acceptodds
Withdrawn from ICLR 2027

Robust Filtering and Correlation-Aligned Diffusion for Unsupervised Tabular Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised anomaly detection trains on unlabeled data that contain anomalies. We address this contamination with two data-centric steps. First, a FastMCD-based pre-training filter removes rows outside the robust covariance support; it improves most detectors we compare. Second, we replace only the isotropic forward corruption of Diffusion Time Estimation (DTE), a diffusion-based detector, with correlation-aligned noise. Estimating this correlation from the retained rows, anomalies included, beats every cleaner set of rows we tested. Anisotropy-aware DTE (ADTE) combines these two steps and, in its categorical variant ADTE-C, a regression head that stabilizes results across seeds. On the 57 ADBench groups under the unsupervised protocol, ADTE-C reaches 77.0 AUCROC, to our knowledge the highest reported. Removing any one of the three components significantly lowers ADTE-C's mean AUCROC.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
unsupervised anomaly detection, outlier detection, tabular data, robust filtering, diffusion time estimation, contaminated training data, robust covariance estimation, anisotropic noise

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_MMw5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Anisotropy-aware Diffusion Time Estimation (ADTE) for unsupervised tabular anomaly detection. It adds to DTE a covariance-based training filter and an auxiliary timestep regression head, as well as a new diffusion noise. The paper shows empirically, using the 57 ADBench datasets, that the filtering is useful to many baseline detectors and that the modified diffusion noise provides some improvements. The paper also includes several ablations and additional results in the appendix.

Critical strengths
  • The modified diffusion noise to DTE is simple and motivated. Shaping corruption using feature correlations is a natural extension of DTE.
  • The empirical evaluation is extensive. The paper compares with several other baselines, with multiple detector families, and evaluates the filtering over those methods as well. There are also ablations for each component, parameter sensitivities, and paired comparisons across datasets.
  • The disclosed limitations are honest and useful. The author clearly describes the unusual categorical training objective, the weaker embedding results, and the extensive AI assistance.
Critical weaknesses

As disclosed by the authors (and is quite obvious when reading the paper), the paper uses generative AI extensively, which I believe explains my following assessment:

  • The supplemental material (codebase) shows a large set of files with many different experiment attempts, as well as agent instructions. One of these states: “Each analysis DEPLOYS only if its own p < 0.05; otherwise it is NULL and stays internal.” While some of these “null” results are presented, it is unclear how many unfavourable outcomes were removed. The scripts also motivate new experiments using previous successes and failures. One justifies restricting an experiment to tabular datasets by noting that the broader comparison was not statistically significant. With that many follow-up hypotheses and experiments on the same benchmark, this raises large concerns over whether the paper presents meaningful results for the problem presented or is simply maximizing results on a benchmark.
  • The proposed method adds several components to DTE, but does not show a significant improvement over simple baselines such as T² and Euclidean kNN. The improvement from the additional regression head is also small and weakly supported across settings. Given the added complexity, it is unclear whether the method provides a meaningful improvement for the problem presented.
  • The paper is very hard to read and often feels like a long experiment log. The main idea is simple, but it gets buried under experiments, subgroup comparisons, and appendix references, and there isn't even a single figure. It is difficult to understand which findings actually support the main claims and which experiments were added to explain earlier results. I would suggest that the authors take over the writing and do a thorough human rewrite: remove the clutter, decide which findings are central, and assemble the results into a clear paper.
Questions for authors
  • How do your reconcilie the small effect of the auxiliary head small effect on performance with its claimed importance?
  • Given the intervals against T² and Euclidean kNN, what evidence supports a practical advantage of the learned model over these alternatives?
1 comment

Thank you for the review.

Package: the quoted line shows you read the supplementary material closely; the version you examined was our full working tree, experiments that did not reach the paper included. Whether results were maximized on the benchmark is a question any paper has to answer with its own numbers, and we do so below: the components and the settings. The package now contains the code and recorded results that regenerate every table of the paper.

Components: each has one purpose, and each is achieved. The filter addresses contamination: removing it from ADTE-C costs 2.79 AUCROC points [95% CI 0.97, 4.87] on the 57 groups. Aligned noise shapes the corruption by the data's own correlation, retained anomalies included (the central novelty, see Baselines): aligned minus isotropic noise, everything else identical, +0.73 [+0.14, +1.44] on the 57 groups and +0.92 [+0.23, +1.73] on the 47 tabular groups. The regression head stabilizes results across seeds (the abstract's words). Stability, not size, is what the paper claims for the head, so a small mean effect is expected. The measured one, +0.33 AUCROC points [+0.03, +0.68] on the 57 groups, is significant in the paper's statistical sense of an interval that excludes zero. The stability shows in the seeds: with the filter and aligned noise, the head narrows a group's AUCROC spread over its five training seeds by 21% and raises each group's lowest seed outcome by an average of +0.63 points [+0.20, +1.22], while the highest seed outcome does not change significantly. Without the filter and under isotropic noise, the head narrows the spread by 34% and raises the lowest outcomes by +1.91 [+0.68, +3.22] (Table 21). A narrower seed spread also makes differences in results easier to measure.

Parameters: DTE's network, diffusion schedule and optimizer are as released; what we set, once for every dataset, is K = 8 bins (DTE's default, 7, is the value its authors tuned over all the datasets), head weight λ = 0.1, score weight ω = 0.5, full alignment (s = 0), and for the filter a support fraction of 0.75 with a 32-dimensional PCA projection only when a dataset has more features. Nearby settings change little, all on the 57 groups: ω anywhere from 0 to 1 changes the mean AUCROC by at most 0.22 points. The effect of a support fraction of 0.70 or a PCA dimension of 16 or 64 is not distinguishable from zero. K from 6 to 10 moves the mean by at most 0.1 point, and the chosen bin count does not even maximize it: K = 10 gives 77.09 against K = 8's 77.02 (Table 5).

Quoted instruction: the scripts the review cites, the quoted line among them, belong to the analyses that explain the results after the detector was fixed and its 77.0 AUCROC obtained; they did not choose the detector or its settings, and the comparison of detectors in Table 1 reports its numbers whatever their significance. The quoted line applies the paper's 5% significance level to those analyses: a result is stated as a finding only when it is significant at that level, and non-significant contrasts are printed with their intervals too, as you note. The split into 47 tabular and 10 embedding groups is the benchmark's own, by data type. On the 10 embedding groups, the paper reports that DTE itself is behind the classical detectors before any alignment (Section 4.6). That is why the analyses of alignment focus on the 47 tabular groups. A variant of the aligned noise recovers part of that gap on the embedding groups, raising ADTE-C there by +0.80 AUCROC points [0.23, 1.45] (Section 4.6). The detector comparison and the alignment effect are reported on both the 57 groups and the 47.

Baselines: the paper claims no significant margin over T² or Euclidean kNN; it claims the highest mean AUCROC among the detectors compared after the same filter, and a significant alignment gain. With the paper's filter, those two detectors are not ordinary baselines. T² (76.8) and Euclidean kNN (76.5) then have a higher mean AUCROC than every other detector in Table 1, including the learned ones and the robust-covariance detector built on the filter's own estimator (75.2 at its best), and only ADTE-C (77.0) scores higher. The learned model's advantage is the central novelty: anomalies help shape the training noise, even better than the normal samples, where T² and Mahalanobis kNN need them removed from their correlation estimate; the evidence is in our reply to Reviewer ZA9R (Novelty). That understanding is what can be leveraged into a significant improvement over the filtered T² on the 47 tabular groups, at lower computation than ADTE-C. The components add no training cost: averaged over the member datasets, the full pipeline trains in 215 seconds against 219 for DTE-C without them, plus about 2.5 seconds per dataset for the filter on a CPU (Section 4.3). We will run any detector you consider a stronger comparator, since we found none.

Reviewer_WAi6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work presents a strategy to extend the anomaly detection capability of DTE (ICLR 2024 [1]) to tabular data. However, the proposed method appears to be more of an engineering-oriented extension or technical report than a conceptually novel research contribution. Specifically, the method first applies PCA to identify and filter out potential anomalies in the tabular data, and then recomputes the covariance matrix used by DTE based on the filtered samples. Finally, to better adapt the model to the filtered data distribution and the updated covariance matrix, the authors introduce a lightweight learnable head to further fine-tune the feature representations.

[1] On diffusion modeling for anomaly detection. ICLR 2024.

Critical strengths

The primary strength of this work lies in its practical strategy for improving DTE in the context of tabular anomaly detection. In addition, the experimental evaluation is fairly comprehensive and demonstrates the effectiveness of the proposed approach.

Critical weaknesses

I discuss the weaknesses one by one below, in descending order of importance.

  1. The novelty of this work appears limited for two main reasons. First, PCA-based anomaly detection has already been extensively explored in recent work, such as SubspaceAD [1] in industrial anomaly detection (IAD). Although the proposed idea may be effective in the context of tabular anomaly detection (TAD), simply transferring a similar PCA-based strategy from IAD to TAD does not, in my view, constitute a sufficiently novel or compelling contribution. Second, the proposed framework relies heavily on DTE, with most of the modifications centered on practical adaptation strategies rather than substantial methodological innovation. As a result, the work reads more like an engineering-oriented extension or technical report built upon DTE than a well-developed research paper with a clearly distinct methodological contribution.

  2. The presentation of this work raises concerns about the extent to which the manuscript may have been heavily assisted by generative AI, although this is difficult to determine conclusively from the paper alone. Several aspects contribute to this impression. First, the paper lacks sufficient visualizations, such as an overall framework diagram and qualitative anomaly detection or localization results, which makes the presentation less informative and less convincing. Second, the writing style is unusually formulaic and, in several places, resembles AI-generated academic prose. For example, expressions such as “(Table 1)” in Line 52 and “We make three contributions” in Line 53 feel somewhat templated and disconnected from the surrounding narrative. Having served as a reviewer for more than six years, I have rarely encountered this particular combination of highly formulaic writing and limited qualitative presentation in manuscripts written in a more conventional academic style. While these observations do not constitute evidence of AI authorship by themselves, they do raise concerns about the originality and quality of the manuscript’s presentation.

[1] SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling. CVPR 2026.

Questions for authors

See weaknesses above.

1 comment

Thank you for the review. Starting with clarification about the summary:

Filter: the filter is not PCA-based. A row is removed when it lies outside the robust covariance support of the FastMCD estimator. PCA is only a projection used to fit that estimator when a dataset has more than 32 features; 33 of the 47 tabular groups are filtered with no PCA at all, the PCA dimension can be halved or doubled with no distinguishable effect on mean AUCROC, and the model trains in the original feature space. So nothing here transfers a PCA-based detector from industrial anomaly detection: PCA is not the detector, and every anomaly score is computed in the original feature space.

Head: the regression head is not a fine-tuning stage. It is trained jointly with the categorical head, in one run, and its role is stability: with the filter and aligned noise it narrows the spread of a group's AUCROC over the five training seeds by 21% and raises each group's lowest seed outcome by an average of +0.63 points [+0.20, +1.22] on the 57 groups, while the highest seed outcome does not change significantly.

Noise: no covariance of DTE is recomputed, because DTE's noise is isotropic. The training noise is based on the correlation of the retained rows, and this carries the gain: aligned minus isotropic noise, everything else identical, +0.92 AUCROC points [95% CI +0.23, +1.73] on the 47 tabular groups.

Contribution: it lies in this noise. The central novelty is that anomalies help shape the training noise, even better than the normal samples. The retained anomalies tilt the noise toward the directions they occupy, so the network learns how far such rows lie from the normal rows. The evidence is in our reply to Reviewer ZA9R (Novelty). DTE is the host because its corruption is a modelling choice that can be shaped: its score measures how far a row lies from the rows it was trained on.

Presentation: visualizations are ready. The use of generative AI is disclosed in the paper's AI use statement, and we take responsibility for the content and the presentation.

Reviewer_ZA9ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Anisotropy-aware Diffusion Time Estimation (ADTE) for unsupervised tabular anomaly detection. Building upon DTE, the method introduces three components: an MCD-based robust filter to reduce training-data contamination, correlation-aligned diffusion noise to capture feature dependencies, and an auxiliary regression head to improve prediction stability. Experiments on 57 ADBench groups demonstrate improved detection performance, achieving an average AUROC of 77.0. The paper also investigates how noise alignment affects different anomaly

Critical strengths

1.The paper is well structured, and the experiments and statistical analyses provide useful insights into the effects of robust filtering and correlation-aligned noise.

2.The proposed method achieves competitive performance across a broad range of anomaly detection benchmarks, with ablations supporting the contributions of its individual components.

Critical weaknesses

1.Limited methodological novelty: The approach mainly combines existing techniques, including FastMCD filtering, correlation-based noise covariance, and auxiliary regression. The algorithmic contributions beyond the original DTE framework appear relatively incremental.

  1. Limited generality of correlation alignment: The experiments show that correlation-aligned noise improves detection for anomalies along leading eigenvector directions but degrades performance for several other anomaly geometries. Since anomaly directions are unknown in practical unsupervised settings, the robustness of this fixed noise design remains unclear.
Questions for authors

How does ADTE generalize to out-of-sample anomaly detection under varying contamination ratios or distribution shifts?

1 comment

Thank you for the review.

Novelty: the central novelty is that anomalies help shape the training noise, even better than the normal samples. The retained anomalies tilt the corruption toward the directions they occupy, so the network is trained on displacements along those directions and learns how far such rows lie from the normal rows. The evidence is in Section 4.5: the correlation that shapes the noise is estimated from the rows the filter retains, anomalies included; estimating it from the labeled normal rows alone, the anomalies left out, costs ADTE-C 0.94 AUCROC points [0.26, 1.83] on the 47 tabular groups, while the same substitution improves the detectors that use the correlation as a distance metric, T² by 4.6 [2.1, 7.3] and Mahalanobis kNN by 1.6 [0.6, 2.7].

Generality: the noise shape is a bet, including isotropic noise; aligned noise is a bet shaped by the data, and its gain comes from matching each dataset's own correlation: rotating the noise correlation by 12°, 23°, 38° or 49° costs 0.5, 0.8, 1.2 and 1.2 AUCROC points on the 47 tabular groups (Table 8). The geometries on which alignment degrades detection are anomalies planted in a chosen direction on synthetic copies, to show which geometry each noise favors (Table 7). On the real data, alignment is a significant gain: aligned minus isotropic noise, everything else identical, +0.92 AUCROC points [95% CI +0.23, +1.73] on the 47 tabular groups (Table 9). The same geometry governs it: the gain is larger where more of the anomalies' deviation lies along the leading correlation directions. A threshold on that share, chosen on half the groups, separates the other half by 1.7 AUCROC points of alignment gain (one-sided p = .01; Section 4.4).

Generalization question: out of sample, with everything fitted on one half of each tabular dataset and the other half scored, ADTE-C with the retained rows' correlation still beats ADTE-C with the labeled normal rows' correlation by +0.67 AUCROC points [+0.10, +1.39] on the 47 groups (Appendix G). Contamination varies within the benchmark itself: every dataset is trained at its native anomaly share, which runs from 0.03% to 39.9% across the 57 groups and exceeds 20% in 12 of the 47 tabular groups. After the filter, 35 of the 47 tabular groups still have anomalies among their retained rows, at a median share of 4%, and the remaining 12 have none (Section 4.5). To answer this question we resampled each dataset's anomalies to 5% and to 15% of its training rows before filtering: alignment's gain is +0.79 AUCROC points [+0.07, +1.66] at 5% and +0.84 [−0.45, +2.32] at 15% on the 47 tabular groups, similar to its gain at the native share. One configuration, fixed once, serves every dataset; nothing is tuned per dataset, and nearby settings change little.

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

Related papers

Discussion (0)

Sign in to comment.