arxiv: 2602.06810 · v2 · submitted 2026-02-06 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Calibrating Tabular Anomaly Detection via Optimal Transport

Hangting Ye , He Zhao , Wei Fan , Xiaozhuang Song , Dandan Guo , Yi Chang , Hongyuan Zha

Authors on Pith no claims yet

Pith reviewed 2026-05-16 06:45 UTC · model grok-4.3

classification 💻 cs.LG

keywords tabular anomaly detectionoptimal transportpost-processing calibrationmodel-agnosticK-means centroidsdistribution disruptionanomaly scoring

0 comments

The pith

Optimal transport measures how a test sample disrupts normal data distributions from random samples and K-means centroids to calibrate any tabular anomaly detector.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a post-processing step that improves any existing tabular anomaly detector by assigning calibration scores based on optimal transport distance. Normal data is modeled through two distributions: one from random sampling and one from cluster centroids. Adding a test point and recomputing the transport cost between these distributions yields low disruption for normals and high disruption for anomalies. A proof shows the transport distance is bounded below by distance to the centroids, so anomalies receive systematically higher scores in expectation. This signal is added to the base detector output and works across density, classification, reconstruction, and isolation methods on 34 datasets.

Core claim

CTAD calibrates any tabular anomaly detector by computing the optimal transport distance between an empirical distribution formed by random normal samples and a structural distribution formed by K-means centroids, both before and after inserting the test sample. The difference quantifies disruption. The authors prove that this distance admits a lower bound proportional to the test sample's distance from the centroids and that anomalies produce higher expected calibration values than normal points, which explains generalization across heterogeneous tabular datasets.

What carries the argument

Optimal transport distance between the joint empirical distribution of random normal samples and the distribution of K-means centroids, used to quantify the incremental disruption introduced by a single test sample.

If this is right

Any existing detector from density estimation, classification, reconstruction, or isolation families receives a performance lift without retraining.
The improvement holds with statistical significance across 34 datasets that span different scales, feature types, and anomaly patterns.
Even deep-learning detectors that already achieve high accuracy on some sets show further gains.
The method requires no dataset-specific hyperparameter search beyond the base detector's own settings.
The calibration signal is additive and can be applied at inference time to existing anomaly score outputs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same disruption idea could be tested on graph or time-series data if suitable centroid constructions are defined for those domains.
Combining the OT calibration with existing ensemble techniques might produce further error reduction without additional model training.
The lower-bound proof on transport distance suggests that the calibration strength scales with how far anomalies lie from normal clusters, offering a way to predict per-dataset improvement magnitude in advance.
In high-stakes settings such as fraud or medical screening, the method could be used to rank flagged cases by their calibrated disruption value rather than raw scores alone.

Load-bearing premise

Normal points are sufficiently well captured by random sampling plus K-means centroids that anomalies always produce measurably larger optimal transport disruption than normal points under the heterogeneity found in tabular data.

What would settle it

On a new collection of tabular datasets, compute the average calibration score for confirmed normal points versus confirmed anomalies; if normals receive higher scores on average or if the method fails to improve at least 80 percent of the tested base detectors, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2602.06810 by Dandan Guo, Hangting Ye, He Zhao, Hongyuan Zha, Wei Fan, Xiaozhuang Song, Yi Chang.

**Figure 2.** Figure 2: Comparison of all models’ performance across different [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 4.** Figure 4: Hyperparameter ablation studies. Top: Effect of centroid [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Tabular anomaly detection (TAD) remains challenging due to the heterogeneity of tabular data: features lack natural relationships, vary widely in distribution and scale, and exhibit diverse types. Consequently, each TAD method makes implicit assumptions about anomaly patterns that work well on some datasets but fail on others, and no method consistently outperforms across diverse scenarios. We present CTAD (Calibrating Tabular Anomaly Detection), a model-agnostic post-processing framework that enhances any existing TAD detector through sample-specific calibration. Our approach characterizes normal data via two complementary distributions, i.e., an empirical distribution from random sampling and a structural distribution from K-means centroids, and measures how adding a test sample disrupts their compatibility using Optimal Transport (OT) distance. Normal samples maintain low disruption while anomalies cause high disruption, providing a calibration signal to amplify detection. We prove that OT distance has a lower bound proportional to the test sample's distance from centroids, and establish that anomalies systematically receive higher calibration scores than normals in expectation, explaining why the method generalizes across datasets. Extensive experiments on 34 diverse tabular datasets with 7 representative detectors spanning all major TAD categories (density estimation, classification, reconstruction, and isolation-based methods) demonstrate that CTAD consistently improves performance with statistical significance. Remarkably, CTAD enhances even state-of-the-art deep learning methods and shows robust performance across diverse hyperparameter settings, requiring no additional tuning for practical deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CTAD adds a practical post-hoc calibration layer to tabular anomaly detectors via OT distances on random samples and K-means centroids, with consistent gains across 34 datasets, though the lower-bound proof looks fragile under typical data heterogeneity.

read the letter

The core idea is straightforward: take any existing TAD detector, then adjust its scores by how much a test point increases the OT distance between a random empirical sample of the data and a K-means structural summary. Normal points keep the two distributions compatible; anomalies push them apart. The paper reports that this calibration lifts performance with statistical significance on 34 tabular datasets and across seven detector families, including deep models, without retraining or extra hyperparameter search for the user. That empirical breadth is the strongest part of the work and makes the method immediately usable in applied settings where tabular data is messy and no single detector dominates. The model-agnostic framing is also clean and avoids the usual trap of proposing yet another detector that only works on its own benchmarks. The main soft spot is the theory. The abstract states a lower bound on OT distance proportional to distance from centroids and claims this produces higher expected scores for anomalies, which is meant to explain the cross-dataset generalization. But the argument appears to rest on K-means centroids remaining stable and the metric making features commensurable. Tabular data routinely violates those conditions through mixed types, scale differences, and outliers that distort centroids, so the proportionality constant could shrink or the separation could fail in expectation. Without explicit regularity conditions in the derivation, the proof does not yet close the gap between the math and the hardest real-world cases. The experiments are broad enough to stand on their own, but the claimed explanation for why the method works everywhere needs tighter justification. This is the kind of paper a practitioner or applied researcher would want to read for the calibration recipe and the dataset coverage. A reader focused on post-processing or OT applications in anomaly detection would get concrete value from the results. It deserves peer review so the full derivation and experimental controls can be checked; the empirical case is already strong enough to justify the time.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces CTAD, a model-agnostic post-processing framework for tabular anomaly detection. Normal data is characterized via an empirical distribution from random sampling and a structural distribution from K-means centroids; optimal transport (OT) distance then quantifies the disruption induced by a test sample. The paper claims a proof that OT distance admits a lower bound proportional to the test sample's distance from the centroids, implying that anomalies receive strictly higher expected calibration scores than normals. Experiments on 34 tabular datasets across seven detector families (density estimation, classification, reconstruction, isolation) report consistent, statistically significant gains, including on deep-learning baselines, with claimed robustness to hyperparameter choice.

Significance. If the lower-bound result holds under the regularity conditions appropriate to heterogeneous tabular data, the work supplies a theoretically grounded, tuning-free calibration layer that can be applied to any existing TAD detector. The combination of an explicit expectation argument with large-scale empirical validation across detector categories would constitute a useful contribution to the tabular anomaly-detection literature.

major comments (2)

[Abstract] Abstract: the claimed proof that OT distance has a lower bound proportional to the test sample's distance from centroids does not state the necessary regularity conditions (bounded support, Lipschitz cost, or normalized metric). Without these, the proportionality constant can become arbitrarily small when K-means centroids are computed on unscaled, mixed-type tabular features, undermining the expectation separation between anomalies and normals.
[Abstract] Abstract and experimental claims: the argument that the two normal-data distributions (random empirical + K-means structural) remain compatible for normal points but become measurably incompatible for anomalies rests on the assumption that K-means centroids adequately capture the normal manifold. In high-dimensional heterogeneous tables this assumption is sensitive to feature scaling and the choice of K; the manuscript must verify that the lower-bound constant remains positive under the data regimes where TAD is hardest.

minor comments (2)

The abstract states that CTAD requires no additional tuning; the manuscript should explicitly list the two free parameters (K-means cluster count and number of random samples) and report the exact ranges used in the robustness experiments.
Ensure that all OT-related equations (cost function, transport plan, lower-bound derivation) are numbered and cross-referenced in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the theoretical statement and its applicability to heterogeneous tabular data. We will revise the manuscript to explicitly include the necessary regularity conditions for the lower-bound result and to provide targeted empirical verification of the constant's positivity in challenging regimes.

read point-by-point responses

Referee: [Abstract] Abstract: the claimed proof that OT distance has a lower bound proportional to the test sample's distance from centroids does not state the necessary regularity conditions (bounded support, Lipschitz cost, or normalized metric). Without these, the proportionality constant can become arbitrarily small when K-means centroids are computed on unscaled, mixed-type tabular features, undermining the expectation separation between anomalies and normals.

Authors: We agree that the regularity conditions must be stated explicitly. In the revised manuscript we will restate the theorem with the assumptions of bounded support, Lipschitz continuity of the cost function, and normalized Euclidean metric after standard feature scaling. Under these conditions the proportionality constant is strictly positive. Because CTAD is applied after the same preprocessing used by the base detectors, the separation between expected scores for anomalies and normals is preserved in the regimes considered in the paper. revision: yes
Referee: [Abstract] Abstract and experimental claims: the argument that the two normal-data distributions (random empirical + K-means structural) remain compatible for normal points but become measurably incompatible for anomalies rests on the assumption that K-means centroids adequately capture the normal manifold. In high-dimensional heterogeneous tables this assumption is sensitive to feature scaling and the choice of K; the manuscript must verify that the lower-bound constant remains positive under the data regimes where TAD is hardest.

Authors: The current experiments already cover 34 datasets that include high-dimensional heterogeneous tables and report statistically significant gains across all detector families. To directly address the request for verification, the revision will add a short subsection that computes empirical lower-bound constants on representative high-dimensional subsets and confirms they remain positive for the chosen K values. The existing hyper-parameter sensitivity study already shows that performance is stable for moderate changes in K, supporting that the manifold-capture assumption holds under the preprocessing used throughout the evaluation. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines the calibration score directly as the OT distance measuring disruption when a test sample is added to two label-free normal-data distributions (random empirical sample and K-means centroids). The claimed lower bound is derived from general OT metric properties and stated to be proportional to distance from centroids; the expectation that anomalies produce higher scores follows from this construction and the assumption that anomalies lie farther from normal centroids. No equations reduce the result to a fitted parameter, self-citation, or ansatz that presupposes the target outcome. The method remains model-agnostic post-processing without reference to anomaly labels, rendering the derivation self-contained against external OT theory.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard properties of optimal transport as a discrepancy measure and the domain assumption that K-means centroids capture structural normality; no new entities are introduced and free parameters are limited to implementation choices for the two distributions.

free parameters (2)

K-means cluster count
Determines the structural distribution representation; value is not stated in abstract and must be chosen or tuned.
number of random samples
Controls the empirical distribution; specific count not given in abstract.

axioms (2)

standard math Optimal transport distance quantifies distributional discrepancy in a manner that reflects sample disruption
Invoked to define the calibration signal from adding a test point.
domain assumption K-means centroids plus random samples together characterize normal tabular data sufficiently for anomaly contrast
Foundation for the two complementary distributions used in the method.

pith-pipeline@v0.9.0 · 5567 in / 1617 out tokens · 49292 ms · 2026-05-16T06:45:49.404461+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We prove that OT distance has a lower bound proportional to the test sample's distance from centroids... OT(P,Q) ≥ 1/(M+1) ∥x_test − C_j*∥
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

K-means centroids... structural distribution Q... empirical P via random sampling

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 1 internal anchor

[1]

Zeeshan Ahmad, Adnan Shahid Khan, Cheah Wai Shiang, Johari Abdullah, and Farhan Ahmad. 2021. Network intrusion detection system: A systematic study of machine learning and deep learning approaches.Transactions on Emerging Telecommunications Technologies32, 1 (2021), e4150

work page 2021
[2]

Mohiuddin Ahmed, Abdun Naser Mahmood, and Md Rafiqul Islam. 2016. A survey of anomaly detection techniques in financial domain.Future Generation Computer Systems55 (2016), 278–288

work page 2016
[3]

Khaled Gubran Al-Hashedi and Pritheega Magalingam. 2021. Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019.Computer Science Review40 (2021), 100402

work page 2021
[4]

Liron Bergman and Yedid Hoshen. 2020. Classification-based anomaly detection for general data.arXiv preprint arXiv:2005.02359(2020)

work page arXiv 2020
[5]

Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. 2000. LOF: identifying density-based local outliers. InProceedings of the 2000 ACM SIGMOD international conference on Management of data. 93–104

work page 2000
[6]

Yang Cao, Haolong Xiang, Hang Zhang, Ye Zhu, and Kai Ming Ting. 2025. Anom- aly detection based on isolation mechanisms: A survey.Machine Intelligence Research22, 5 (2025), 849–865

work page 2025
[7]

Varun Chandola, Arindam Banerjee, and Vipin Kumar. 2009. Anomaly detection: A survey.ACM computing surveys (CSUR)41, 3 (2009), 1–58

work page 2009
[8]

Zhaomin Chen, Chai Kiat Yeo, Bu Sung Lee, and Chiew Tong Lau. 2018. Autoencoder-based network anomaly detection. In2018 Wireless telecommu- nications symposium (WTS). IEEE, 1–5

work page 2018
[9]

Lei Cui, Youyang Qu, Gang Xie, Deze Zeng, Ruidong Li, Shigen Shen, and Shui Yu

work page
[10]

Security and privacy-enhanced federated learning for anomaly detection in IoT infrastructures.IEEE Transactions on Industrial Informatics18, 5 (2021), 3492–3500

work page 2021
[11]

Roy De Maesschalck, Delphine Jouan-Rimbaud, and Désiré L Massart. 2000. The mahalanobis distance.Chemometrics and intelligent laboratory systems50, 1 (2000), 1–18

work page 2000
[12]

2020.Machine learning in finance

Matthew F Dixon, Igor Halperin, Paul Bilokon, et al. 2020.Machine learning in finance. Vol. 1170. Springer

work page 2020
[13]

Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. 2021. Deep learning for medical anomaly detection–a survey. ACM Computing Surveys (CSUR)54, 7 (2021), 1–37

work page 2021
[14]

Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. 2020. DROCC: Deep robust one-class classification. InInternational conference on machine learning. PMLR, 3711–3721

work page 2020
[15]

Songqiao Han, Xiyang Hu, Hailiang Huang, Mingqi Jiang, and Yue Zhao. 2022. ADBench: Anomaly Detection Benchmark.arXiv preprint arXiv:2206.09426 (2022)

work page arXiv 2022
[16]

Ki Hyun Kim, Sangwoo Shim, Yongsub Lim, Jongseob Jeon, Jeongwoo Choi, Byungchan Kim, and Andre S Yoon. 2019. Rapp: Novelty detection with re- construction along projection pathway. InInternational Conference on Learning Representations

work page 2019
[17]

Zheng Li, Yue Zhao, Xiyang Hu, Nicola Botta, Cezar Ionescu, and George Chen

work page
[18]

Ecod: Unsupervised outlier detection using empirical cumulative distribu- tion functions.IEEE Transactions on Knowledge and Data Engineering(2022)

work page 2022
[19]

Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation forest. In2008 eighth ieee international conference on data mining. IEEE, 413–422

work page 2008
[20]

Victor Livernoche, Vineet Jain, Yashar Hezaveh, and Siamak Ravanbakhsh. 2024. On Diffusion Modeling for Anomaly Detection. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id= lR3rk7ysXz

work page 2024
[21]

Minh-Nghia Nguyen and Ngo Anh Vien. 2018. Scalable and interpretable one- class svms with deep learning and random fourier features. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 157–172

work page 2018
[22]

Guansong Pang, Longbing Cao, Ling Chen, and Huan Liu. 2018. Learning rep- resentations of ultrahigh-dimensional data for random distance-based outlier detection. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 2041–2050

work page 2018
[23]

Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. 2021. Deep learning for anomaly detection: A review.ACM Computing Surveys (CSUR) 54, 2 (2021), 1–38

work page 2021
[24]

Emanuel Parzen. 1962. On estimation of a probability density function and mode. The annals of mathematical statistics33, 3 (1962), 1065–1076

work page 1962
[25]

Gabriel Peyré, Marco Cuturi, et al. 2019. Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning11, 5-6 (2019), 355–607

work page 2019
[26]

Emanuele Principi, Fabio Vesperini, Stefano Squartini, and Francesco Piazza. 2017. Acoustic novelty detection with adversarial autoencoders. In2017 International Joint Conference on Neural Networks (IJCNN). IEEE, 3324–3330

work page 2017
[27]

Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph

work page
[28]

InInternational conference on machine learning

Neural transformation learning for deep anomaly detection beyond images. InInternational conference on machine learning. PMLR, 8703–8714

work page
[29]

Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. 2000. Efficient algo- rithms for mining outliers from large data sets. InProceedings of the 2000 ACM SIGMOD international conference on Management of data. 427–438

work page 2000
[30]

Shebuti Rayana. 2016. ODDS library.Stony Brook University, Department of Computer Sciences(2016)

work page 2016
[31]

Stephen Roberts and Lionel Tarassenko. 1994. A probabilistic resource allocating network for novelty detection.Neural Computation6, 2 (1994), 270–284

work page 1994
[32]

Lukas Ruff, Jacob R Kauffmann, Robert A Vandermeulen, Grégoire Montavon, Wojciech Samek, Marius Kloft, Thomas G Dietterich, and Klaus-Robert Müller

work page
[33]

IEEE109, 5 (2021), 756–795

A unifying review of deep and shallow anomaly detection.Proc. IEEE109, 5 (2021), 756–795

work page 2021
[34]

Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. 2018. Deep one-class classification. InInternational conference on machine learning. PMLR, 4393–4402

work page 2018
[35]

Maziar Sanjabi, Jimmy Ba, Meisam Razaviyayn, and Jason D Lee. 2018. On the convergence and robustness of training gans with regularized optimal transport. Advances in Neural Information Processing Systems31 (2018)

work page 2018
[36]

Thomas Schlegl, Philipp Seeböck, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. 2017. Unsupervised anomaly detection with generative ad- versarial networks to guide marker discovery. InInternational conference on information processing in medical imaging. Springer, 146–157

work page 2017
[37]

Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. 1999. Support vector method for novelty detection.Advances in neural information processing systems12 (1999)

work page 1999
[38]

Tom Shenkar and Lior Wolf. 2022. Anomaly Detection for Tabular Data with Inter- nal Contrastive Learning. InInternational Conference on Learning Representations. https://openreview.net/forum?id=_hszZbt46bT

work page 2022
[39]

2003.A novel anomaly detection scheme based on principal component classifier

Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and LiWu Chang. 2003.A novel anomaly detection scheme based on principal component classifier. Technical Report. Miami Univ Coral Gables Fl Dept of Electrical and Computer Engineering

work page 2003
[40]

Titouan Vayer, Laetitia Chapel, Rémi Flamary, Romain Tavenard, and Nicolas Courty. 2018. Optimal transport for structured data with application on graphs. arXiv preprint arXiv:1805.09114(2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[41]

Hu Wang, Guansong Pang, Chunhua Shen, and Congbo Ma. 2019. Unsuper- vised representation learning by predicting random distances.arXiv preprint arXiv:1912.12186(2019)

work page arXiv 2019
[42]

David H Wolpert and William G Macready. 1997. No free lunch theorems for optimization.IEEE transactions on evolutionary computation1, 1 (1997), 67–82

work page 1997
[43]

Hongzuo Xu, Guansong Pang, Yijie Wang, and Yongjun Wang. 2023. Deep Isolation Forest for Anomaly Detection.IEEE Transactions on Knowledge and Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Hangting Ye, He Zhao, Wei Fan, Xiaozhuang Song, Dandan Guo, Yi Chang, and Hongyuan Zha Data Engineering35, 12 (2023), 12591–12604. https://doi.org/10.1109/TKDE...

work page doi:10.1109/tkde.2023 2023
[44]

Hangting Ye, Wei Fan, Xiaozhuang Song, Shun Zheng, He Zhao, Dan dan Guo, and Yi Chang. 2024. PTaRL: Prototype-based Tabular Representation Learn- ing via Space Calibration. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=G32oY4Vnm8

work page 2024
[45]

Hangting Ye, Zhining Liu, Xinyi Shen, Wei Cao, Shun Zheng, Xiaofan Gui, Huishuai Zhang, Yi Chang, and Jiang Bian. 2023. UADB: Unsupervised Anomaly Detection Booster.arXiv preprint arXiv:2306.01997(2023)

work page arXiv 2023
[46]

Hangting Ye, He Zhao, Wei Fan, Mingyuan Zhou, Dan dan Guo, and Yi Chang

work page
[47]

InThe Thirteenth International Conference on Learning Representations

DRL: Decomposed Representation Learning for Tabular Anomaly Detection. InThe Thirteenth International Conference on Learning Representations

work page
[48]

Jiaxin Yin, Yuanyuan Qiao, Zitang Zhou, Xiangchao Wang, and Jie Yang. 2024. MCM: Masked Cell Modeling for Anomaly Detection in Tabular Data. InThe Twelfth International Conference on Learning Representations

work page 2024
[49]

Mikhail Yurochkin, Sebastian Claici, Edward Chien, Farzaneh Mirzazadeh, and Justin M Solomon. 2019. Hierarchical optimal transport for document represen- tation.Advances in neural information processing systems32 (2019)

work page 2019
[50]

Daochen Zha, Kwei-Herng Lai, Mingyang Wan, and Xia Hu. 2020. Meta-AAD: Active anomaly detection with deep reinforcement learning. In2020 IEEE Inter- national Conference on Data Mining (ICDM). IEEE, 771–780

work page 2020
[51]

Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. 2020. Deepemd: Few- shot image classification with differentiable earth mover’s distance and structured classifiers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12203–12213

work page 2020
[52]

Yue Zhao, Zain Nasrullah, and Zheng Li. 2019. Pyod: A python toolbox for scalable outlier detection.Journal of machine learning research20, 96 (2019), 1–7

work page 2019
[53]

Increase (%)

Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. 2018. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. InInternational conference on learning represen- tations. Calibrating Tabular Anomaly Detection via Optimal Transport Conference acronym ’XX, June 03–05, 2018, Woodstock, NY A ...

work page arXiv 2018