Pith. sign in

REVIEW 1 major objections 6 minor 47 references

Data Depth as a Risk

T0 review · 1 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper proves that halfspace depth is exactly the minimum expected 0-1 loss of a linear classifier on an artificial two-class problem where the query point is the single negative example, and builds a family of loss depths from this…

desk verdict A useful reframing of data depth as classification risk, with the caveat that the flagship halfspace-depth identity needs an atomlessness/tie-break repair before it is stated as exact. read the letter →

arxiv 2507.08518 v1 pith:OGUDTLX3 submitted 2025-07-11 stat.ML cs.LG

classification stat.MLcs.LG
keywords datadepthhalfspacelosssupportvectormachinelogisticregressionanomalydetectionRademachercomplexitystatisticallearning
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

This paper sets out to show that data depth, a classical unsupervised measure of how central a point is in a distribution, can be recast as a supervised learning risk. Concretely, the halfspace depth of a query $z$ with respect to a distribution $Q$ is exactly the smallest expected zero-one loss a linear classifier can achieve when half the mass is $Q$ labelled positive and $z$ itself is the single negative example. By substituting other losses and classifier families, the authors obtain a general class of "loss depths," including logistic-regression depth and SVM depth, that directly inherit the optimization speed and statistical convergence guarantees of those classifiers. The payoff is a unified view in which depth values become easy to compute in high dimensions, with uniform convergence rates that do not depend on the dimension.

What carries the argument

The load-bearing object is the artificially labelled distribution $P_{Q+|z-} = \tfrac12 Q\otimes \delta_1 + \tfrac12 \delta_z\otimes \delta_{-1}$, together with the identity $D_{l,\mathcal{H}}(z|Q)=\inf_{f\in\mathcal{H}} \mathbb{E}_{(x,y)\sim P_{Q+|z-}} l(f(x),y)$. For halfspace depth, $\mathcal{H}$ is the set of non-null affine linear classifiers and $l$ is the zero-one loss, so the depth becomes the best achievable classification error when the query is the only negative example. The generalisation-error lemmas (Lemma 4.1 and Lemma 4.2) transfer statistical rates from supervised learning to depths: the empirical depth error is bounded by the uniform deviation of the positive-part loss, and the Rademacher complexity of linear classes with bounded features and weights supplies the dimension-free $n^{-1/2}$ bound. Regularisation with $\lambda\|w\|^2$ ensures strict convexity and a unique minimiser, which defines the regularised depth and lets the proof control the regularisation-error term in the convergence argument.

What would settle it

Run logistic-regression depth or SVM depth on a distribution with unbounded support, such as a heavy-tailed $t$-distribution or a Cauchy distribution in $\mathbb{R}^2$, and plot $\sup_z |D(z|Q_n)-D(z|Q)|$ against $n$ on a log-log scale; if the slope is not $-1/2$ or if the supremum does not converge to zero uniformly, then Assumption B is essential and the theorem's stated rate fails beyond it.

Watch

Extended reading notes

Core claim

The central claim is Proposition 2.2: for any distribution $Q$ and query point $z$, $$\mathrm{HD}(z|Q) = \inf_{f \in \mathrm{Lin}^*} \mathbb{E}_{(x,y)\sim P_{Q+|z-}} 2 L_{0-1}(f(x),y),$$ where the artificially labelled distribution $P_{Q+|z-}$ puts half its mass on $Q$ with label $+1$ and half on the Dirac mass $\delta_z$ with label $-1$, and $\mathrm{Lin}^*$ is the set of non-null affine linear classifiers. In words, halfspace depth is the minimum misclassification rate of a linear rule separating the data from the single query point; the proof shifts the separating hyperplane through $z$ and normalizes its normal vector. Replacing the zero-one loss by any loss $l$ and the linear class by any class $\mathcal{H}$ defines the loss depth $D_{l,\mathcal{H}}(z|Q)=\inf_{f\in\mathcal{H}}\mathbb{E}_{(x,y)\sim P_{Q+|z-}} l(f(x),y)$. With a convex loss and an $L^2$ regulariser $\lambda\|w\|^2$, the optimization is $2\lambda$-strongly convex and has a unique minimiser, so the regularised depth is well defined. Under bounded-feature and bounded-weight assumptions, the framework yields uniform $O(n^{-1/2})$ convergence rates for logistic-regression depth and SVM depth, dimension-independent, and the resulting depths are bounded in $[0,1]$, convex for plain linear features, Lipschitz, and computable by standard algorithms.

Load-bearing premise

The uniform $n^{-1/2}$ convergence rates require the data distribution to have bounded support (all feature norms at most $B$) and, for unregularised logistic depth, bounded weights; for distributions with unbounded support, the stated uniform rate is not established.

Editorial extensions

If this is right

  • Computing the depth of a point is no harder than fitting one regularised classifier, so loss depths scale to high-dimensional data where classical halfspace-depth algorithms become computationally prohibitive.
  • Any classifier with a convex loss and Lipschitz features inherits a dimension-free $O(n^{-1/2})$ convergence rate for its associated depth, not just logistic and hinge losses.
  • Anomaly detection with loss depths becomes robust to label contamination because each query point is scored by its own optimisation problem rather than by a single one-class model trained on potentially dirty data.
  • The depths are bounded in $[0,1]$, convex in $z$ for plain linear features, and Lipschitz in both $z$ and the distribution $Q$, so depth regions are closed and small perturbations of the input or the distribution cannot cause large score swings.
  • On MNIST and Fashion-MNIST, a single logistic neuron used as a depth reaches AUC scores competitive with deep architectures, suggesting that per-point optimisation can sometimes replace memorising the dataset in a deep network.

Reading between the lines

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

  • If depth is a risk, then the spread of depth values over a dataset becomes a measure of how much structure the chosen classifier class can describe; using this spread to select classifier complexity is a model-selection criterion the paper only gestures at.
  • The identity also suggests that lower bounds or impossibility results for linear classification transfer directly to halfspace depth on the same distribution, connecting depth theory to minimax learning theory.
  • A testable extension is to replace the single negative Dirac with a small negative class or a kernel density estimate of the positive class, yielding a family of localised depths with tunable sensitivity to outliers.
  • Rescaling the mixture weight away from $1/2$ would define a parameterised depth with controlled robustness to contaminated regions; the paper fixes the weight but does not explore this degree of freedom.
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

1 major / 6 minor

Summary. The paper proposes to view data depth as the minimum expected loss of a classifier trained on an artificial labelled distribution in which the query point z is a single negative example and the reference distribution Q is the positive class. The authors prove (Prop. 2.2) that for linear classifiers and 0-1 loss this recovers halfspace depth, then define a general family of 'loss depths' obtained by varying the loss and classifier family. They focus on logistic-regression and SVM instantiations, prove Lipschitzness, convexity, boundedness, and O(n^{-1/2}) statistical convergence rates under bounded-support assumptions, and provide anomaly-detection experiments on tabular and image data.

Significance. If the central equivalence were valid, the paper would make a valuable connection between data depth and supervised learning, giving depths computational efficiency, optimization guarantees, and dimension-independent statistical rates. The proposed loss-depth family is a genuinely novel and useful construct, and the theoretical analysis of logistic and SVM depths is a solid contribution. The paper is not circular: the depth is defined as a risk, but the link to halfspace depth is proved against an external object, and the convergence results are standard generalization bounds. However, the advertised exact identification with halfspace depth is false for atomic distributions, which are exactly the empirical distributions used throughout the experiments; the framework itself can be repaired, but the stated result needs correction.

major comments (1)
  1. [Section 2, Proposition 2.2] The equality in Eq. (2) is false for atomic distributions. With L0-1(f(x),y)=1[y f(x)<0], a classifier with f(z)=0 incurs zero negative error, and its positive error is Q({x: f(x)<0}), an open halfspace, whereas HD in (1) counts the closed halfspace {x: <u,x> >= <u,z>}. The proof's step 'then for u = w ... the expressions (1) and (2) will be equal' silently assumes no atom of Q lies on the separating hyperplane. Counterexample: let Q be uniform on {(1,0),(-1,0),(0,1)} and z=(0,0). Then HD(z|Q)=1/3, attained for u=(1,-1)/sqrt(2), but the linear classifier f(x,y)=y has zero positive errors and f(z)=0, so the right-hand side of (2) is 0. Since the paper's statistical theory and experiments work with the empirical distribution Qn, which is atomic, the advertised interpretation of halfspace depth as an exact risk is not valid in the finite-sample setting. Please repair the statement, e.g. by assuming Q atomless, by replacing the closed halfspace in (1) with an open one, or by introducing a deterministic tie-break in the loss so that boundary points are counted consistently, and discuss the finite-sample status.
minor comments (6)
  1. [Section 5, logistic regression definition] The logistic loss is written as '(1 + e^{-y f(x)})'; the logarithm is missing and it should be log(1 + e^{-y f(x)}).
  2. [Proposition 5.2 proof] For the logistic loss, l_+(0) = l_-(0) = log 2, so the value of the objective at w=0 is log 2, not 1; the conclusion that the depth lies in [0,1] still holds because log 2 < 1.
  3. [Lemma 4.1 proof] The displayed equality '... = E_{Q_n} l_+(f*_epsilon(x)) - E_Q l_+(f*_epsilon(x)) + epsilon' drops the factor 1/2 that appears in the definition of Gamma; the bound in the lemma's statement is still valid because omitting the factor only makes the right-hand side larger.
  4. [Appendix A.5, proof of Lemma 4.2] The expression '|lambda(||w*_lambda||^2 - ||w*_lambda||^2)|' should read '|lambda(||hat w*_lambda||^2 - ||w*_lambda||^2)|'; as written, the difference is identically zero and the subsequent bound does not follow.
  5. [Theorem 5.6] The first rate is written with 'sup_z' but the second is not; both rates are presumably uniform over z, and the statement should be made consistent.
  6. [Abstract and conclusion] The abstract and conclusion state dimension-independent convergence rates without qualification; the rates in Theorem 5.6 are conditional on Assumptions B and W (bounded features and weights), and this caveat should be carried more prominently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the loss-depth construction is definitional, but the halfspace-depth equivalence and the convergence rates are proven against independent quantities.

full rationale

The paper's central identity, Proposition 2.2, is not circular. It proves that halfspace depth HD(z|Q), defined by the independent geometric expression in (1), equals the infimum of an expected 0-1 loss over linear classifiers on the constructed distribution P_{Q+|z-}. The risk formulation is derived from the halfspace-depth definition rather than assumed. Definition 3.1 then introduces 'loss depths' directly as risks, but this is an explicit construction of a new family, not a claim that an independently defined quantity is predicted by itself. The statistical rates in Theorem 5.6 follow from Lemma 4.1/4.2 combined with standard Rademacher complexity bounds and stated assumptions (B and W); no parameter is fitted to the depth value being predicted. Assumptions B and W are conditions on the distribution or weight set, not fitted outcomes relabeled as predictions. The self-citations in the manuscript (e.g., Castellanos et al. 2023, Mozharovskyi et al. 2020, Mozharovskyi and Valla 2022) appear only as contextual references for applications or prior extensions and do not carry the load of the main derivation. The only substantive concern raised by the reader's take, the closed-versus-open halfspace boundary issue in Proposition 2.2 for atomic distributions, is a correctness or technical-condition matter, not a circularity, and therefore does not affect the circularity score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The theoretical results rest on standard learning-theoretic tools and on two explicit regularity assumptions (bounded feature norm, bounded weights). The experimental hyperparameters lambda and gamma are data-dependent but do not enter the theoretical claims.

free parameters (2)
  • lambda (L2 regularization weight) = 1 (experiments); also 0.01 for MLP depth
    Chosen by hand; controls the complexity of the classifier and hence the depth values. Not fitted to the depth target, but affects all experimental results.
  • gamma (kernel bandwidth for SVM D and GPD) = 1 (contamination experiment); data-dependent heuristics otherwise
    For the main experiment gamma=1; in general gamma is set to the inverse median squared distance (SVM) or half the inverse first-quartile squared distance (GPD). This data-dependent choice influences depth values and rankings.
assumptions (4)
  • standard math RKHS theory and Rademacher complexity bounds (Kakade et al. 2008, Bartlett-Mendelson 2002, Steinwart-Christmann 2008)
    Used in the proof of Theorem 5.6 for the generalisation error and the norm-convergence bound.
  • domain assumption Assumption B: all x in support of Q satisfy ||phi(x)|| <= B
    Required for the Rademacher complexity bound and the uniform bounded loss constant C in the proof of Theorem 5.6.
  • domain assumption Assumption W: all optimal weights satisfy ||w|| <= W
    Required for the Rademacher bound; for regularised depths it is implied by Prop 5.1 via the candidate w=0, but for unregularised LRD it is an explicit extra condition.
  • standard math Convexity and 1-Lipschitzness of the logistic and hinge losses
    Used to apply the vector-contraction inequality for Rademacher complexities and to prove strong convexity and smoothness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data Depth as a Risk." pith.science (2026). https://pith.science/paper/OGUDTLX3

@misc{pith2026250708518,
  author       = {Pith},
  title        = {Pith review of: Data Depth as a Risk},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OGUDTLX3}},
  note         = {Machine review of arXiv:2507.08518}
}
read the original abstract

Data depths are score functions that quantify in an unsupervised fashion how central is a point inside a distribution, with numerous applications such as anomaly detection, multivariate or functional data analysis, arising across various fields. The halfspace depth was the first depth to aim at generalising the notion of quantile beyond the univariate case. Among the existing variety of depth definitions, it remains one of the most used notions of data depth. Taking a different angle from the quantile point of view, we show that the halfspace depth can also be regarded as the minimum loss of a set of classifiers for a specific labelling of the points. By changing the loss or the set of classifiers considered, this new angle naturally leads to a family of "loss depths", extending to well-studied classifiers such as, e.g., SVM or logistic regression, among others. This framework directly inherits computational efficiency of existing machine learning algorithms as well as their fast statistical convergence rates, and opens the data depth realm to the high-dimensional setting. Furthermore, the new loss depths highlight a connection between the dataset and the right amount of complexity or simplicity of the classifiers. The simplicity of classifiers as well as the interpretation as a risk makes our new kind of data depth easy to explain, yet efficient for anomaly detection, as is shown by experiments.

Figures

Figures reproduced from arXiv: 2507.08518 by the authors.

Figure 1
Figure 1. Equivalence between halfspace depth and the zero-one loss on [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Loss depths on a bigaussian dataset with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Score heatmaps with top 5 quantile zones of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Outstanding samples for the ‘sandal’ class of Fashion-MNIST on the left side and on the right side the regression coefficients used to compute their Logistic Regression Depth LRD [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Rank correlations w.r.t. the true density for a bigaussian distribution (n = 200) [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 25 canonical work pages

  1. [1]

    Latent Space Autoregression for Novelty Detection

    Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. AND: autoregressive novelty detectors. CoRR, abs/1807.01653, 2018. URL http://arxiv.org/abs/1807.01653

  2. [2]

    Theory of reproducing kernels

    Nachman Aronszajn. Theory of reproducing kernels. Transactions of the American mathematical society, 68 0 (3): 0 337--404, 1950

  3. [3]

    Learning theory from first principles

    Francis Bach. Learning theory from first principles. MIT press, 2024

  4. [4]

    Bartlett and Shahar Mendelson

    Peter L. Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk bounds and structural results. J. Mach. Learn. Res., 3: 0 463--482, 2002. URL https://jmlr.org/papers/v3/bartlett02a.html

  5. [5]

    Classification-based anomaly detection for general data

    Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net, 2020. URL https://openreview.net/forum?id=H1lK\_lBtvS

  6. [6]

    Reproducing kernel Hilbert spaces in probability and statistics

    Alain Berlinet and Christine Thomas-Agnan. Reproducing kernel Hilbert spaces in probability and statistics. Springer Science & Business Media, 2011

  7. [7]

    Christopher M. Bishop. Pattern recognition and machine learning, 5th Edition. Information science and statistics. Springer, 2007. ISBN 9780387310732. URL https://www.worldcat.org/oclc/71008143

  8. [8]

    Random rotation ensembles

    Rico Blaser and Piotr Fryzlewicz. Random rotation ensembles. Journal of Machine Learning Research, 17 0 (4): 0 1--26, 2016. URL http://jmlr.org/papers/v17/blaser16a.html

Show all 47 references
  1. [9]

    Convex Optimization

    Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004

  2. [10]

    Breunig, Hans-Peter Kriegel, Raymond T

    Markus M. Breunig, Hans-Peter Kriegel, Raymond T. Ng, and J\" o rg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, SIGMOD '00, page 93–104, New York, NY, USA, 2000. Association for Com...

  3. [11]

    Fast kernel half-space depth for data with non-convex supports

    Arturo Castellanos, Pavlo Mozharovskyi, Florence d'Alch \'e Buc, and Hicham Janati. Fast kernel half-space depth for data with non-convex supports. arXiv preprint arXiv:2312.14136, 2023

  4. [12]

    Beyond mahalanobis distance for textual ood detection

    Pierre Colombo, Eduardo Dadalto, Guillaume Staerman, Nathan Noiry, and Pablo Piantanida. Beyond mahalanobis distance for textual ood detection. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, vol...

  5. [13]

    Handbook of convergence theorems for (stochastic) gradient methods

    Guillaume Garrigos and Robert M Gower. Handbook of convergence theorems for (stochastic) gradient methods. arXiv preprint arXiv:2301.11235, 2023

  6. [14]

    Deep anomaly detection using geometric transformations

    Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. URL...

  7. [15]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Sch \"o lkopf, and Alexander Smola. A kernel two-sample test. Journal of Machine Learning Research, 13 0 (25): 0 723--773, 2012. URL http://jmlr.org/papers/v13/gretton12a.html

  8. [16]

    Faster algorithms for structured linear and kernel support vector machines

    Yuzhou Gu, Zhao Song, and Lichen Zhang. Faster algorithms for structured linear and kernel support vector machines. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net, 2025. URL https://openreview.ne...

  9. [17]

    Hadsell, S

    R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735--1742, 2006. doi:10.1109/CVPR.2006.100

  10. [18]

    On the complexity of linear prediction: Risk bounds, margin bounds, and regularization

    Sham M Kakade, Karthik Sridharan, and Ambuj Tewari. On the complexity of linear prediction: Risk bounds, margin bounds, and regularization. Advances in neural information processing systems, 21, 2008

  11. [19]

    A new measure of rank correlation

    Maurice G Kendall. A new measure of rank correlation. Biometrika, 30 0 (1-2): 0 81--93, 1938

  12. [20]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Yoshua Bengio and Yann LeCun, editors, 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014. URL http://arxiv.org/a...

  13. [21]

    Mnist handwritten digit database

    Yann LeCun and Corinna Cortes. Mnist handwritten digit database

  14. [22]

    Model compression for deep neural networks: A survey

    Zhuo Li, Hengyi Li, and Lin Meng. Model compression for deep neural networks: A survey. Computers, 12 0 (3), 2023. ISSN 2073-431X. doi:10.3390/computers12030060. URL https://www.mdpi.com/2073-431X/12/3/60

  15. [23]

    Risk bounds and calibration for a smart predict-then-optimize method

    Heyuan Liu and Paul Grigas. Risk bounds and calibration for a smart predict-then-optimize method. In Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS '21, Red Hook, NY, USA, 2021. Curran Associates Inc. ISBN 9781713845393

  16. [24]

    Liu and Kesar Singh and

    Regina Y. Liu and Kesar Singh and. A quality index based on data depth and multivariate rank tests. Journal of the American Statistical Association, 88 0 (421): 0 252--260, 1993. doi:10.1080/01621459.1993.10594317. URL https://doi.org/10.1080/01621459.1993.10594317

  17. [25]

    Beyond least-squares: Fast rates for regularized empirical risk minimization through self-concordance

    Ulysse Marteau - Ferey, Dmitrii Ostrovskii, Francis Bach, and Alessandro Rudi. Beyond least-squares: Fast rates for regularized empirical risk minimization through self-concordance. In Alina Beygelzimer and Daniel Hsu, editors, Proceedings of the Thirty-Second Conference on Le...

  18. [26]

    A vector-contraction inequality for rademacher complexities

    Andreas Maurer. A vector-contraction inequality for rademacher complexities. In Ronald Ortner, Hans Ulrich Simon, and Sandra Zilles, editors, Algorithmic Learning Theory - 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings , volume 9925 of L...

  19. [27]

    Anomaly detection using data depth: multivariate case

    Pavlo Mozharovskyi and Romain Valla. Anomaly detection using data depth: multivariate case. arXiv preprint arXiv:2210.02851, 2022

  20. [28]

    Nonparametric imputation by data depth

    Pavlo Mozharovskyi, Julie Josse, and François Husson and. Nonparametric imputation by data depth. Journal of the American Statistical Association, 115 0 (529): 0 241--253, 2020. doi:10.1080/01621459.2018.1543123. URL https://doi.org/10.1080/01621459.2018.1543123

  21. [29]

    Combining statistical depth and fermat distance for uncertainty quantification

    Hai-Vy Nguyen, Fabrice Gamboa, Reda Chhaibi, Sixin Zhang, Serge Gratton, and Thierry Giaccone. Combining statistical depth and fermat distance for uncertainty quantification. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advance...

  22. [30]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in P ython. Journal of Machine Learnin...

  23. [31]

    Ocgan: One-class novelty detection using gans with constrained latent representations

    Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocgan: One-class novelty detection using gans with constrained latent representations. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2893--2901, 2019. doi:10.1109/CVPR.2019.00301

  24. [32]

    A halfspace-mass depth-based method for adversarial attack detection

    Marine Picot, Federica Granese, Guillaume Staerman, Marco Romanelli, Francisco Messina, Pablo Piantanida, and Pierre Colombo. A halfspace-mass depth-based method for adversarial attack detection. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://open...

  25. [33]

    Outlier detection datasets (odds) library, 2016

    Shebuti Rayana. Outlier detection datasets (odds) library, 2016. URL https://odds.cs.stonybrook.edu

  26. [34]

    Reid and Robert C

    Mark D. Reid and Robert C. Williamson. Information, divergence and risk for binary experiments. Journal of Machine Learning Research, 12 0 (22): 0 731--817, 2011. URL http://jmlr.org/papers/v12/reid11a.html

  27. [35]

    Deep one-class classification

    Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel M \"u ller, and Marius Kloft. Deep one-class classification. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine ...

  28. [36]

    Waldstein, Ursula Schmidt - Erfurth, and Georg Langs

    Thomas Schlegl, Philipp Seeb \" o ck, Sebastian M. Waldstein, Ursula Schmidt - Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In Marc Niethammer, Martin Styner, Stephen R. Aylward, Hongtu Zhu, Ipek Oguz,...

  29. [37]

    Platt, John Shawe-Taylor, Alex J

    Bernhard Schölkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the support of a high-dimensional distribution. Neural Computation, 13 0 (7): 0 1443--1471, 2001. doi:10.1162/089976601750264965

  30. [38]

    Spearman

    C. Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15 0 (1): 0 72--101, 1904. ISSN 00029556. URL http://www.jstor.org/stable/1412159

  31. [39]

    Support Vector Machines

    Ingo Steinwart and Andreas Christmann. Support Vector Machines. Information science and statistics. Springer, 2008. ISBN 978-0-387-77241-7

  32. [40]

    Mathematics and the picturing of data

    John W Tukey. Mathematics and the picturing of data. In Proceedings of the International Congress of Mathematicians, Vancouver, 1975, volume 2, pages 523--531, 1975

  33. [41]

    Conditional image generation with pixelcnn decoders

    Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, koray kavukcuoglu, Oriol Vinyals, and Alex Graves. Conditional image generation with pixelcnn decoders. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing System...

  34. [42]

    V. N. Vapnik and A. Ya. Chervonenkis. The uniform convergence of frequencies of the appearance of events to their probabilities. Dokl. Akad. Nauk SSSR, 181: 0 781, 1968

  35. [43]

    Optimal transport: Old and New , volume 338

    C \'e dric Villani. Optimal transport: Old and New , volume 338. Springer, 2009

  36. [44]

    One-class anomaly detection via novelty normalization

    Jhih-Ciang Wu, Sherman Lu, Chiou-Shann Fuh, and Tyng-Luh Liu. One-class anomaly detection via novelty normalization. Computer Vision and Image Understanding, 210: 0 103226, 2021. ISSN 1077-3142. doi:https://doi.org/10.1016/j.cviu.2021.103226. URL https://www.sciencedirect.com/...

  37. [45]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017

    Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017

  38. [46]

    Deep structured energy based models for anomaly detection

    Shuangfei Zhai, Yu Cheng, Weining Lu, and Zhongfei Zhang. Deep structured energy based models for anomaly detection. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Mac...

  39. [47]

    Deep autoencoding gaussian mixture model for unsupervised anomaly detection

    Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Dae - ki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, Apri...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.