acceptodds
Under review as a conference paper at ICLR 2027

Local–Global Alignment for One-Class Tabular Anomaly Detection

Abstract

In this work, we address anomaly detection in tabular data under a one-class classification setting, where only normal records are available for training. We introduce local-global Representation Alignment for Tabular Anomaly Detection (Lo-GRAD), a method that detects anomalies by measuring how poorly a record's local feature representations align with its global representation. Our method assumes that anomalous records present a weaker alignment between their local and global representations. A Transformer Encoder produces both types of representations, which are optimized by a contrastive objective that increases the dot-product alignment between local-global pairs from the same record and decreases it for pairs from different records. Two auxiliary objectives strengthen this alignment: masked-feature modeling for the local representations and a reconstruction objective for the global representation. At inference, records with elevated local-global misalignment receive higher anomaly scores. We find that multiple scoring functions can effectively measure this misalignment, suggesting that the learned representations capture an anomaly signal that is not tied to a specific scoring function. Lo-GRAD also naturally provides feature-level anomaly attributions, since each local-global pair score is associated with a single feature. Across 25 ADBench datasets and without dataset-specific tuning, Lo-GRAD achieves the best macro-averaged AUROC and AUPRC compared with classical and state-of-the-art baselines, ranking first on 14 datasets for each metric.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-supervised learning, anomaly detection, local-global alignment, representation learning, feature attribution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CU1QOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes a new Anomaly Detection method for tabular data. The proposed approach combines two main objectives:

  • a contrastive objective,
  • a mask reconstruction objective.

The proposed approach involves generating a dual representation from a single sample, first local features representations that assign a token representation to each feature of the sample, and a global representation of the feature obtained from a CLS token added to the sequence. Mask reconstruction is performed in two ways:

  • masked features are reconstructed from the global feature representation.
  • masked features are reconstructed from their corresponding feature representation. The contrastive objective consists in minimizing average distance between the local tokens and the global token for a sample, while maximizing the same distance for negative pairs.

The proposed method is tested in 25 datasets from ADBench a widely used benchmark for AD on tabular data and shows strong performance.

Critical strengths

Q1: Paper is well organized and well written.

Q2: Method is clearly explained.

Q3: Ablations are thorough and motivate well each component of the method.

Q4: Experiments are ran several times to provide robust metrics.

Critical weaknesses

W1: Experimental setup differs from what is usually used in the literature.

W2: Apart from performance optimization, the paper lacks intuitions/analysis/explanation as to why combining the objectives is relevant.

W3: Without any clear rational only a subset of the benchmark was used to evaluate the method.

W4: A wide range of methods tested in the benchmark's paper are not included in the paper.

Questions for authors

Q1: Line 211 authors mention how they use per-feature reconstruction heads for the global row reconstruction. This implies a linear mapping for each feature. How does this scale on datasets that have a larger number of features? This seems like a lot of parameters to be trained.

Q2: The method disentangles local and global feature representations. Opposite to Q1, have the authors investigated how their model performs on datasets that have very few features (e.g. http dataset from ADBench)? Intuitively the global representation would contain very little additional information than the 3 existing tokens.

Q3: Why were only 25 datasets from ADBench selected ?

Q4:The vanilla k-nn AD approach is a strong baseline that was omitted from the benchmark. It would be very interesting to include it and analyze how it compares to the proposed method. This method is particularly interesting as it is very cheap to compute.

Q5: The training setup that includes a validation and test set is different from the ones usually used in papers. What is the rational? This prevents from including previously ran model on the same benchmark. Also, this is not really aligned with the AD objective that assumes that one does not dispose of anomalies at training.

Q6: it would be interesting to analyze the global token representation space within normal distribution and between normal and anomaly distribution. How much does this representation differ from a simple average of the feature tokens (for the unmasked case used in inference)?

Overall the proposed approach is interesting and well explained but the main rational for the proposed method is that it performs well on a subset of datasets from a benchmark while omitting some strong baselines previously included. (e.g. Knn method). We encourage the authors to extend their benchmark and better analyze the obtained representation from their method to provide a better understanding of why their method might perform well. We recommend rejection but we are open to discussion.

Small typo: Line 216, authors mention xjCLSx_j^{CLS} while it is not mentioned in the previous equation.

2 comments

Addressing weaknesses

  • W1: In our evaluation, we adopt a 70/10/20 split of normal records, no anomaly labels are used during training or validation stages on any experiment. Every baseline in our study was evaluated under the exact same conditions. We have added an explicit discussion of this protocol in the revised manuscript (response to Q5).
  • W2: We added a discussion explaining why combining the objectives is relevant for AD in Section 3.4.
  • W3: We evaluated on 25 ADBench datasets (26 now) dataset diversity against the computational cost of training all methods across five random seeds. We have documented this rationale in the revised manuscript (see Q3).
  • W4: Following your suggestion in Q4, we have now added KNN to the other methods included in ADBench (Autoencoder, Isolation Forest, OCSVM, ECOD).

Q1 - Reconstruction heads scaling

Thank you for raising this point. Each feature has a reconstruction mapping, so the parameter count grows linearly with the number of features.

We have clarified this in Sec. 3.3 where we introduce the implication on parameter growth and reference a section addressing this overhead in Appendix C of the revised manuscript. This section reports exact parameter counts and shows that the total reconstruction parameters account for less than 7% of the total parameters on the largest datasets.

Q2 – Performance with very few features

Thank you for this question. To address the suggested example, we evaluated the http dataset using the same model configuration and evaluation protocol of the main work (0.9997 AUROC). Our findins show that in datasets with few features there is no degradation in performance and it does not suffer from representational collapse. We addressed this in Section 4, and main evaluation tables of the revised manuscript.

Q3 - Selected ADBench datasets

The selected datasets balance dataset diversity with computational cotst (relates to your concern in Q1). Selection spans multiple domains, which range from 80 to 49,097 records and from 6 to 512 features (without http). We included diverse datasets so that the aggregate results reflect performance across varied characteristics. Related AD methods were also evaluated using a similar benchmark size (eg: MCM and Disent-AD use 20 datasets).

However, we acknowledge that the rationale for this selection was not explicitly stated in the original submission, and we have clarified it in Section 4 “Datasets” of the revised manuscript.

Q4 - knn

Thank you for this suggestion. We have evaluated KNN under the exact same one-class evaluation protocol across the selected ADBench datasets, and integrated the results into Section 4, Table 1 (macro averages), and Appendix A of the revised manuscript. In our experiments, KNN was found to be the strongest classical baseline. KNN achieves a macro AUROC of 0.8386 and a macro AUPRC of 0.7401, outperforming all other classical methods (Isolation Forest, OCSVM, and ECOD) as well as every deep baseline except MCM. However, Lo-GRAD substantially outperforms KNN across the benchmark, improving macro AUROC by 4.42 percentage points and macro AUPRC by 9.01 percentage points.

Q5 - Training setup

Thank you for raising this point. Our rationale for the 70/10/20 split of normal records is to separate model fitting, validation, and performance assessment, where both training and validation contain exclusively normal records. For Lo-GRAD and all deep baselines the train and validation set does not contain or use any anomaly labels. Although differences in evaluation protocols can limit direct comparison with the scores reported in other studies (eg: ADBench uses 50/50 split), the comparisons within our submission follow the same protocol and splits, without a per-dataset parameter tunning.

We have clarified that our method does not use anomalies during training or validation and the rationale for data splitting and evaluation protocol in Section 4 and Appendix F.1 of the revised manuscript.

Q6 - Global representation

Thank you for this suggestion. We also agree that this analysis addresses a gap in the original manuscript. To address this question, we compared CLS against the average of the feature tokens at inference on the three ablation datasets. The two representations exhibit similar inter-record distance distributions in Yeast, however, this distance increases substantially in the others.

We also evaluated whether the feature-token average can replace CLS as the global reference at inference. Using the mean Chebyshev scorer, we show that replacing CLS with mean pooling reduces macro AUROC by 11.43 percentage points and macro AUPRC by 13.99 percentage points. We have added this analysis to Section 4.2 and Appendix D.5 of the revised manuscript.

Typo in line 216 fixed, thank you.


Please let us know if these responses resolve your concerns, or if any additional experiments or clarifications would be helpful.

Reviewer_CU1QOpenReview ↗

Answer to Rebuttal

We thank the authors for trying to address the different points we raised in our review. In particular adding KNN, clarifying the training setup and investigating other representations than the CLS token (e.g. mean pooling).

For this reason, I'm happy to increase my score.

Reviewer_MTm1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Lo-GRAD, a one-class tabular anomaly detection method based on local-global representation alignment. A Transformer encoder produces feature-level and record-level representations, which are trained using a contrastive objective with masked-feature modeling and global reconstruction as auxiliary tasks. Anomaly scores are computed from discrepancies between local and global representations. Experiments on 25 ADBench datasets demonstrate competitive AUROC and AUPRC performance, along with feature-level anomaly attribution capabilities.

Critical strengths
  • Competitive empirical performance: Lo-GRAD achieves the highest average AUROC and AUPRC among the evaluated baselines across 25 datasets. The results suggest that the proposed framework is effective for one-class tabular anomaly detection.

  • Meaningful anomaly scoring framework: The use of feature-level and record-level representation discrepancies provides an intuitive approach to anomaly detection and enables feature-level attribution. The authors also provide ablation studies and attribution experiments to support the approach.

Critical weaknesses
  • Insufficient justification of the relationship between contrastive learning and anomaly scoring: The central contribution is local-global alignment, but the contrastive objective optimizes alignment in the projected space, while the best detection results are obtained using distances in the encoder space. In fact, the projected-space scores perform substantially worse. This raises concerns about whether the reported improvements are genuinely attributable to the proposed alignment objective.

  • Limited evidence for the specific contribution of local-global contrastive learning: The proposed contrastive pair construction follows established local-global representation learning approaches. Although the full model performs well, the ablation study examining the individual training objectives is limited to three datasets. More comprehensive evidence is needed to establish that the contrastive objective provides consistent benefits beyond masked-feature modeling and global reconstruction.

Questions for authors
  • Contribution of contrastive alignment: Can the authors provide an ablation study across the main benchmark datasets comparing the full model with an otherwise identical model trained without the local-global contrastive objective? This would help clarify its specific contribution to anomaly detection performance.

  • Discrepancy between representation spaces: Why does the projected space, where alignment is explicitly optimized, yield substantially worse anomaly detection performance than the encoder space? Can the authors provide further evidence explaining how the learned alignment contributes to the encoder-space anomaly scores?

  • Negative pair construction (clarification): Since different records are always treated as negative pairs, have the authors investigated potential false negatives arising from semantically similar normal records or low-cardinality categorical features? Although contextualized representations may mitigate this issue, it would be useful to understand its empirical impact.

1 comment

Addressing weaknesses

  • W1 - justification of training and AD: we discuss that the training objective is not necessarily the optimal AD signal, in line with prior AD literature (Sohn et al., 2021). Because the lower-dimensional projection space is optimized solely by the contrastive signal, it might modify representation’s geometry and discard relevant details for AD with the selected scorers. We expanded this discussion in the manuscript (question 2).
  • W2 - contribution of alignment: Following your suggestion, we expanded the ablation study comparing the full model against an identical model trained solely with the auxiliary objectives across the evaluation benchmark over five seeds. We show that local-global alignment improves macro AUROC and AUPRC by 6.33 and 13.14 percentage points respecively (question 1).

1st question - Contribution of contrastive alignment

We thank the reviewer for this suggestion. We evaluated a model trained only with the auxiliary objectives across the evaluation datasets over five seeds. The Transformer architecture, splits, preprocessing, hyperparameters, and scoring remains identical to Lo-GRAD.

The experiments over the selected ADBench datasets show that including local-global alignment improves macro AUROC by 6.33 percentage points and macro AUPRC by 13.14 percentage points. At the dataset level, Lo-GRAD outperforms the auxiliary-only variant on 22 of 26 datasets in AUROC and on 23 of 26 datasets in AUPRC. These results confirm that optimizing local-global alignment improves the anomaly signal. We have added this analysis to Section 4.2 “Contribution of alignment” and the complete breakdown in Appendix F.5 of the revised manuscript.

2nd question - Discrepancy between representation spaces

We thank the reviewer for raising this point. Because Reviewer bsRi also noted the distinction between the training objective and the inference scoring function and representations, we have revised Section 3.5, Section 4.3, and Section 5 of the revised manuscript to explain the relationship between the two spaces, and the empirical selection of the anomaly scorer, addressing both comments. In the submitted manuscript in sec. 5 we also mention the following regarding the relationship between the representation learning and anomaly scoring: “the dot-product score derived from training is one of the weakest scorers, showing that its role during training does not translate into superior performance. This finding reinforces the observation that the function used for representation learning is not necessarily the best choice for AD (Gong et al., 2019; Sohn et al., 2021)”. Because the projected space is a lower-dimensional projection of the encoder space optimized solely by the contrastive training signal, the projection mapping may discard some usefgul information for AD. Reviewer bsRi also indicated that “The poorer performance of distance-based scoring in the projected space may therefore also be related to the scale or geometry of the representations”, which is a plausible hypothesis of why the projected space achieves worse results.

3rd question - Negative pair construction

We thank the reviewer for this comment. To evaluate the impact of false-negative pairs, we conducted a controlled sensitivity analysis on the Cardio dataset in Appendix F.5 (referenced in Section 4.2 “Negative pair construction”). In this experiment, for each contrastive alignment calculation, we replaced a set of original normal training records of the batch (‘C’ in the manuscript formulas) with sets of the source record (‘a’ in the manuscript formulas) with 10, 20, 30, 40 and 50 slightly perturbed copies. We used zero-mean Gaussian noise (sigma = 0.01) for the perturbations. We use the configuration for Lo-GRAD for the manuscript across five seeds.

Additional Copies (k)dot_mean AUROCdot_mean AUPRCchebyshev_mean AUROCchebyshev_mean AUPRC
00.9450±0.01110.9450 \pm 0.01110.9398±0.01310.9398 \pm 0.01310.9443±0.01550.9443 \pm 0.01550.9415±0.01900.9415 \pm 0.0190
100.9472±0.01720.9472 \pm 0.01720.9357±0.01800.9357 \pm 0.01800.9434±0.01710.9434 \pm 0.01710.9391±0.01290.9391 \pm 0.0129
200.9388±0.01110.9388 \pm 0.01110.9332±0.01240.9332 \pm 0.01240.9428±0.01690.9428 \pm 0.01690.9406±0.01600.9406 \pm 0.0160
300.9302±0.02300.9302 \pm 0.02300.9264±0.02400.9264 \pm 0.02400.9354±0.01870.9354 \pm 0.01870.9324±0.01660.9324 \pm 0.0166
400.9383±0.02350.9383 \pm 0.02350.9305±0.01110.9305 \pm 0.01110.9296±0.02520.9296 \pm 0.02520.9215±0.02630.9215 \pm 0.0263
500.9324±0.02080.9324 \pm 0.02080.9298±0.02120.9298 \pm 0.02120.9303±0.02180.9303 \pm 0.02180.9220±0.02740.9220 \pm 0.0274

As shown in the table above, Lo-GRAD AD performance remains robust to a high presence of near-duplicate negative pairs. At moderate duplication levels, AUROC and AUPRC remain mostly unchanged. However, high duplication shows a modest decline.


Please let us know if these responses resolve your concerns, or if any additional experiments or clarifications would be helpful.

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

This paper proposes Lo-GRAD for tabular anomaly detection using only normal samples for training. The method uses a Transformer to generate both record-level global representations and feature-level local representations. Training combines local–global contrastive learning, masked-feature modeling, and global reconstruction, while anomaly scores are computed from discrepancies between the local and global representations. The authors evaluate the method on 25 ADBench datasets and report better average performance than the compared baselines.

Critical strengths

The method is clearly designed and has practical value. Detecting anomalies through the compatibility between feature-level and record-level representations is a reasonable and intuitive approach. The paper is well written and generally clear.

Critical weaknesses

1.Section 4.3 presents the performance difference between the encoder and projected spaces as an important finding. However, Section 4.2 explicitly states that evaluations in the projected space use different loss weights, batch sizes, projection dimensions, and masking rates. Therefore, my understanding is that the comparison in Table 2 changes not only the representation space but also the model’s training configuration. For Table 4, the main text does not clearly explain whether the results for the two spaces come from the same trained model. I am therefore somewhat uncertain about how much of the observed difference comes from the projection itself and how much comes from the different training configurations. Furthermore, the training objective optimizes dot-product matching in the projected space, whereas the main anomaly score uses Chebyshev distance in the encoder space before projection. These two measures are not directly equivalent. The poorer performance of distance-based scoring in the projected space may therefore also be related to the scale or geometry of the representations. 2.Section 4.2 describes the advantages of the Transformer over an FFNN. I would like to see experimental evidence supporting this choice of encoder.

Questions for authors

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

2 comments

1st question – configurations of different spaces

Thank you for identifying this ambiguity. Your understanding is correct, Tables 2 and 4 report results from two separately trained configurations, selected for scoring in RD\mathbb{R}^D and Rp\mathbb{R}^p. Therefore, the two spaces were not evaluated using the same trained checkpoints.

We have clarified the distinction between the model configurations for each space in Section 4.2 “Hyperparameters”, updated the captions of all tables evaluating both spaces (Tables 2, 4, 12-37, and 40-46), clarified scoring function selection in Appendix F.2, and detailed the underlying rationale in Appendix F.1 of the revised manuscript. Additionally, Appendix F.1 now includes cross-configuration evaluations comparing both spaces under each model configuration. We show that the encoder space optimization configuration favors encoder space scoring, while the projection space configuration achieves higher performance in the projected space. This result confirms that AD optimization in just one space might not benefit both spaces.

Optimized SpaceDatasetEncoder space (RD) AUROCEncoder space (RD) AUPRCProjected space (Rp) AUROCProjected space (Rp) AUPRC
RD\mathbb{R}^DYeast0.53410.86090.51490.7357
RD\mathbb{R}^DFault0.71790.92310.42570.7065
RD\mathbb{R}^DCardio0.94500.93980.58460.4500
RD\mathbb{R}^DMacro0.73230.90790.50840.6307
Rp\mathbb{R}^pYeast0.42070.70060.60590.8808
Rp\mathbb{R}^pFault0.58670.76060.65050.8971
Rp\mathbb{R}^pCardio0.92930.80000.86240.8675
Rp\mathbb{R}^pMacro0.64560.75370.70630.8818

2nd question – training objective vs anomaly scoring

We agree projected dot-product and Chebyshev distance are not equivalent measures. In Section 4 we mention that mean Chebyshev was selected because it achieved the best results on the development datasets by a tight margin, since the other functions also achieved very similar results. We do not claim that the training objective directly minimizes this distance. In section 5 we also mention the following regarding the relationship between the representation learning and anomaly scoring function: “the dot-product score derived from training is one of the weakest scorers, showing that its role during training does not translate into superior performance. This finding reinforces the observation that the function used for representation learning is not necessarily the best choice for AD (Gong et al., 2019; Sohn et al., 2021)”, indicating that the behavior observed in our experiments is consistent with previous work. The selection of the anomaly scoring function is completely empirical and does not imply that optimizing dot-product produces small Chebyshev distances.

However, we acknowledge that the distinction between the training objective and the inference-time scoring function, as well as how the projection head alters the scale and geometry of the representations, could be stated more explicitly in the manuscript (a point also noted by Reviewer MTm1). Accordingly, we have revised Sections 3.5, 4.3, and 5 to explicitly distinguish the projected dot-product objective used during training from the empirical selection of mean Chebyshev distance for inference, and to clarify how geometric differences between the encoder and projected spaces affect the evaluated scorers.

3rd question - Transformer vs FFNN

Thank you for raising this point. The submitted manuscript already included an experimental comparison in Appendix F.4, “Encoder Ablation”. In this experiment, we replaced the Transformer with an FFNN while keeping the masking procedure, auxiliary objectives, training protocol, and anomaly-scoring function unchanged.

Although we acknowledge that these experiments were not clearly explained in Section 4.2 “Attention”, and we updated that section (renamed to Section 4.2 “Encoder Ablation”) to explicitly cite results from Appendix F.4 of the revised manuscript.


Please let us know if these responses resolve your concerns, or if any additional experiments or clarifications would be helpful.

Thank you for the response.After considering it,I’ll keep my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.