Pith. sign in

REVIEW 4 major objections 5 minor 69 references

Investigating Mask-aware Prototype Learning for Tabular Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learnable masks and optimal-transport prototypes lift tabular anomaly detection by about 4% over the previous best method.

desk verdict Solid empirical integration of mask modeling and prototype learning for tabular AD, but the inference-time OT scoring function is underspecified and must be fixed before the central 4% claim can be trusted. read the letter →

arxiv 2506.02757 v1 pith:NDA6SP4C submitted 2025-06-03 cs.LG cs.AI

Ruiying Lu , Jinhan Liu , Chuan Du , Dandan Guo This is my paper
classification cs.LGcs.AI
keywords tabularanomalydetectionmaskmodelingprototypelearningoptimaltransportone-classclassificationreconstruction-basednon-parametrictransformerassociationprototypes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes PTAD, a reconstruction-based method for tabular anomaly detection that trains only on normal samples. It argues that standard reconstruction fails when representations are entangled and when the shared patterns across normal samples are ignored. PTAD masks inputs both in the raw data space and in a projection space spanned by orthogonal basis vectors, then decodes the masked representations and learns cross-feature association prototypes. Both the basis vectors and the prototypes are learned as optimal transport distribution-matching problems, and their transport distances are added to the reconstruction error as calibration scores. Across 20 tabular benchmarks, the method reports an average AUC-PR of 0.7513 and an average AUC-ROC of 0.9064, around 4% better than MCM, the second-best comparison method.

What carries the argument

The central mechanism is a two-stage masking-plus-prototype reconstruction loop built on the Non-Parametric Transformer (NPT), an attention architecture that alternates attention between datapoints (ABD) with attention between attributes (ABA). In the first stage, a data-space soft mask is produced by a small network and applied to the input; an NPT encoder maps the masked input into a projection space spanned by K learnable orthogonal basis vectors, and for each basis vector a binary mask is built by thresholding elementwise squared distances between the latent representation and that basis vector. In the second stage, the decoder reconstructs the K masked branches in parallel and, from its attribute-attention maps, forms association vectors; M association prototypes summarize normal cross-feature correlation patterns. Both basis-vector learning and association-prototype learning are posed as entropic optimal transport problems between empirical uniform distributions, with the Sinkhorn-style transport plans defining per-sample calibration distances, and the orthogonality constraint on the basis vectors is meant to keep the projection space disentangled.

What would settle it

On a synthetic tabular benchmark where normal samples are drawn from two well-separated Gaussian clusters and anomalies are placed along a direction orthogonal to the first principal component of the normal data, train PTAD and compare its AUC-PR and AUC-ROC against the same model without the calibration distances. If the calibrated scores add no improvement, or if the projection-space masks cannot be visually distinguished between normal and anomalous samples, the assumption that orthogonal basis vectors capture normal patterns is not doing the work the paper claims.

Watch

Extended reading notes

Core claim

The paper claims that tabular anomaly detection is improved by making the mask itself learnable and by adding two distribution-matching calibration terms to the reconstruction error. A learned soft mask is applied in the data space, and then, in a projection space spanned by a group of learnable orthogonal basis vectors, multiple binary masks are generated by thresholding the squared distance of each latent coordinate to each basis vector, so that positions far from the learned normal patterns are masked away. An NPT encoder and decoder capture sample-sample and feature-feature dependencies while reconstructing the masked branches in parallel. The decoder's attribute attention also yields association vectors, and the paper learns association prototypes summarizing normal cross-feature correlation patterns. Optimal transport between latent representations and basis vectors, and between association vectors and association prototypes, gives transport plans whose per-sample costs are added to the reconstruction loss as calibration distances. On 20 tabular datasets, the full method achieves an average AUC-PR of 0.7513 and average AUC-ROC of 0.9064, about 4% above MCM, and ablations attribute the gains to data-space masking, multiple projection-space masks, association prototypes, the orthogonality constraint, and optimal transport distances.

Load-bearing premise

The load-bearing premise is that a small set of orthogonal reference vectors learned from normal data captures the global normal patterns of every tabular dataset, and that the latent coordinates farthest from those vectors are the ones that should be masked to expose anomalies.

Editorial extensions

If this is right

  • Reconstruction-based tabular anomaly detection can be improved by replacing random or fixed masks with masks learned in both the data space and a projection space spanned by orthogonal basis vectors.
  • Anomaly scores need not come from reconstruction error alone: transport costs from a sample's latent representation to learned normal basis vectors, and from its feature-association vector to learned association prototypes, can be added as calibration terms.
  • On the 20 tested benchmarks, the method reports the best or second-best AUC-ROC on 14 datasets and the largest per-dataset gains on Optdigits and Wbc, with 8.12% and 6.56% AUC-PR improvements respectively.
  • The masking and association-prototype modules are model-agnostic: in the paper's experiments they improve MLP, vanilla Transformer, and NPT backbones.
  • Using optimal transport instead of MSE to learn the basis vectors and prototypes yields a clear gain, 0.7513 versus 0.6649 average AUC-PR, suggesting that distribution-level matching is doing real work.
  • The method detects dependency and cluster anomalies better than the compared baselines on the cardiotocography-based robustness study, consistent with its explicit modeling of cross-feature associations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method's success likely depends on normal data being approximately linear near a low-dimensional subspace in the projection space; a natural untested extension would be learning a mixture of orthogonal basis sets to handle multimodal normal data.
  • Because the projection-space masks assign a masking rate to each feature, the method produces a per-feature explanation of why a sample is anomalous, which could be used directly in fraud or medical triage even though the paper does not evaluate explanation quality.
  • The association-prototype calibration distance quantifies which cross-feature dependencies are abnormal, so pairing it with the feature-level distance could yield a two-level explanation: which features deviate and which feature-feature relationships are anomalous.
  • The paper fixes five basis vectors and five association prototypes for all datasets; testing whether optimal counts scale with dataset dimensionality or cluster structure would clarify how much of the gain depends on these hyperparameters.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes PTAD, a mask-aware prototype learning method for tabular anomaly detection in the one-class setting. It introduces a learnable soft mask in the data space, multiple hard masks in a projection space defined by orthogonal basis vectors, and association prototypes extracted from cross-feature attention in a decoder. The learning of basis vectors and association prototypes is formulated as two optimal transport problems, and the resulting transport calibration distances are combined with reconstruction error into an anomaly score. The method is evaluated on 20 tabular benchmarks against 13 baselines, reporting average AUC-PR 0.7513 vs. 0.7111 for the second-best method MCM, and average AUC-ROC 0.9064 vs. 0.8914, along with ablations, robustness to four anomaly types, computational cost comparisons, and visualizations.

Significance. If the empirical claims hold, PTAD would offer a competitive and conceptually interesting approach to tabular anomaly detection, integrating mask modeling, prototype learning, and optimal transport in a unified framework. The paper's strengths include a wide experimental coverage (20 datasets, 13 baselines, ablations, four anomaly types, cost comparisons) and a clear presentation of the training objectives. However, the reported improvements come without error bars, statistical tests, or code/data release, and the inference-time scoring procedure is incompletely specified. These gaps are fixable and do not undermine the core idea, but they currently limit the verifiability of the central claim.

major comments (4)
  1. [Section IV, Eq. (13)] The anomaly score s_cab_n in Eq. (13) uses s_bv_n and s_ap_n defined through the transport plans T* and \hat T* that solve the OT problems in Eqs. (8) and (10) over training-batch distributions with N samples. The manuscript does not specify how these scores are computed at inference time for a single test sample x. Per-sample evaluation with N=1 is mathematically feasible (the row and column constraints in Eq. (8) are satisfied by T_1k = 1/K for all k, so the reviewer's claim that the constraints are unsatisfiable is incorrect), but the paper never states this. If instead the plan is recomputed over a test batch, then the score of each sample depends on the composition and size of the batch, contradicting the per-sample scoring function S: R^d -> R defined in Section III. The authors must specify the exact inference procedure—per-sample N=1, or batch recomputation with a clear definition of the batch—and justify that the reported results follow from a single well-defined scoring function.
  2. [Section IV-B, Eq. (10)] The association vector pi_n is defined from the query and key matrices of "the n-th sample in the ABA layer" of the decoder, but the decoder is applied to K masked representations H_k (k=1..K) to produce K reconstructions per sample. The paper does not specify which of the K branches (or whether an unmasked forward pass) yields the pi_n used in the OT problem of Eq. (10) and in the calibration score s_ap_n. This ambiguity affects both training and inference, since Eq. (10) and the final anomaly score depend directly on pi_n. Please clarify the exact forward pass from which pi_n is extracted.
  3. [Section V-A, Tables I and II] The paper reports only averages over three runs without standard deviations or statistical tests. The average improvement over MCM is 0.0402 in AUC-PR and 0.0150 in AUC-ROC, but on 5 of 20 datasets (e.g., Cardiotocography, Fraud, Satellite, Shuttle, Thyroid) the proposed method underperforms MCM on AUC-PR. Without per-seed results or a paired significance test (e.g., Wilcoxon signed-rank over datasets), the claim of "around 4% improvement" is not statistically supported. Please provide variance estimates or significance tests, or temper the claim accordingly.
  4. [Section IV-A, Eq. (5)] The projection-space mask in Eq. (5) is a hard binary mask defined by a threshold on squared distances. The manuscript states that the model is trained end-to-end under loss L in Eq. (12), but the hard mask is non-differentiable with respect to its inputs. The paper does not describe how gradients are handled at the mask selection step (e.g., straight-through estimator, Gumbel-softmax, or treating the mask as a non-learnable operation during backpropagation). This is a load-bearing implementation detail for reproducibility and should be clarified.
minor comments (5)
  1. [Section V-A] The phrase "around 4% improvement" should specify whether this is an absolute increase in percentage points (4.02 pp AUC-PR) or a relative improvement (5.65% AUC-PR relative to MCM's 0.7111).
  2. [Section III] There is a typo: "stochastic masking matrix with the same dimention as input" should be "dimension".
  3. [Section V] For reproducibility, please provide version numbers or commit hashes for the PYOD and DeepOD libraries used in the baseline implementations.
  4. [Section V-D] The discussion of Figs. 8 and 9 describes performance as "stable" or "sensitive" without quantitative support; consider adding error bars or numerical values.
  5. [Appendix A] The phrase "Noting that we take the single vector as example" should be "Noting that we take the single vector as an example".

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the anomaly score is a self-supervised reconstruction/prototype-distance objective trained only on normal data, and the reported benchmark gains are external empirical comparisons not implied by the paper's equations or by the authors' prior work.

full rationale

The paper's derivation chain is: encode masked input into a projection space; learn orthogonal basis vectors and association prototypes by optimal transport matching against the normal training batch; then score each test point by reconstruction error plus OT calibration distances. Equation 13 defines s_cab = s_rec + κ s_bv + α s_ap, where s_bv and s_ap are the same per-sample transport costs minimized in Equations 8 and 10. Reusing a training objective as an anomaly score is standard one-class anomaly detection and is not circular: training uses only normal samples, no anomaly labels or test labels enter, and the prototypes are fitted solely to normal data. A separate specification gap, unrelated to circularity, is that Equation 13 invokes transport plans T* and \hat{T}* defined over batches through the constraint sets in Equations 8 and 10, so per-sample inference scoring is underspecified unless a test batch is scored jointly; this is a reproducibility/correctness issue rather than a tautology. The paper does contain self-citations to the authors' prior prototype/OT work, including references [20], [31], and [39]–[45], but the load-bearing evidence for the central claim is the benchmark comparison against 13 baselines on 20 datasets, which is external to those citations and not implied by them. The basis-vector assumption is an empirical hypothesis tested by the experiments, not an input that forces the reported result. I assign score 2 rather than 0 only because several self-citations appear, but none is load-bearing and no derivation step reduces to its own inputs.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central method rests on one-class assumptions, reconstruction assumptions, and two modeling postulates (orthogonal basis vectors and association prototypes). All are internal to the method; none are validated by external benchmarks beyond the reported experiments.

free parameters (5)
  • Number of basis vectors K = 5
    Set to 5 as a trade-off between performance and cost based on Fig. 7; controls mask granularity.
  • Number of association prototypes M = 5
    Set to 5 from Fig. 7; controls capacity of cross-feature prototype matching.
  • Entropic regularization lambda = 0.1
    Used in both OT problems (Eqs. 8 and 10); chosen by hand, affects smoothness of transport plans.
  • Loss weights lambda_bv, lambda_ap, lambda_orth = 1, 1, 0.1
    Weights in Eq. 12; orthogonality weight chosen from Fig. 8.
  • Score weights kappa and alpha = 0.01 each
    Weights in Eq. 13; chosen from Fig. 9 as insensitive values.
assumptions (6)
  • domain assumption Training set contains only normal samples (one-class setting).
    Section III formulates the problem this way; all training losses are computed on normal data only.
  • domain assumption Reconstruction-based AD assumption: a model trained only on normal samples will reconstruct normal inputs well and abnormal inputs poorly.
    Section I states this assumption explicitly and it underlies L_rec as the main anomaly score.
  • domain assumption NPT layers capture sample-sample and feature-feature dependencies that are useful for tabular data.
    Section III summarizes NPT; the encoder and decoder are both NPT layers.
  • ad hoc to paper Normal samples share global intrinsic patterns representable by K orthogonal basis vectors.
    Section IV-A states: 'we assume that they share some global intrinsic characteristic patterns in the P-space'; this is a modeling postulate specific to PTAD.
  • ad hoc to paper Association vectors extracted from attention maps can represent cross-feature correlations, and M prototypes summarize their normal distribution.
    Section IV-B constructs association vectors from query-key products and matches them to learned prototypes via OT.
  • standard math Discrete uniform distributions over samples, basis vectors, and prototypes are appropriate for OT matching.
    Eqs. 6-7 and Eq. 10 use uniform empirical distributions; OT theory supports matching discrete measures, but the uniform weighting is a modeling choice.
invented entities (2)
  • Projection space basis vectors B
    purpose: Prototypes representing shared normal patterns; used for mask generation and OT-based anomaly calibration.
    Learned parameters internal to the model; only evidence is benchmark performance, no external falsifiable prediction.
  • Association prototypes Upsilon
    purpose: Summarize normal cross-feature correlation patterns from decoder attention; used in anomaly scoring.
    Learned parameters; no independent handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigating Mask-aware Prototype Learning for Tabular Anomaly Detection." pith.science (2026). https://pith.science/paper/NDA6SP4C

@misc{pith2026250602757,
  author       = {Pith},
  title        = {Pith review of: Investigating Mask-aware Prototype Learning for Tabular Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDA6SP4C}},
  note         = {Machine review of arXiv:2506.02757}
}
read the original abstract

Tabular anomaly detection, which aims at identifying deviant samples, has been crucial in a variety of real-world applications, such as medical disease identification, financial fraud detection, intrusion monitoring, etc. Although recent deep learning-based methods have achieved competitive performances, these methods suffer from representation entanglement and the lack of global correlation modeling, which hinders anomaly detection performance. To tackle the problem, we incorporate mask modeling and prototype learning into tabular anomaly detection. The core idea is to design learnable masks by disentangled representation learning within a projection space and extracting normal dependencies as explicit global prototypes. Specifically, the overall model involves two parts: (i) During encoding, we perform mask modeling in both the data space and projection space with orthogonal basis vectors for learning shared disentangled normal patterns; (ii) During decoding, we decode multiple masked representations in parallel for reconstruction and learn association prototypes to extract normal characteristic correlations. Our proposal derives from a distribution-matching perspective, where both projection space learning and association prototype learning are formulated as optimal transport problems, and the calibration distances are utilized to refine the anomaly scores. Quantitative and qualitative experiments on 20 tabular benchmarks demonstrate the effectiveness and interpretability of our model.

Figures

Figures reproduced from arXiv: 2506.02757 by the authors.

Figure 1
Figure 1. Overall framework: The input data is first masked with the generated soft mask in the data space and then encoded through an NPT layer. Next, in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 4
Figure 4. The average ranking results are also shown in Fig. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 2
Figure 2. AUC-PR of models over 20 datasets (↑). DeepSVDD GOAD IForest NeuTralAD ICL GMM DTE-C LOF OCSVM NPT-AD KNN MCM MCM + NPTOurs 0 2 4 6 8 10 Rank ( ) 10.80 9.90 8.95 8.70 8.35 8.35 8.00 7.20 7.05 6.75 6.25 5.60 5.50 3.55 Average Ranking of Models based on AUC-PR [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Ranking of model based on AUC-ROC(↓). vi) Orthogonality Constrain: We further validate the or￾thogonality constrain of basis vectors, the performance gap showcases the orthogonality contributes to anomaly detection by forming disentangled features. Overall, the compreh…
Figure 7
Figure 7. Figure 7: Performance with different numbers of basis vectors (left) and [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison Results of different weights of orthogonal constrain [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Comparison Results of different weights of calibration distances [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Visualization of P-space masks. The left figure corresponds to the normal sample, and the right figure refers to the abnormal sample, both possessing [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Visualization of the soft masks. The left figure corresponds to the normal sample, and the right figure refers to the abnormal sample, both possessing [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 47 canonical work pages

  1. [1]

    Random partitioning forest for point-wise and collective anomaly detection—application to network intrusion detection,

    P.-F. Marteau, “Random partitioning forest for point-wise and collective anomaly detection—application to network intrusion detection,”IEEE Transactions on Information Forensics and Security, vol. 16, pp. 2157– 2172, 2021

    work page 2021

  2. [2]

    Network intrusion detection system: A systematic study of machine learning and deep learning approaches,

    Z. Ahmad, A. Shahid Khan, C. Wai Shiang, J. Abdullah, and F. Ahmad, “Network intrusion detection system: A systematic study of machine learning and deep learning approaches,”Transactions on Emerging Telecommunications Technologies, vol. 32, no. 1, p. e4150, 2021

    2021

  3. [3]

    Uadb: Unsupervised anomaly detection booster,

    H. Ye, Z. Liu, X. Shen, W. Cao, S. Zheng, X. Gui, H. Zhang, Y . Chang, and J. Bian, “Uadb: Unsupervised anomaly detection booster,” in2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2023, pp. 2593–2606

    work page 2023

  4. [4]

    Generating synthetic data in finance: opportunities, challenges and pitfalls,

    S. A. Assefa, D. Dervovic, M. Mahfouz, R. E. Tillman, P. Reddy, and M. Veloso, “Generating synthetic data in finance: opportunities, challenges and pitfalls,” inProceedings of the First ACM International Conference on AI in Finance, 2020, pp. 1–8

    work page 2020

  5. [5]

    Deep learning for medical anomaly detection–a survey,

    T. Fernando, H. Gammulle, S. Denman, S. Sridharan, and C. Fookes, “Deep learning for medical anomaly detection–a survey,”ACM Com- puting Surveys (CSUR), vol. 54, no. 7, pp. 1–37, 2021

    2021

  6. [6]

    Financial fraud detection ap- plying data mining techniques: A comprehensive review from 2009 to 2019,

    K. G. Al-Hashedi and P. Magalingam, “Financial fraud detection ap- plying data mining techniques: A comprehensive review from 2009 to 2019,”Computer Science Review, vol. 40, p. 100402, 2021

    2009

  7. [7]

    Deepsyslog: Deep anomaly detection on syslog using sentence embedding and metadata,

    J. Zhou, Y . Qian, Q. Zou, P. Liu, and J. Xiang, “Deepsyslog: Deep anomaly detection on syslog using sentence embedding and metadata,” IEEE Transactions on Information Forensics and Security, vol. 17, pp. 3051–3061, 2022

    work page 2022

  8. [8]

    An empirical evaluation of deep learning for network anomaly detection,

    R. K. Malaiya, D. Kwon, S. C. Suh, H. Kim, I. Kim, and J. Kim, “An empirical evaluation of deep learning for network anomaly detection,” IEEE Access, vol. 7, pp. 140 806–140 817, 2019

    work page 2019

Show all 69 references
  1. [9]

    Transformation based deep anomaly de- tection in astronomical images,

    E. Reyes and P. A. Est ´evez, “Transformation based deep anomaly de- tection in astronomical images,” in2020 International Joint Conference on Neural Networks (IJCNN). IEEE, 2020, pp. 1–8

    2020

  2. [10]

    A unifying review of deep and shallow anomaly detection,

    L. Ruff, J. R. Kauffmann, R. A. Vandermeulen, G. Montavon, W. Samek, M. Kloft, T. G. Dietterich, and K.-R. M ¨uller, “A unifying review of deep and shallow anomaly detection,”Proceedings of the IEEE, vol. 109, no. 5, pp. 756–795, 2021

    2021

  3. [11]

    Decouple and resolve: Transformer-based models for online anomaly detection from weakly labeled videos,

    T. Liu, C. Zhang, K.-M. Lam, and J. Kong, “Decouple and resolve: Transformer-based models for online anomaly detection from weakly labeled videos,”IEEE Transactions on Information Forensics and Secu- rity, vol. 18, pp. 15–28, 2023

    2023

  4. [12]

    Anomalynet: An anomaly detection network for video surveillance,

    J. T. Zhou, J. Du, H. Zhu, X. Peng, Y . Liu, and R. S. M. Goh, “Anomalynet: An anomaly detection network for video surveillance,” IEEE Transactions on Information Forensics and Security, vol. 14, no. 10, pp. 2537–2550, 2019

    2019

  5. [13]

    Data- efficient and interpretable tabular anomaly detection,

    C.-H. Chang, J. Yoon, S. ¨O. Arik, M. Udell, and T. Pfister, “Data- efficient and interpretable tabular anomaly detection,” inProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 190–201

    2023

  6. [14]

    Neural transformation learning for deep anomaly detection beyond images,

    C. Qiu, T. Pfrommer, M. Kloft, S. Mandt, and M. Rudolph, “Neural transformation learning for deep anomaly detection beyond images,” in International conference on machine learning. PMLR, 2021, pp. 8703– 8714

    2021

  7. [15]

    Anomaly detection for tabular data with internal contrastive learning,

    T. Shenkar and L. Wolf, “Anomaly detection for tabular data with internal contrastive learning,” inInternational conference on learning representations, 2022

    2022

  8. [16]

    Beyond individual input for deep anomaly detection on tabular data,

    H. Thimonier, F. Popineau, A. Rimmel, and B.-L. DOAN, “Beyond individual input for deep anomaly detection on tabular data,” inForty- first International Conference on Machine Learning, 2024

    2024

  9. [17]

    Self-attention between datapoints: Going beyond individual input- output pairs in deep learning,

    J. Kossen, N. Band, C. Lyle, A. N. Gomez, T. Rainforth, and Y . Gal, “Self-attention between datapoints: Going beyond individual input- output pairs in deep learning,”Advances in Neural Information Pro- cessing Systems, vol. 34, pp. 28 742–28 756, 2021

    2021

  10. [18]

    Mcm: Masked cell modeling for anomaly detection in tabular data,

    J. Yin, Y . Qiao, Z. Zhou, X. Wang, and J. Yang, “Mcm: Masked cell modeling for anomaly detection in tabular data,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  11. [19]

    Deep learning for anomaly detection: A review,

    G. Pang, C. Shen, L. Cao, and A. V . D. Hengel, “Deep learning for anomaly detection: A review,”ACM computing surveys (CSUR), vol. 54, no. 2, pp. 1–38, 2021

    2021

  12. [20]

    Ptarl: Prototype-based tabular representation learning via space calibration,

    H. Ye, W. Fan, X. Song, S. Zheng, H. Zhao, D. dan Guo, and Y . Chang, “Ptarl: Prototype-based tabular representation learning via space calibration,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  13. [21]

    Sup- port vector machines,

    M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Sup- port vector machines,”IEEE Intelligent Systems and their applications, vol. 13, no. 4, pp. 18–28, 1998

    1998

  14. [22]

    Revisiting deep learning models for tabular data,

    Y . Gorishniy, I. Rubachev, V . Khrulkov, and A. Babenko, “Revisiting deep learning models for tabular data,”Advances in Neural Information Processing Systems, vol. 34, pp. 18 932–18 943, 2021

    2021

  15. [23]

    Deep weakly- supervised anomaly detection,

    G. Pang, C. Shen, H. Jin, and A. van den Hengel, “Deep weakly- supervised anomaly detection,” inProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 1795–1807

    2023

  16. [24]

    Rapp: Novelty detection with reconstruction along projection pathway,

    K. H. Kim, S. Shim, Y . Lim, J. Jeon, J. Choi, B. Kim, and A. S. Yoon, “Rapp: Novelty detection with reconstruction along projection pathway,” inInternational Conference on Learning Representations, 2019

    2019

  17. [25]

    Adbench: Anomaly detection benchmark,

    S. Han, X. Hu, H. Huang, M. Jiang, and Y . Zhao, “Adbench: Anomaly detection benchmark,”Advances in Neural Information Processing Sys- tems, vol. 35, pp. 32 142–32 159, 2022

    2022

  18. [26]

    Unsupervised anomaly detection with generative adversarial networks to guide marker discovery,

    T. Schlegl, P. Seeb ¨ock, S. M. Waldstein, U. Schmidt-Erfurth, and G. Langs, “Unsupervised anomaly detection with generative adversarial networks to guide marker discovery,” inInternational conference on information processing in medical imaging. Springer, 2017, pp. 146– 157

    2017

  19. [27]

    Applications of generative adversarial networks in anomaly detection: a systematic literature review,

    M. Sabuhi, M. Zhou, C.-P. Bezemer, and P. Musilek, “Applications of generative adversarial networks in anomaly detection: a systematic literature review,”Ieee Access, vol. 9, pp. 161 003–161 029, 2021

    2021

  20. [28]

    Neural prototype trees for interpretable fine-grained image recognition,

    M. Nauta, R. Van Bree, and C. Seifert, “Neural prototype trees for interpretable fine-grained image recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 933–14 943

    2021

  21. [29]

    Rethinking seman- tic segmentation: A prototype view,

    T. Zhou, W. Wang, E. Konukoglu, and L. Van Gool, “Rethinking seman- tic segmentation: A prototype view,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2582–2593

    2022

  22. [30]

    Prototype-representations for training data filtering in weakly-supervised information extraction,

    N. Zalmout and X. Li, “Prototype-representations for training data filtering in weakly-supervised information extraction,” inProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2022, pp. 467–474. 13

    2022

  23. [31]

    Hierarchi- cal vector quantized transformer for multi-class unsupervised anomaly detection,

    R. Lu, Y . Wu, L. Tian, D. Wang, B. Chen, X. Liu, and R. Hu, “Hierarchi- cal vector quantized transformer for multi-class unsupervised anomaly detection,”Advances in Neural Information Processing Systems, vol. 36, pp. 8487–8500, 2023

    2023

  24. [32]

    Vague prototype-oriented diffusion model for multi- class anomaly detection,

    Y . Li, Y . Feng, B. Chen, W. Chen, Y . Wang, X. Hu, C. Qu, M. Zhouet al., “Vague prototype-oriented diffusion model for multi- class anomaly detection,” inInternational Conference on Machine Learning, 2024

    2024

  25. [33]

    Computational op- timal transport: Complexity by accelerated gradient descent is better than by sinkhorn’s algorithm,

    P. Dvurechensky, A. Gasnikov, and A. Kroshnin, “Computational op- timal transport: Complexity by accelerated gradient descent is better than by sinkhorn’s algorithm,” inInternational conference on machine learning. PMLR, 2018, pp. 1367–1376

    2018

  26. [34]

    Optimal transport for domain adaptation,

    N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy, “Optimal transport for domain adaptation,”IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 9, pp. 1853–1865, 2016

    2016

  27. [35]

    Learning generative models with sinkhorn divergences,

    A. Genevay, G. Peyr ´e, and M. Cuturi, “Learning generative models with sinkhorn divergences,” inInternational Conference on Artificial Intelligence and Statistics, 2018, pp. 1608–1617

    2018

  28. [36]

    Optimal transport for deep gener- ative models: state of the art and research challenges,

    V . Huynh, D. Phung, and H. Zhao, “Optimal transport for deep gener- ative models: state of the art and research challenges,” inInternational Joint Conference on Artificial Intelligence 2021, 2021, pp. 4450–4457

    2021

  29. [37]

    Optimal transport for conditional domain matching and label shift,

    A. Rakotomamonjy, R. Flamary, G. Gasso, M. E. Alaya, M. Berar, and N. Courty, “Optimal transport for conditional domain matching and label shift,”Machine Learning, pp. 1–20, 2022

    2022

  30. [38]

    Multi-source domain adaptation via weighted joint distributions optimal transport,

    R. Turrisi, R. Flamary, A. Rakotomamonjy, and M. Pontil, “Multi-source domain adaptation via weighted joint distributions optimal transport,” in Uncertainty in Artificial Intelligence. PMLR, 2022, pp. 1970–1980

    2022

  31. [39]

    Learning to re-weight examples with optimal transport for imbalanced classifica- tion,

    D. Guo, Z. Li, M. Zheng, H. Zhao, M. Zhou, and H. Zha, “Learning to re-weight examples with optimal transport for imbalanced classifica- tion,” inAdvances in Neural Information Processing Systems, 2022, pp. 25 517–25 530

    2022

  32. [40]

    Enhancing minority classes by mixing: an adaptative optimal transport approach for long-tailed classification,

    J. Gao, H. Zhao, Z. Li, and D. Guo, “Enhancing minority classes by mixing: an adaptative optimal transport approach for long-tailed classification,”Advances in Neural Information Processing Systems, vol. 36, 2023

    2023

  33. [41]

    Distribution alignment optimization through neural collapse for long-tailed classification,

    J. Gao, H. Zhao, D. dan Guo, and H. Zha, “Distribution alignment optimization through neural collapse for long-tailed classification,” in Forty-first International Conference on Machine Learning, 2024

    2024

  34. [42]

    Modality- consistent prompt tuning with optimal transport,

    H. Ren, F. Tang, H. Zheng, H. Zhao, D. Guo, and Y . Chang, “Modality- consistent prompt tuning with optimal transport,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  35. [43]

    Adaptive distribution calibration for few-shot learning with hierarchical optimal transport,

    D. Guo, L. Tian, H. Zhao, M. Zhou, and H. Zha, “Adaptive distribution calibration for few-shot learning with hierarchical optimal transport,” Advances in neural information processing systems, vol. 35, pp. 6996– 7010, 2022

    2022

  36. [44]

    Learning prototype-oriented set representations for meta-learning,

    D. dan Guo, L. Tian, M. Zhang, M. Zhou, and H. Zha, “Learning prototype-oriented set representations for meta-learning,” inInterna- tional Conference on Learning Representations, 2022

    2022

  37. [45]

    Representing mixtures of word embeddings with mixtures of topic embeddings,

    G. D. Wang, dongsheng, H. Zhao, H. Zheng, K. Tanwisuth, B. Chen, M. Zhouet al., “Representing mixtures of word embeddings with mixtures of topic embeddings,” inInternational Conference on Learning Representations, 2022

    2022

  38. [46]

    Computational optimal transport: With applications to data science,

    G. Peyr ´e, M. Cuturiet al., “Computational optimal transport: With applications to data science,”Foundations and Trends® in Machine Learning, vol. 11, no. 5-6, pp. 355–607, 2019

    2019

  39. [47]

    Sinkhorn distances: Lightspeed computation of optimal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,”Advances in neural information processing systems, vol. 26, 2013

    2013

  40. [48]

    ODDS library,

    S. Rayana, “ODDS library,” 2016. [Online]. Available: https: //odds.cs.stonybrook.edu

    2016

  41. [49]

    Deep autoencoding gaussian mixture model for unsupervised anomaly detection,

    B. Zong, Q. Song, M. R. Min, W. Cheng, C. Lumezanu, D. Cho, and H. Chen, “Deep autoencoding gaussian mixture model for unsupervised anomaly detection,” inInternational conference on learning representa- tions, 2018

    2018

  42. [50]

    Classification-based anomaly detection for general data,

    L. Bergman and Y . Hoshen, “Classification-based anomaly detection for general data,”arXiv preprint arXiv:2005.02359, 2020

    2005 arXiv

  43. [51]

    Large batch optimiza- tion for deep learning: Training bert in 76 minutes,

    Y . You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C.-J. Hsieh, “Large batch optimiza- tion for deep learning: Training bert in 76 minutes,”arXiv preprint arXiv:1904.00962, 2019

    1904 arXiv

  44. [52]

    Lookahead optimizer: k steps forward, 1 step back,

    M. Zhang, J. Lucas, J. Ba, and G. E. Hinton, “Lookahead optimizer: k steps forward, 1 step back,”Advances in neural information processing systems, vol. 32, 2019

    2019

  45. [53]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  46. [54]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,”arXiv preprint arXiv:1608.03983, 2016

    2016 arXiv

  47. [55]

    Efficient algorithms for mining outliers from large data sets,

    S. Ramaswamy, R. Rastogi, and K. Shim, “Efficient algorithms for mining outliers from large data sets,” inProceedings of the 2000 ACM SIGMOD international conference on Management of data, 2000, pp. 427–438

    2000

  48. [56]

    Isolation forest,

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in2008 eighth ieee international conference on data mining. IEEE, 2008, pp. 413–422

    2008

  49. [57]

    Lof: identifying density-based local outliers,

    M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof: identifying density-based local outliers,” inProceedings of the 2000 ACM SIGMOD international conference on Management of data, 2000, pp. 93–104

    2000

  50. [58]

    Support vector method for novelty detection,

    B. Sch ¨olkopf, R. C. Williamson, A. Smola, J. Shawe-Taylor, and J. Platt, “Support vector method for novelty detection,”Advances in neural information processing systems, vol. 12, 1999

    1999

  51. [59]

    Detecting anomalies in cross-classified streams: a bayesian approach,

    D. Agarwal, “Detecting anomalies in cross-classified streams: a bayesian approach,”Knowledge and information systems, vol. 11, no. 1, pp. 29– 44, 2007

    2007

  52. [60]

    Deep one-class classification,

    L. Ruff, R. Vandermeulen, N. Goernitz, L. Deecke, S. A. Siddiqui, A. Binder, E. M ¨uller, and M. Kloft, “Deep one-class classification,” inInternational conference on machine learning. PMLR, 2018, pp. 4393–4402

    2018

  53. [61]

    On diffusion modeling for anomaly detection,

    V . Livernoche, V . Jain, Y . Hezaveh, and S. Ravanbakhsh, “On diffusion modeling for anomaly detection,”arXiv preprint arXiv:2305.18593, 2023

    2023 arXiv

  54. [62]

    Pyod: A python toolbox for scalable outlier detection,

    Y . Zhao, Z. Nasrullah, and Z. Li, “Pyod: A python toolbox for scalable outlier detection,”Journal of Machine Learning Research, vol. 20, no. 96, pp. 1–7, 2019. [Online]. Available: http://jmlr.org/papers/v20/ 19-011.html

    2019

  55. [63]

    Deep isolation forest for anomaly detection,

    H. Xu, G. Pang, Y . Wang, and Y . Wang, “Deep isolation forest for anomaly detection,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 12, pp. 12 591–12 604, 2023

    2023

  56. [64]

    Calibrated one-class classification for unsupervised time series anomaly detection,

    H. Xu, Y . Wang, S. Jian, Q. Liao, Y . Wang, and G. Pang, “Calibrated one-class classification for unsupervised time series anomaly detection,” IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  57. [65]

    Benchmarking unsupervised outlier de- tection with realistic synthetic data,

    G. Steinbuss and K. B ¨ohm, “Benchmarking unsupervised outlier de- tection with realistic synthetic data,”ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 15, no. 4, pp. 1–20, 2021

    2021

  58. [66]

    An algorithm for generating artificial test clusters,

    G. W. Milligan, “An algorithm for generating artificial test clusters,” Psychometrika, vol. 50, pp. 123–127, 1985

    1985

  59. [67]

    Pair-copula construc- tions of multiple dependence,

    K. Aas, C. Czado, A. Frigessi, and H. Bakken, “Pair-copula construc- tions of multiple dependence,”Insurance: Mathematics and economics, vol. 44, no. 2, pp. 182–198, 2009

    2009

  60. [68]

    Fault detection and diag- nosis in nonlinear systems,

    R. Martinez-Guerra and J. L. Mata-Machuca, “Fault detection and diag- nosis in nonlinear systems,”Understanding Complex Systems, Springer International Publishing, Cham, 2014

    2014

  61. [69]

    Hastie, R

    T. Hastie, R. Tibshirani, J. H. Friedman, and J. H. Friedman,The elements of statistical learning: data mining, inference, and prediction. Springer, 2009, vol. 2

    2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.