acceptodds
Under review as a conference paper at ICLR 2027

Anomaly Detection for Sparse and Irregular Multivariate Time Series with Latent SDEs

Abstract

Multivariate time series anomaly detection (MTSAD) is critical for a wide range of application areas, such as industrial monitoring, cybersecurity, or healthcare. Real-world data is often sparse, irregularly sampled or partially observed, yet existing methods assume uniformly sampled time series. We propose a generative approach based on Latent SDEs that projects the observed time series on a continuous-time stochastic dynamical system, directly being able to handle missing observations and irregular sampling, while also naturally capturing possible cyclic behavior that many real-world use cases inherently possess. Experiments on six anomaly benchmark datasets show that our proposed method ranks first among state-of-the-art baselines. We further demonstrate that our method remains robust under severe data sparsity, while performance significantly degrades for the tested baseline methods. These results highlight latent SDEs as a natural inductive bias for anomaly detection in multivariate time series, especially in presence of real-world irregularities.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Time Series Anomaly Detection, SDE, Irregular Multivariate Time Series

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

This paper introduces Latent Stochastic Differential Equation (LSD) for anomaly detection. It is a reconstruction-based generative framework for multivariate time-series anomaly detection, which primarily models normal data using continuous-time stochastic dynamics in a low-dimensional latent space. The method is built upon latent SDEs on homogeneous spaces and employs masked likelihoods to handle irregular sampling and missing data, thereby avoiding the need for imputation or resampling. The paper also investigates two types of latent geometric structures: Euclidean spaces and hyperspheres. Experiments on six widely used benchmark datasets and one industrial dataset with recurring dynamic patterns demonstrate that LSD achieves a competitive overall ranking.

Critical strengths
  1. Leverageing latent SDEs on homogeneous spaces for Multivariate Time Series Anomaly Detection (MTSAD), to my knowledge thus far, is the first time.
  2. Broad benchmark coverage. Ablations and runtime notes (albeit limited) and a sparsity study that concretely tests the core claim of handling irregular/sparse data.
  3. The used codebase seems available, which is good for reproducing the experiment.
Critical weaknesses
  1. I find that the proposed method essentially amounts to adapting the algorithm in [1] to a simple time-series anomaly detection framework. Moreover, Neural/latent SDE/CDE approaches for irregular time series [2,3,4] have already been extensively studied. Therefore, neither the motivation nor the proposed method itself offers the authors' claimed novelty.
  2. A bigger issue lies in the organization and storytelling of the paper. First, regarding the Introduction section, I think most of the preceding content overly redundant, especially considering the subsequent Related Work section. In contrast, only a small portion at the end introduces the contributions without providing sufficient methodological motivation or context. I suggest substantially shortening the earlier discussion to make room for content that is more important to the overall story, such as why the authors came up with this particular approach and what limitations of other latent ODE methods motivate it. Second, regarding the Method section, I question the necessity of the first three paragraphs, which mainly provide an overview. Their length is almost comparable to that of the formal description in Section 3.2, while the latter is overly concise. I suggest that the authors refer to [1] and revise and elaborate on the algorithmic details. Overall, the paper currently requires a major restructuring of its content and presentation.
  3. Likelihood-based reconstruction scoring is known to suffer from calibration/pathologies in deep generative models; while the weak decoder and structured dynamics mitigate this, the paper does not probe failure modes or compare to compliance/GOF-based tests that can be more robust.
  4. Computational cost of latent SDE integration is high; solver and step-size choices, stability, and trade-offs are not deeply examined in the manuscript.
  5. In sparsity experiments, baselines are evaluated only after imputation, which may hand LSD an advantage; a fairer setup would include baselines that can ingest irregular data directly. Best-F1 is selected on test by sweeping all thresholds (optimistic). Segment-level metrics or fixed-thresholding (e.g., POT/MAD) could better reflect deployment use.

References:
[1] Zeng, Sebastian, Florian Graf, and Roland Kwitt. "Latent sdes on homogeneous spaces." Advances in Neural Information Processing Systems 36 (2023): 76151-76180.
[2] Verma, Yogesh, Markus Heinonen, and Vikas Garg. "Climode: Climate and weather forecasting with physics-informed neural odes." International Conference on Learning Representations. Vol. 2024. 2024.
[3] Jeon, Jinsung, et al. "GT-GAN: General purpose time series synthesis with generative adversarial networks." Advances in Neural Information Processing Systems 35 (2022): 36999-37010.
[4] Jhin, Sheo Yon, Jaehoon Lee, and Noseong Park. "Precursor-of-anomaly detection for irregular time series." Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining. 2023.

Questions for authors
  1. See the Critical Weaknesses.
  2. Was min-max scaling fit strictly on training data (per-variable) and then applied to validation/test, or on the full dataset? Please clarify to rule out leakage.
  3. To what extent do LSD’s gains at 1% retention persist if baselines that can handle irregular sampling without imputation are included? If you re-run with such baselines, how do rankings change?
4 comments
AuthorsOpenReview ↗

Response to Reviewer TWUM - Part 1

We thank the reviewer for the thorough and critical review. We take the concerns seriously and have revised the manuscript substantially in response: we restructured the introduction and method section, now motivating the approach through the limitations of latent NeuralODEs and stating explicitly which components are adopted from prior work; we added a latent NeuralODE as a baseline that operates directly on irregular data, an ablation isolating the stochastic component, and a qualitative analysis of failure modes; and we clarified the solver, the normalization, and the evaluation protocol. We address each point below.


▶ W1: Novelty relative to Zeng et al. (2023) and irregular time series modeling

We agree that LSD builds on the latent SDE framework of Zeng et al. (2023), and that the modeling of irregular time series with NeuralODEs, SDEs and CDEs has been studied extensively. However, our goal is not a new time series model, but address a different research question: We use the latent SDE as a tool for multivariate time series anomaly detection: modeling the benign time series is the surrogate, unsupervised task, and the learned model is then used to measure how likely each observation is under the learned latent dynamics. The works in [2] and [3] address forecasting and generative synthesis, respectively, and do not consider anomaly detection.

Within anomaly detection, to the best of our knowledge, LSD is the first approach based on neural differential equations that operates on sparse and irregularly sampled data without imputation or interpolation as a preprocessing step. Methodologically, beyond adopting the framework of Zeng et al. (2023), our contributions are the mask-based training and scoring on partially observed inputs, the instantiation and comparison of two latent geometries (the realization of the GL(n)GL(n) action on Rn\mathbb{R}^n, complementing the original SO(n+1)SO(n+1) action on Sn\mathbb{S}^n), a deliberately weak decoder, and a point-wise anomaly score based on the expected conditional log-likelihood (Section 3.3, Eq. 2). We now state explicitly in Section 3 which components are adopted from Zeng et al. (2023) and which are new.

Regarding [4] and Sun et al. (2026), we considered both as baselines. However, both methods construct a continuous input path by cubic spline interpolation in a first stage, which reintroduces exactly the interpolation step LSD is designed to avoid and which would place them in the same position as the imputed baselines in the sparsity experiment (see also Q3). We agree, though, that a simpler model that also operates directly on the sparse data is the right reference to justify our more complex one. We have therefore added the latent NeuralODE of Rubanova et al. (NeurIPS 2019) as a baseline on all benchmarks and in the sparsity experiment, sharing the recognition network, decoder, and anomaly score of LSD. LSD outperforms it on all benchmarks except QAPPD with full observations, where it ranks between our two LSD variants; under burst subsampling with 1 % of the data retained, it falls behind both. To further separate the contribution of stochasticity from that of the structured dynamics, we added an ablation that removes only the diffusion term (Appendix C.3, see also our response to Reviewer xop4, W3). On QAPPD, the stochastic model outperforms its deterministic counterpart for both geometries, while on PSM the picture is mixed.


▶ W2: Organization and storytelling We thank the reviewer for this comment. The structure of the paper reflects a trade-off between different needs: being accessible to readers from the anomaly detection community who are less familiar with latent SDEs, being self-contained, and avoiding redundancy for experts. We aimed to preserve the overall structure, but we agree that it can be improved in the direction suggested by the reviewer. We therefore revised the manuscript as follows:

  • Introduction: We shortened the general discussion and now motivate the approach more directly, including the limitations of latent NeuralODEs that motivate the use of structured latent SDEs.
  • Related work: We removed the paragraph on challenges in industrial time series, which repeated the motivation of the introduction, so that the related work now focuses on MTSAD approaches and neural differential equations.
  • Method: We shortened the overview paragraphs at the beginning of Section 3 to a brief roadmap, which also states which components are adopted from Zeng et al. (2023) and which are new.
  • Algorithmic details: Following [1], we elaborated the formal description of the method, including the training objective and a dedicated section on the anomaly score (Section 3.3, Eq. 2), and added an algorithm for training and scoring (Algorithm 1, Appendix B).

With these changes, we hope to have found a structure that serves both readers new to the field and experienced readers.

AuthorsOpenReview ↗

Response to Reviewer TWUM - Part 2

▶ W3: Likelihood-based scoring and failure modes We thank the reviewer for this suggestion. First, the known pathologies of likelihood-based detection (e.g., Nalisnick et al., 2019) mainly concern the marginal likelihood p(x) of deep generative models, which can be dominated by low-level input statistics. Our score is instead the expected conditional log-likelihood of the observations given posterior sample paths under a Gaussian emission model with fixed variance. As shown in Eq. 2, it decomposes into the deviation from the mean reconstruction and the variance of the reconstructions across sample paths, and is therefore not inflated by low input complexity in the same way. We agree, however, that reconstruction-based scores share a related failure mode: anomalies that leave the observed dynamics largely unchanged remain well reconstructed.

To probe this, we added a qualitative analysis on QAPPD (Appendix C.4), where the anomaly types are documented by Nosrati et al. (2027), evaluating both LSD variants and the latent Neural ODE per trace and anomaly type. Delay anomalies are detected reliably by all methods (AUC between 92 and 99), and pick-up position anomalies are detected well on most traces but less reliably on one. Voltage anomalies, in contrast, are not detected by any method (AUC between 53 and 57), which we attribute to the voltage deviations having only a marginal effect on the observed trajectories. This is precisely the failure mode described above, which we now discuss as a limitation in Section 5. As this anomaly type occurs in a single trace only, we refrain from broader conclusions.

We agree that conformal and goodness-of-fit-based tests are a promising direction, as they could also flag observations that are reconstructed suspiciously well. As the per-variate scores and sample paths of LSD provide a natural basis for such tests, we now mention them as future work.


▶ W4: Computational cost of latent SDE

Following Zeng et al. (2023), we use a fixed-step geometric Euler–Maruyama scheme for both the SO(n+1)SO(n+1) and the GL(n)GL(n) variant. Each step applies a matrix exponential to the latent state, preserving the group structure at any step size: the update matrix is orthogonal for SO(n+1)SO(n+1) and invertible for GL(n)GL(n). The step size trades discretization accuracy against the number of updates, and reparameterizing the Gaussian increments allows backpropagating directly through the updates, so neither adaptive step-size control nor a separate adjoint solve is required.

Training is dominated by the forward and backward pass through the solver and scales linearly with the number of solver steps, sample paths, and the window length. Table 8 (Appendix C.1) reports the resulting training times, which are substantially higher than for the baselines; scoring requires no backpropagation and is considerably cheaper. We acknowledge this as the primary limitation of LSD in Section 5.


▶ W5: Fairness of the sparsity experiment and thresholding

Baselines after imputation. We agree, and have added the latent Neural ODE of Rubanova et al. (2019), which ingests irregular data directly, to the sparsity experiment; see W1 and Q3.

Best-F1 thresholding. We follow the protocol of Sarfraz et al. (ICML 2024), which selects the F1-maximizing threshold, as is common in the literature. We agree that this is optimistic and does not reflect deployment, where thresholds are application dependent. For this reason, two of our three metrics, AUC-ROC and AUC-PR, are threshold-free, accounting for two thirds of the ranking in Table 3.


▶ Q2: Min-max scaling

The min-max scaling is fitted per variable exclusively on the training data and applied unchanged to the validation and test data, which rules out leakage through this step.


▶ Q3: Gains at 1 % retention with baselines that handle irregular sampling natively

The latent Neural ODE of Rubanova et al. (2019) is such a baseline: it operates directly on the irregularly observed data and is evaluated under the same protocol as LSD. On QAPPD with full observations, it ranks between our two LSD variants; under burst subsampling down to 1 %, however, it falls behind both. Hence, LSD's gains in the sparsity experiment are not merely an artifact of imputing the baselines (Section 4.2).


We thank the reviewer again for the detailed feedback, which has substantially improved both the presentation and the empirical analysis of the paper. We acknowledge that LSD builds on an existing model class; we hope, however, that the revised manuscript makes clearer why this model class is well suited for anomaly detection on sparse and irregular data, and how our design choices and evaluation contribute in this setting. We would be grateful if the reviewer could take these revisions into account in the final assessment.

Thank the authors for their response. The writting of the manuscript has now improved (while I suggest highlighting the revised parts in a different, more distinguishable color, e.g., in blue). Some of my concerns have also been addressed, and I have accordingly increased my rating.

We sincerely thank the reviewer for the follow-up and for increasing the rating. Regarding the highlighting of revisions, we had assumed that the automatically generated pdfdiff, which according to the ICLR 2027 author guidelines compares each revision against the original submission, would serve this purpose, and therefore did not additionally mark the changes in color. We apologize for any inconvenience this may have caused. For orientation, the main revisions concern the introduction and the method section (Sections 1 and 3), the latent Neural ODE baseline in all experiments (Sections 4.1 and 4.2), and the new analyses in Appendices C.3 to C.6.

Reviewer_xop4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent SDE Anomaly Detection (LSD), a framework for unsupervised multivariate time-series anomaly detection. The authors combine mask-aware inference with structured stochastic dynamics in Euclidean and hyperspherical latent spaces, reformulating anomaly detection as evaluating observation likelihoods under learned latent trajectories. Experiments on multiple benchmark datasets show competitive results against classical and deep-learning methods, particularly under sparse observations.

Critical strengths
  1. Modeling sparse and irregular time series through continuous-time stochastic dynamics is an improvement over conventional methods that rely on imputation or resampling.
  2. The authors demonstrate the effectiveness of the framework across multiple benchmarks and under different sparsity levels, while the use of different latent geometries improves its flexibility and applicability.
Critical weaknesses
  1. Although the paper proposes a new anomaly detection framework, existing continuous-time methods for irregular time series are not included as primary experimental competitors, leaving the comparison with closely related approaches insufficiently explored.

  2. I am concerned that the choice of latent geometry is highly dependent on the dataset, as the hyperspherical variant performs better under extreme sparsity but worse with full observations, potentially limiting its generalizability to unseen scenarios.

  3. The ablation study does not isolate the contribution of stochasticity, as the Neural ODE baseline simultaneously changes the latent geometry and dynamics parameterization.

  4. There are some inconsistencies in the paper, e.g., Table 3 appears to mislabel PCA as kNN, and the reproducibility statement references a nonexistent Section 3.3.

Questions for authors

Q1. The authors consider Euclidean and hyperspherical latent spaces and argue that the latter provides a beneficial inductive bias for repetitive dynamics. Is there any empirical justification for selecting the appropriate latent geometry for unseen datasets?

Q2. The paper defines anomaly scores using the expected negative log-likelihood under an approximate posterior conditioned on the input observations. Could the authors explain how this scoring mechanism reliably distinguishes anomalous behavior from normal variations?

4 comments
AuthorsOpenReview ↗

Response to Reviewer xop4 - Part 1

We thank the reviewer for the careful review and concrete suggestions. In response, we added a latent Neural ODE as a continuous-time baseline on all benchmarks and in the sparsity experiment (W1), added ablations of the stochastic component and the decoder capacity (W3), and corrected the reported inconsistencies (W4). We address all points below.


▶ W1: Continuous-time methods for irregular time series as competitors

We have added the latent Neural ODE of Rubanova et al. (NeurIPS 2019), designed specifically for irregularly sampled time series, as a baseline. It is evaluated under the same protocol as LSD, including the likelihood-based score, on all seven benchmarks and in the sparsity experiment. LSD outperforms it on all benchmarks except QAPPD with full observations, where it ranks between our two LSD variants; under burst subsampling with 1 % of the data retained, it falls behind both (Sections 4.1 and 4.2; see also our response to Reviewer t5Wx, W1).

We also considered the continuous-time anomaly detection methods discussed in our related work (Jhin et al., 2023; Sun et al., 2026). However, these require evenly sampled inputs or an interpolation step in preprocessing, which reintroduces exactly the step LSD is designed to avoid, and/or lack an implementation that allows reproducing their results in our setting. A latent Neural ODE under an identical scoring protocol is therefore, in our view, the closest reproducible competitor.


▶ W2: Dependence of the latent geometry on the dataset

Indeed, the latent geometry influences performance, and different geometries can favor different types of data; determining the appropriate geometry a priori is non-trivial, which we discuss as an open question in Section 5. In our experiments, neither variant dominates on all benchmarks, but the differences are mostly small and a less suitable choice does not destabilize the model: switching the geometry changes the results only moderately (Tables 1, 2, and 4), and in the overall ranking (Table 3), the two LSD variants rank first and second.

The baselines are considerably more sensitive to the data characteristics: classical methods perform well on single-trace but poorly on multi-trace benchmarks, whereas the opposite holds for most deep learning methods, while both LSD variants lead in both settings (Table 3). Similarly, under extreme sparsity (Tables 5 and 9), all methods degrade, but both LSD variants retain the top ranks, whereas the ranking of the baselines changes substantially.

We therefore keep the geometry as a hyperparameter that allows encoding prior knowledge, with Rn\mathbb{R}^n as a safe default. In both variants, the drift is constrained by the group action on the latent space, which makes the dynamics more controllable than in a vanilla latent Neural ODE with an unconstrained MLP vector field.


▶ W3: Isolating the contribution of stochasticity

We agree that the latent Neural ODE changes both geometry and parameterization of the dynamics. We have therefore added an ablation that removes only the diffusion term from LSD, keeping geometry, recognition network, decoder, and training protocol identical, for both Rn\mathbb{R}^n and Sn\mathbb{S}^n, as well as a decoder ablation replacing the weak decoder by an over-parameterized one. Both are performed on PSM and QAPPD (Appendix C.3); the latent Neural ODE is now a baseline in all experiments instead.

  • Stochasticity. On QAPPD, the stochastic model outperforms its deterministic counterpart on all metrics for both geometries. On PSM, the picture is mixed: for Sn\mathbb{S}^n, the stochastic model is ahead on AUC and F1 and ranks better overall, whereas for Rn\mathbb{R}^n the deterministic variant performs better. Beyond performance, we retain the stochastic formulation because it contains the deterministic model as the special case of zero diffusion and represents variability of normal behavior along the trajectory.

  • Decoder capacity. On PSM, the weak decoder outperforms the over-parameterized one on all metrics for both geometries. On QAPPD, the differences lie within one standard deviation: for Rn\mathbb{R}^n, the weak decoder is ahead on AUPRC and F1, for Sn\mathbb{S}^n the stronger decoder on AUC and F1. In line with Chen et al. (ICLR 2017), we attribute the advantage on PSM to the expressive decoder modeling the training data without relying on the latent representation, whereas the restricted decoder forces the latent SDE to capture the underlying dynamics (see also Q2).

AuthorsOpenReview ↗

Response to Reviewer xop4 - Part 2

▶ W4: Inconsistencies

We thank the reviewer for spotting these and apologize for them. The mislabeling of PCA as kNN in Table 3 and the incorrect section reference in the reproducibility statement have been corrected. We additionally proofread the manuscript for similar issues.


▶ Q1: Selecting the latent geometry for unseen datasets

As discussed in W2, the Euclidean variant Rn\mathbb{R}^n serves as a safe default, and the performance difference between the two geometries is small throughout our experiments, so a suboptimal choice comes at limited cost. The hyperspherical variant is motivated when prior knowledge indicates bounded, recurrent dynamics, such as repetitive production cycles. Beyond this, the geometry can be treated as a hyperparameter and selected without anomaly labels, e.g., via the likelihood on a held-out split of benign training data. For a discussion of the structural requirements and further admissible latent geometries, we kindly refer the reviewer to our response to Reviewer t5Wx, Q2.


▶ Q2: How does the score distinguish anomalies from normal variation?

LSD is trained in an unsupervised manner: only benign, non-anomalous data is used for training and validation. From this data, LSD learns a distribution over latent dynamics, and the score measures how likely each observation is under these learned dynamics: for each observed entry, it is the expected negative log-likelihood under the approximate posterior conditioned on the context window, estimated by Monte Carlo over 5 sampled latent paths. Normal variation, i.e., variability present in the benign training data (such as noise), different operating states or recurring mode switches, is accounted for by the learned dynamics and the emission model. Behavior that is outside the variations observed in the training data, and therefore inconsistent with the learned dynamics, cannot be explained by any plausible latent trajectory and receives a low likelihood. Of course, this relies on the two assumptions that variations within the training data are representative, and that they can reliably be modeled by the latent SDE. Two design choices support the latter: the low-dimensional latent space and the deliberately weak decoder restrict the model's capacity to reconstruct arbitrary inputs, counteracting the tendency of reconstruction-based detectors that otherwise also reconstruct anomalies well. Consistently, replacing the weak decoder by an over-parameterized one degrades detection performance on PSM, while on QAPPD the differences lie within one standard deviation (Appendix C.3, see W3).

As for any unsupervised detector, dynamics that are absent from the training data are flagged as anomalous by definition. In practice, this assumes that the training data was curated, e.g. by a domain expert; that this assumption does not always hold even for established benchmarks is discussed in detail by Wu and Keogh (IEEE TKDE, 2023).

Empirically, the separation is reflected in the reported F1 scores, which already account for false positives and false negatives through precision and recall. For convenience, we additionally report the false positive rate (FPR) and false negative rate (FNR) on QAPPD at the F1-maximizing threshold (in %, mean ± std over 5 runs):

ModelFPR ↓FNR ↓
LSD on Rn\mathbb{R}^n (ours)6.62 ± 2.5933.45 ± 2.98
LSD on Sn\mathbb{S}^n (ours)9.17 ± 1.9836.63 ± 0.13

Both LSD variants keep the FPR below 10 % while detecting about two thirds of the anomalous observations.


We are grateful for the reviewer's comments, which led to a more rigorous evaluation, in particular through the continuous-time baseline and the targeted ablation of the stochastic component. We hope that these additions, together with the clarifications on the latent geometry and the anomaly score, resolve the reviewer's concerns, and we would kindly ask the reviewer to consider them in the final assessment. We are happy to provide further details if needed.

Thank you for the additional experiments. Two concerns still require direct empirical evidence:

  1. W2/Q1 Geometry selection. Higher likelihood on benign validation data does not necessarily imply better anomaly detection. A controlled experiment comparing benign variability, such as normal noise or legitimate mode changes, with genuine anomalies could help clarify whether the score reliably distinguishes the two.
  2. Q2 Normal variability versus anomalies. Since the posterior conditions on the observations being scored, it may also adapt to anomalous inputs. The weak-decoder argument provides a motivation but does not establish the required separation. Evaluating the proposed selection criterion across datasets and sparsity levels could help establish whether it provides reliable guidance for choosing the latent geometry.

We thank the reviewer for the follow-up and agree that both points require empirical evidence. We have added two experiments to the revised appendix, which we summarize below. As the suggested experiments appear to be swapped between the two points, we address each concern with the experiment that targets it.

W2/Q1: Geometry selection (Appendix [C.5]). We evaluate whether the validation loss reliably selects the better latent geometry across datasets and levels of data availability. For each of the 10 QAPPD traces at 100 %, 5 %, and 1 % of the data retained (five seeds each), and for PSM (three seeds), we compare the validation loss and the validation negative log-likelihood of both geometries with their test AUC and AUPRC. Both validation criteria always select the same geometry. On QAPPD, the selected geometry also achieves the better test performance in the following number of cases (out of 50 per setting):

Data retainedAUCAUPRC
100 %34/5028/50
5 %33/5030/50
1 %29/5032/50

Overall, the selection agrees in 96 of 150 cases (64 %) for AUC and 90 of 150 cases (60 %) for AUPRC. On PSM, it agrees for AUPRC but not for AUC. In all cases where the criterion selects the worse geometry, its performance remains within one standard deviation of the better one. We therefore conclude that the validation loss provides guidance better than chance, but not a reliable selection criterion; given that a suboptimal choice comes at limited cost, we recommend the Euclidean variant as a default. We have revised the corresponding statement in the manuscript accordingly.

Q2: Normal variability versus anomalies (Appendix [C.6]). Simulating realistic benign variability, such as legitimate mode changes, in a controlled way is difficult, as it requires knowledge of the underlying process beyond the recorded data. As a controllable proxy, we add Gaussian noise with standard deviation 0.05, 0.1, and 0.2 (relative to the normalized feature range) to the benign test data of QAPPD and compare the resulting scores with those of the genuine anomalies, per trace. We keep the decision threshold fixed at the value determined on the noise-free data and report the resulting false positive rates. On the original data, benign and anomalous scores are clearly separated on all traces except trace 7, which contains only voltage anomalies and is discussed as a failure case in Appendix C.4. Notably, this separation holds although the posterior is conditioned on the anomalous windows themselves, which indicates that the adaptation of the posterior to anomalous inputs is limited in practice. With increasing noise, the benign scores shift upwards: at 0.05, the distributions remain largely separable on most traces, although the false positive rate at the fixed threshold already increases noticeably on some traces; from 0.1 onwards, it increases sharply, and at 0.2, the distributions largely overlap. This is expected for a likelihood-based score, as noise beyond the variability in the training data is inconsistent with the learned model of normal behavior. The experiment shows that the ranking of scores is reasonably robust to small additional noise, whereas a fixed threshold is not; in deployment, the training data should therefore reflect the expected noise level, or the threshold should be recalibrated when it changes.

We thank the reviewer again for these suggestions, which have helped clarify both the reliability of the geometry selection and the behavior of the score under benign variability.

Reviewer_t5WxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Latent SDE Anomaly Detection (LSD) a novel method for unsupervised anomaly detection in multivariate time-series. The proposed methodology leverages a low-dimensional latent space SDE to reconstruct the observed time series. The learning problem is cast into this latent space and the latent SDE is well suited for irregular time series. The unsupervised anomaly score is based on reconstruction likelihood given sampled path from the trained latent SDE, aggregated over a window per-variate. The empirical evaluation is based on standard diverse benchmarks (SWaT, WaDi, PSM, MSL, SMAP, SMD) and the more specialized QAPPD benchmark (benchmark based on a real cyber-physical system performing a repetitive task). The baselines are broad and cover a diverse set of family of methods. Across the multiple datasets, LSD appear as best, second, or third best and vast majority of metrics.

Critical strengths
  1. Well motivated and principled novel formulation of multivariate anomaly detection for irregular timesteps. The construction incorporates masking and observation times directly into the model. The latent SDE formulation serves the intended purpose by construction.
  2. Well developed mathematical and probabilistic modeling choices and design.
  3. Strong empirical evidence that the approach is useful and outperforms strong and varied baselines.
  4. Broad and transparent empirical evaluation.
Critical weaknesses
  1. The experiments in the main text weakly isolates the advantaged of latent SDE for irregular timeseries modeling. I would be interested to see some discussion of the comparison with "native" irregular-time alternatives such as latent ODEs, Neural CDEs, etc and evaluated on QAPPD on the main text.
  2. Amount of empirical evidence evaluating distinct regimes of robustness/sparsity and irregularity of observations. Table 5 covers QAPPD at 1% and demonstrates better relative rank for LSD, and on the appendix for 5%. In my understanding this means randomly sampled irregular point, but it could be interesting to include some forms of structural irregularity (random segments of fixed size for example, or with size randomly selected in an interval, etc).
Questions for authors
  1. Could you run the method against the benchmark of the TSB-AD leaderboard https://thedatumorg.github.io/TSB-AD/#leaderboard and report how it performs compared to top performers?
  2. How central to the method is the Euclidean or hyperspherical geometry? Would the method continue to make sense and work with others spaces?
1 comment
AuthorsOpenReview ↗

Full Response to Reviewer t5Wx

We thank the reviewer for the positive and constructive assessment, and for recognizing the principled formulation, the probabilistic modeling design, and the breadth of the empirical evaluation. Following the reviewer's suggestion, the revised manuscript now includes a latent NeuralODE as a native irregular-time baseline across all benchmarks and in the sparsity experiment, and describes the structured masking procedure more prominently. We address the individual points below.


▶ W1: Comparison with native irregular-time alternatives

We agree that a direct comparison with native irregular-time alternatives better isolates the contribution of the latent SDE formulation. We have therefore added the latent NeuralODE approach of Rubanova et al. (NeurIPS 2019) as an additional baseline, evaluated on all seven benchmark datasets including QAPPD. This model is designed specifically for irregularly sampled time series and, like LSD, consumes the observations and their timestamps directly; it models deterministic latent dynamics in Euclidean space. LSD outperforms the latent NeuralODE on all benchmarks except QAPPD with full observations, where the latent NeuralODE ranks between our two LSD variants; under burst subsampling down to 1 %, however, it ranks behind both. We have added the results to the experimental evaluation of the revised manuscript, both in the main comparison (Section 4.1) and in the sparsity experiment (Section 4.2), and discuss them accordingly.


▶ W2: Regimes of sparsity and structural irregularity

We believe there is a misunderstanding here, which indicates that this part of the manuscript needs to be presented more clearly. The masks in the sparsity experiment are not sampled as independent random points; to the contrary, the masks are structured to model typical bursts of missing data. The details are given in Appendix B.6: the mask of each variate follows a two-state Markov chain: lengths of consecutive missing observations and lengths of consecutive available observations are drawn from geometric distributions, with an additional parameter controlling the burst length. This produces missing segments of random length, independently per variate, i.e. both bursty and asynchronous across features. It is designed to simulate sensor failures and transmission gaps as observed in real industrial systems. We have revised the manuscript to describe this procedure more prominently in the main text.


▶ Q1: TSB-AD leaderboard

We thank the reviewer for this valuable input. Five of the seven benchmarks in our evaluation (SWaT, PSM, SMAP, MSL, SMD) are already part of the multivariate TSB-AD-M suite; only WaDi and QAPPD are not. During dataset selection, we specifically focused on benchmarks representative of cyber-physical systems, which is the application domain targeted by LSD. Should the reviewer consider a further aspect or dataset of TSB-AD-M particularly relevant for our setting, we would be grateful for the suggestion.


▶ Q2: Role of the latent geometry

LSD does not depend on one particular geometry. It only requires the latent space to be homogeneous, i.e. a manifold on which a matrix Lie group acts transitively. Using the Lie algebra and the exponential map, SDEs and driftless prior processes of the required form can then be defined on the group and transferred to the latent space. In addition, a suitable family of initial distributions is needed, preferably one with a closed-form KL divergence to the prior distribution.

In this work, we use the SO(n+1)SO(n+1) action on Sn\mathbb{S}^n and the GL(n)GL(n) action on Rn∖{0}\mathbb{R}^n \setminus \lbrace 0 \rbrace. Extensions to other settings are possible in principle, such as the Rn\mathbb{R}^n action on the flat torus, the O(n)O(n) action on real projective space, or the SO0(n,1)SO_0(n,1) action on hyperbolic space, but we have not implemented or evaluated them.


We thank the reviewer again for the helpful suggestions, which have strengthened the empirical comparison in particular. We hope that the additional baseline and the clarification of the masking procedure address the raised concerns, and we would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.