acceptodds
Under review as a conference paper at ICLR 2027

RGLD: Randomized Global-Local Density Estimation for Tabular Anomaly Detection

Abstract

Unsupervised tabular anomaly detection requires methods that are accurate, robust across heterogeneous datasets, and computationally efficient. Classical statistical detectors are often efficient, but they usually rely on a fixed data view and a single notion of abnormality. Deep anomaly detectors can learn more flexible scoring functions, but they are substantially slower and difficult to tune in unsupervised settings due to the lack of a reliable supervisory signal. We propose RGLD, a randomized global-local density estimator for efficient unsupervised tabular anomaly detection. RGLD combines a global random-feature density branch, which identifies samples in broadly low-density regions, with a local neighbor branch, which detects samples that are weakly supported by nearby observations. Both branches operate over feature-bagged randomized views, allowing RGLD to expose anomaly evidence that may be hidden in any single representation. We evaluate RGLD on 47 tabular datasets against 23 statistical and deep anomaly detection baselines in the fully unsupervised setting. RGLD achieves competitive detection accuracy, with 77.50 AUROC and 38.59 AUPRC, while obtaining the most AUROC dataset wins. It runs 4.9x faster than ADERH and 50x–580x faster than the evaluated deep detectors, yielding a strong accuracy–efficiency tradeoff.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Unsupervised Learning, Anomaly Detection, Density Estimation, Randomized Algorithms

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kDj1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RGLD, a fully unsupervised method for tabular anomaly detection that combines global density support and local neighborhood isolation. The global branch approximates multi-scale kernel density scores using random Fourier features and mean embeddings, while the local branch computes nearest-neighbor distances against sampled reference sets. Both branches operate on feature-bagged randomized views, and their outputs are combined through hierarchical rank aggregation.

Critical strengths
  • A practically useful accuracy–efficiency tradeoff with identifiable computational savings.
  • Substantial empirical validation, including a useful controlled comparison.
  • The paepr is well-written and easy to follow.
Critical weaknesses
  • The comparative evaluation leaves preprocessing and computational-budget effects insufficiently separated from the core method. Appendix G describes an adaptive, label-free preprocessing procedure that selects among standard scaling, robust scaling, and quantile transformation. Section 6 states that baselines use official default configurations, but does not clearly establish whether the strongest baselines receive the same preprocessing.

  • RGLD seems to be a ensemble of KDE and KNN. In this case, why not compare it with other ensemble-based methods?

  • Data split is not clearly stated. The experimental setting of this paper seems to be unsupervised anoamly detection or outlier detection, where data split is not needed. Thus, what does data split in this paper mean? Morevoer, it seems most recent baselines in this paper are originally used for semi-supervised anoamly detection where a normal training set is needed. The authors should compare RGLD with other recent OD methods like [1] and [2].

  • Insufficient justification for the global–local distinction. The paper does not sufficiently justify its distinction between “global density” and “local neighborhood support”. The locality of KDE is controlled by its bandwidth, and nearest-neighbor density estimation is closely related to adaptive-bandwidth kernel estimation.

[1] Kim D, Hwang J, Lee J, et al. ODIM: Outlier detection via likelihood of under-fitted generative models[J]. arXiv preprint arXiv:2301.04257, 2023.

[2] Zhao Y, Chen Q, Tan Y, et al. CORE: In-Context Reconstruction for Unified Tabular Anomaly Detection[J]. arXiv preprint arXiv:2607.27615, 2026.

Questions for authors

See Weaknesses.

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

This paper proposes RGLD, an efficient unsupervised tabular anomaly detection method combining global density estimation and local neighbor-based isolation. It employs random Fourier features, feature bagging, random projections, and rank-based aggregation to capture anomalies across multiple views and scales. Experiments on 47 ADBench datasets against 23 baselines demonstrate competitive detection performance with substantially lower computational cost than several strong baselines.

Critical strengths
  • Competitive accuracy–efficiency trade-off: RGLD achieves competitive detection performance with substantially lower runtime than ADERH and deep learning-based methods. This is particularly valuable for practical unsupervised anomaly detection.

  • Extensive empirical evaluation: The paper evaluates the method on 47 datasets and provides ablation studies and theoretical analyses to support its design choices.

Critical weaknesses
  • Insufficient justification of key design choices: Although the individual components are reasonable, the paper does not sufficiently establish why their specific combination is appropriate. In particular, the contributions of random projection, multi-scale estimation, and rank-based aggregation are not adequately isolated or compared with alternatives. This makes it difficult to assess the necessity of the proposed design.

  • Limited validation against recent methods and theoretical contributions: Despite including 23 baselines, the evaluation covers few recent competitive approaches, particularly those published in 2025–2026. In addition, the theoretical results largely follow established principles or standard approximation bounds, without providing new guarantees directly related to the proposed anomaly detection algorithm. These limitations weaken the evidence for the paper's overall contribution.

Questions for authors
  • Design justification: Can the authors provide additional ablation studies isolating the effects of random projection and rank-based aggregation, including comparisons with reasonable alternatives? Such results would help establish whether these design choices are essential.

  • Recent baselines: Can the authors compare RGLD with recent methods such as Deep Noise Evaluation (AAAI 2025) under a consistent fully unsupervised evaluation protocol? This would help clarify its competitiveness against more recent approaches.

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

This paper proposes RGLD, an efficient method for fully unsupervised tabular anomaly detection. RGLD combines a global branch that estimates density support through random-feature kernel approximation with a local branch that measures neighborhood support through sampled-reference nearest neighbors. The two branches operate over independently sampled feature-bagged random projections, and their scores are merged through rank-based aggregation. Evaluated on 47 ADBench datasets against 23 statistical and deep baselines, RGLD achieves competitive detection accuracy with the most AUROC dataset wins. Although ADERH attains slightly higher mean AUROC and AUPRC, the paired difference is not statistically significant. Meanwhile, RGLD runs approximately 4.9$\times$ faster than ADERH and 50$\times$ to 580$\times$ faster than the evaluated deep detectors.

Critical strengths
  1. The paper conducts extensive experiments on 47 tabular datasets from ADBench against 23 statistical and deep learning baselines.
  2. RGLD shows a substantial advantage in total runtime over the deep learning baselines and the strongest statistical baseline ADERH.
Critical weaknesses
  1. The comparison with recent methods is insufficient. Although the paper includes deep learning based baselines, the most recent one was published in 2023 and many recent methods are missing, such as MCM [1], DRL [2], and diffusion-based approaches [3]. Moreover, a large portion of the compared methods are classical baselines. Since the central claim of the paper is a favorable accuracy and efficiency tradeoff, the absence of recent deep learning methods weakens the reliability of this claim.

  2. The runtime claim appears overstated. The introduction motivates efficiency with time-sensitive applications. Under this motivation, I believe that inference time excluding training should be compared. However, the reported runtime appears to be the sum of training and inference time, and inference time is not reported separately. Hence it is unclear whether RGLD retains its advantage at inference time. In particular, deep learning methods may be considerably faster at inference once trained, so the reported speedup may shrink substantially under an inference only comparison.

  3. The data split protocol requires a detailed explanation. The paper states that it follows the ADBench protocol but does not specify the split. It is unclear whether the paper adopts the protocol commonly used in tabular anomaly detection, where 50% of the normal samples are used for training and the remaining data are used for testing (semi-supervised setting). If the ADBench protocol was followed, the training data presumably contain anomalies, which implies a contaminated setting. In this case, I would like to see whether RGLD remains competitive under the 50% normal training protocol. In addition, several deep learning baselines are reported with considerably low performance (see Table 6 of the original DRL paper for comparison), and this discrepancy requires an explanation. If the baselines are underestimated due to the protocol difference, the baseline results should be re-evaluated. Alternatively, the paper could explicitly frame the contribution under contaminated training data and support this framing with controlled experiments on the contamination ratio.

  4. The bound in Proposition 2 may not be meaningful in practice. The implementation uses D=10D=10 and with δ=0.05\delta=0.05 the bound 8log⁡(2/δ)/D\sqrt{8\log(2/\delta)/D} evaluates to approximately 1.72. This exceeds the range of the target quantity, which is an average of Gaussian kernel values and thus lies in (0,1](0,1], so the bound does not certify the approximation quality actually used in the experiments.

Due to the weaknesses above, I believe the paper cannot be accepted with major revisions in its current form, and therefore I recommend a clear reject.

References [1] Yin, Jiaxin, et al. "Mcm: Masked cell modeling for anomaly detection in tabular data." International Conference on Learning Representations. Vol. 2024. 2024. [2] Ye, Hangting, et al. "DRL: Decomposed representation learning for tabular anomaly detection." International Conference on Learning Representations. Vol. 2025. 2025. [3] Livernoche, Victor, et al. "On diffusion modeling for anomaly detection." International Conference on Learning Representations. Vol. 2024. 2024.

Questions for authors

Please see the Weaknesses.

1 comment

As there was no further author discussion, I will maintain my original rating based on the submitted response and review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.