness of the proposed approach in accurately identifying cross-series dependencies.

To address the reviewer's concern, although this normalization was already described in the original submission, we have further clarified its purpose in Section 4.3 (page 6, lines: 270-275 of the revised manuscript.)

AuthorsOpenReview ↗

Responses to W4,W5

W4

  • Bullet point 1 & 2 : We thank the reviewer for this comment. We acknowledge that the NeuralRos-R score for discrete variables depends on an auxiliary random variable, as required by the randomized Rosenblatt transformation to preserve its theoretical uniformity property. Importantly, this randomization is restricted to the conditional CDF interval corresponding to the observed value, whose width equals its conditional probability. Therefore, observations with smaller conditional probabilities have narrower randomization intervals. Although this introduces variability in individual scores, our extensive experimental evaluation over 30 independent runs demonstrates the effectiveness and robustness of NeuralRos-R in anomaly localization (see Table 2 of the main manuscript). Furthermore, the consistently low standard deviations reported across the evaluation metrics (see Table 6 of Appendix B) provide additional evidence of the stability and robustness of the proposed approach across independent runs.

  • Bullet point 3: We thank the reviewer for this valuable observation. We acknowledge that probability densities and probability masses are not directly comparable and that continuous-variable densities depend on measurement units, potentially affecting cross-series localization rankings. To address this concern, we have introduced a per-series normalization of the negative log-likelihood scores using statistics estimated exclusively from normal training data. This allows each score to be evaluated relative to its corresponding series' normal behavior, improving comparability across continuous and discrete variables. In the revised manuscript, we have included a sensitivity analysis in Appendix B (Table 9, page 28) demonstrating that NeuralRos-L's localization performance is generally insensitive to the use of per-series normalization, although slight improvements are observed in some evaluation metrics when normalization is applied. To ensure better-defined and more comparable anomaly scores across time series, and given the slight performance improvements, we adopt per-series normalization as the default setting in our methodology and use it for the results reported in Table 2 of the revised manuscript.

 

W5

We thank the reviewer for this comment. However, we respectfully disagree with how this concern is presented. The Related Work section was structured to highlight the connection between the two complementary tasks addressed in our study: cross-series dependency learning and anomaly localization. Our motivation is to jointly address these tasks by emphasizing accurate dependency learning to improve anomaly localization.

While additional relevant studies can certainly be included, we aimed to cover the most representative methods within the scope of our work. To address this concern, we have expanded the Related Work section to include MTGFlow and CATCH and added a concluding paragraph explicitly discussing how NeuralRos differs from existing approaches. These revisions further clarify the novelty and methodological contributions of our work. The idea of NeuralRos, which transforms the Rosenblatt transformation into a learnable algorithm with learnable variable ordering, builds on the theoretical properties of the Rosenblatt transformation, and extends to mixed-type multivariate time-series data, is fundamentally different from existing approaches, and our experimental results demonstrate its effectiveness.

We have additionally included MTGFlow in our experimental evaluation for both cross-series dependency identification and anomaly localization. NeuralRos outperforms MTGFlow across most datasets and metrics. Nevertheless, MTGFlow proves to be a strong baseline, often achieving the second-best performance and outperforming several other methods. We thank the reviewer for suggesting its inclusion, as it further strengthens our experimental evaluation.

Regarding CATCH, we were unable to include it in our experimental evaluation due to time constraints. We prioritized MTGFlow because its architecture and publicly available implementation allowed us to more readily adapt it to the datasets and evaluation metrics used in our study, whereas CATCH was originally designed for anomaly detection and would require more extensive modifications to support both tasks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.