acceptodds
Under review as a conference paper at ICLR 2027

Kurtosis-Guided Denoising Score Matching for Tabular Anomaly Detection

Abstract

Denoising models provide a way to learn data distributions from noise-corrupted samples. We show that at a fixed default noise scale, denoising score matching (DSM) provides strong tabular anomaly detection by evaluating the score norm on clean inputs in one forward pass. However, its performance depends on the noise scale: insufficient noise leaves sparse regions poorly represented, while excessive noise blurs local structure. A shared scale also overlooks differences in marginal shape, and tuning this trade-off is difficult without anomaly labels. We introduce kurtosis-guided denoising score matching (K-DSM), which allocates feature-specific noise scales from rearranged marginal histograms of the training data. Rearrangement separates marginal concentration from the locations of modes, while a local coverage analysis motivates assigning larger noise scales to profiles with higher kurtosis. The network trains on the original joint observations and scores clean inputs by the norm of its predicted correction in noise units. On 57 ADBench datasets with normal-only training, K-DSM raises mean AUC-PR from 0.585 to 0.627 and outperforms single-scale DSM at every tested base noise scale. It achieves the highest mean AUC-PR among evaluated baselines, with competitive AUC-ROC, while requiring one forward pass per input compared with fifteen for the strongest AUC-PR baseline. For contaminated training data, an exponential moving average teacher excludes likely anomalies from training updates, raising K-DSM’s mean AUC-PR from 0.189 to 0.343 while preserving one-pass inference.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
anomaly detection, denoising score matching, tabular data, kurtosis, score matching

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_R3HgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies tabular anomaly detection and proposes assigning fixed, feature-specific training noise scales based on the kurtosis of rearranged marginal histograms. The model learns the joint data structure through denoising training and performs single-pass scoring using the magnitude of its predicted correction in noise units. The authors also introduce EMA-teacher filtering to reduce the influence of anomalous samples in the training data.

Critical strengths

The problem has practical significance. Reducing inference cost while maintaining good detection performance is important for anomaly detection. The paper’s single-pass scoring procedure has clear value for deployment. The paper is well written and generally clear.

Critical weaknesses

1.The main methodological contribution is to compute kurtosis after rearranging marginal histograms and use it to allocate noise scales. The control experiments in the main text support the effectiveness of feature-specific noise allocation. My question is: why is it necessary to use kurtosis computed after rearrangement? The shared-scale and shuffled-scale controls examine the benefits of using different scales and assigning them to corresponding features, but they do not separately assess the necessity of rearrangement. I would like the authors to explain: with all other settings held fixed, how does performance differ between using ordinary marginal kurtosis and using kurtosis after rearrangement? 2.The rearrangement-based coverage analysis in Section 2.3 provides a reasonable motivation for the method. However, the affine kurtosis-based calibration relies on a specific near-Gaussian shape assumption, while the actual method applies this approximation to a broader range of distributions and incorporates empirical clipping and special-case handling. More importantly, coverage is an intermediate objective. The authors note that increasing noise may improve training coverage but may also weaken anomaly signals, and that marginal coverage does not guarantee sufficient support in the joint space. Therefore, even if calibration improves coverage, this alone does not explain why anomaly rankings improve. I would like the authors to clarify: which datasets or features show improved coverage, whether these improvements correspond to better detection performance, and why the method remains effective when the near-Gaussian assumption is clearly violated.

Questions for authors

Please see the specific concerns and questions raised in the Weaknesses section.

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

The paper proposed a method to use kurtosis for deriving the necessary noise scale for computing an outlier score from a denoising score matching network. The main idea is that different features may have very different marginal distributions, and features with heavier or longer tails may require larger diffusion noise scales so that low-density regions can be better covered when computing anomaly scores. Using the base method as a standard DSM, the author scale each feature during training and test by these precomputed scales which is assumed to expand the reach of noising processes for feature dimensions having a long-tail distribution. The authors compute this kurtosis on a distribution created by rearranging the histogram bins to closely match a standard normal. Then derive the target scale by a 4th Hermite polynomial approximation to the kurtosis. Experiments show better performance than the baselines on the 57 ADBench datasets.

Critical strengths

The paper addresses an important issue in tabular anomaly detection. The heterogeneous feature distributions have entropies vary widely among them and which affect the outlier detection algorithm significantly.

The approach seems to benefit from precomputed scales derived via a kurtosis information with longer tail feature distributions need larger scale.

The experiments on the ADBench datasets seems to support the method. The idea may be useful in practice.

Critical weaknesses

I think the novelties is limited and the method sounds like a data preprocessing step plus a DSM step. The contribution is a heuristic for selecting noise scales rather than an advance in diffusion-based anomaly detection.

My major concern is the theoretical justification for using kurtosis computed from rearranged histogram bins. This ignores the original spatial locations of the bins, which may have useful information about spread for outlier detection.

The experiments does not demonstrate on the larger-scale, high-volume datasets motivated in the introduction such as large network traffic datasets.

Comparison to some recent foundation models outlier detection methods would help the impact of the method.

Questions for authors

Does using the kurtosis for standardising the data as a preprocessing step improve all methods?

What are the consequences of removing the spatial information when rearrange the histogram bins?

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

This paper looks at denoising score matching as an anomaly detector for tabular data and asks how the training noise should be chosen. The proposal (K-DSM) is to give each feature its own fixed noise scale: heavier-tailed marginals get more noise, and the scale is computed from the kurtosis of a rearranged (symmetrized) version of the feature's marginal histogram. There is a coverage analysis behind this (a rearrangement bound plus a first-order expansion around a Gaussian), though the paper is upfront that the theory motivates rather than guarantees the implemented rule. Scoring is the norm of the network output at the clean input, so inference costs one forward pass. On ADBench with normal-only training this clearly improves over shared-scale DSM and comes out on top of the compared baselines, including DenoiserAD, with somewhat lower AUC-ROC than KPCA. Matched controls show that assigning the scales to the right features matters, not just the total amount of noise. For contaminated training the authors add an EMA-teacher filter that drops a fixed fraction of each batch by score, which gives the best PR/F1 among the denoising methods, though GMM and MCD still win on ROC. The appendix is long and unusually honest, including failures on MVTec/VisA and a full compute accounting.

Critical strengths
  1. The experimental evaluation is comprehensive in scope, but the main text is highly condensed, requiring readers to consult the appendix for much of the key evidence.

  2. The method has a simple and intuitive motivation. It requires neither labels nor a validation set and performs only a single forward pass at test time.

Critical weaknesses
  1. K-DSM’s AUC-PR advantage over DenoiserAD is small and not statistically established, KPCA retains the best AUC-ROC, and the combined variant changes several components, preventing clear attribution. The strongest supported claim is therefore a consistent improvement over plain DSM with much cheaper inference. I suggest qualifying the abstract’s ``highest mean AUC-PR'' statement accordingly.

  2. The empirical scope also seems narrower than the framing suggests. Although gains persist without fallbacks, 25.5% of feature coordinates use heuristic rules rather than measured kurtosis, and the allocation advantage is coupled to noise-unit scoring. K-DSM also underperforms DSM on MVTec-AD and VisA. The paper should therefore clarify its dependence on the scoring rule and avoid implying that the ADBench gains necessarily transfer to small-sample dense embeddings.

Questions for authors
  1. Appendix K shows the allocation advantage only under the D-weighted scoring metric. If a practitioner trained K-DSM but scored with raw residuals, would you expect no gain over scalar noise? Put differently, is the scoring rule part of the contribution?

  2. Did you check sensitivity to the fixed histogram resolution, especially for small training sets where the MVTec/VisA failures happen? You mention capping the fraction of features that hit the upper noise clip as a possible safeguard; was that tried at all?

  3. The EMA filter always discards a fixed fraction of each batch regardless of the actual contamination rate, which under-filters on high-contamination datasets. How much of the performance loss there is due to the fixed cutoff?

  4. Ssince the calibration only looks at marginals while some anomalies are purely interactional, it would be interesting to label the ADBench datasets as marginal- vs. interaction-anomaly and see whether the gain over DSM tracks that distinction.

2 comments
Reviewer_cAEhOpenReview ↗

Response

Thank you for your reply. It partially solved my problem. I will maintain my score.

Thank you for reading our response and for your reply. We're glad that we were able to partially address your concerns.

About your comment on the density of the main text, we have an extra page for this version of the paper, so we were able to bring some of the experiments from the appendix into the main text and to make it less dense. In the new version:

  • The experiments are split into new subsections in the main results in (Section 4.2) and the ablations and analysis (Section 4.3).

  • The comparison of noise allocations is now a table in the main text (Table 1). Before, it was only given in a sentence.

  • A new paragraph and Table 2 show which features gain coverage with K-DSM and whether detection improves with it (brought from Appendix C.7).

  • The results with OUTFORMER and on the full-size datasets are now reported in Section 4.2.

  • We reworded parts of the paper and split longer paragraph to make the reading easier.


Could you let us know which points remain unresolved? Your feedback has been valuable, as it allowed us to figure out the issue with VisA and MVTec-AD. We would be happy to provide further clarification or modifications to the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion and final remarks

Dear Area Chair and reviewers,

Before the discussion period ends, we wanted to summarize where things stand, with our final thoughts.

We ran the experiments that the reviewers asked for and updated the paper based on the results (list of changes below). As far as we can tell, the main concern that remains is novelty, raised by Reviewer Gbxb. While we have written a thorough response to them, which we encourage you to read, we wanted to summarize and give a final defence of novelty and significance.

During the discussion, Reviewer Gbxb pointed out that choosing a noise scale for each feature is equivalent to rescaling the features (with our noise scaling rule derived from coverage theory) and running standard DSM. The equivalence is correct, and we added it to the paper. But we disagree that it weakens novelty. The fact that it can be applied as preprocessing is simply a property of DSM. It does not change any of the insights gained from coverage theory, the rule itself, or the improved performance (best mean AUC-PR on ADBench, on par with DenoiserAD, while using one forward pass instead of fifteen).

As for significance, prior works in anomaly detection only tried to solve coverage with multiple noise scales. To our knowledge, our paper is the first to explore coverage with a single noise scale, and a solution for it keeping single-pass inference. This is a novel insight that other/future denoising detectors could use to keep inference fast and improve performance (and beyond them, the same scales also improve kNN, LOF and KPCA when used to rescale the features). We derived a rule motivated by coverage theory (Theorem 1), and found that the scale of each feature should follow the shape of its marginal, not its spread or the locations of its modes, as rescaling by spread or without rearrangement does worse.

We also answered every weakness and question raised by Reviewers cAEh and R3Hg, with new experiments where needed. For Reviewer cAEh, we modified the claims of the abstract, stated that the gain depends on the scoring rule, and added the sensitivity analyses on histogram resolution and noise clipping. For Reviewer R3Hg, we showed that rearranged kurtosis does better than ordinary kurtosis, and that the gain is larger on the datasets where coverage improves the most.

Thank you for taking the time to evaluate our paper.

Changes to the paper

The revised paper contains the following additions. For clarity, we have marked all changes in blue in the revised paper.

  • The equivalence between scaling the noise and scaling the features (Appendix G.7), with DSM trained in the rescaled form and the rescaled features given to twelve other detectors (Section 4.3, Appendix K.5).
  • An ablation of ordinary against rearranged kurtosis (Section 4.3, Table 1, Appendix K.1).
  • A coverage analysis over all features and datasets (Section 4.3, Table 2, Appendix C.7).
  • Experiments at full size on http and donors, with 567k and 619k observations. K-DSM reaches 0.982 and 0.978 AUC-PR and scores each test set in under a second (Section 4.2, Appendix H.1).
  • A comparison with the foundation model OUTFORMER. K-DSM has a higher mean AUC-PR (0.627 against 0.589) and is about 250 times faster at inference (Section 4.2, Appendix N).
  • The sensitivity to the histogram resolution and to the maximum noise scale (Appendices E.4 and J.3).
  • An analysis of the gap with DSM on the MVTec-AD and VisA embeddings. The gap comes from the output layer of the network and disappears when this layer is rescaled (Appendix M).
  • The two references on kurtosis in outlier detection suggested by Reviewer Gbxb (Appendix F).

We also revised the abstract, and Section 4.3 now states that the gain depends on the anomaly score. With the additional page, the experiments are split into the main results (Section 4.2) and the ablations and analysis (Section 4.3), and several results have been moved from the appendix to the main text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.