Pith. sign in

REVIEW 4 major objections 6 minor 22 references

The EE-Classifier: A classification method for functional data based on extremality indexes

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that extremality indexes—how often a curve lies above or below other curves—can replace depth as the basis for classifying functional data, with accuracy competitive against the DDG-classifier and around 60% out-of-sample…

desk verdict A reasonable depth-classifier analog whose empirical claims rest on an underspecified cross-validation protocol; fix that and the math slip, and it's a publishable incremental contribution. read the letter →

arxiv 2411.14999 v1 pith:UUYBA4QF submitted 2024-11-22 stat.ME math.FA

Catalina Lesmes , Francisco Zuluaga , Henry Laniado , Andres Gomez , Andrea Carvajal This is my paper · ORCID
classification stat.MEmath.FA MSC 62R1062H30
keywords functionaldatasupervisedclassificationextremalityepigraphindexhypographEE-plotdepthS&P500
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 extremality indexes can do for functional-data classification what depth has already been shown to do. The proposed EE-classifier assigns to each curve a pair of numbers—its modified epigraph index or modified hypograph index computed with each of two reference groups as the reference—and plots those points in a two-dimensional EE-plot; standard classifiers such as LDA, QDA, kNN, SVM, and random forests then separate the groups in that planar space. The authors argue that the visual patterns in the EE-plot encode distributional similarity, with a diagonal cluster for identical distributions, chromosome-like shapes for overlapping sinusoids, and L-shapes for non-overlapping curves. They report accuracy on six synthetic experiments and three real functional datasets that is competitive with the depth-based DDG-classifier, often with tighter interquartile spreads, and an application to S&P 500 direction prediction with reported out-of-sample accuracies ranging from about 0.43 to 0.71 across configurations.

What carries the argument

The central objects are the modified epigraph index and modified hypograph index. For curves $x_1,\dots,x_n$ on an interval $I$, the modified hypograph index $MHI(x) = \frac{1}{n\lambda(I)}\sum_{i=1}^n \lambda\{t\in I: x_i(t)\le x(t)\}$ measures the average fraction of time other sample curves fall at or below $x$, and $MEI(x) = \frac{1}{n\lambda(I)}\sum_{i=1}^n \lambda\{t\in I: x_i(t)\ge x(t)\}$ measures the average fraction of time they fall at or above $x$. Each index lies in $[0,1]$, and the two are linearly dependent, $MHI(x)-MEI(x)=1/n$. The EE-classifier computes one of these indexes twice for every curve—once with the first group as reference and once with the second—to produce the two-dimensional EE-plot, and that reduction to a planar space is what lets ordinary classifiers such as LDA, QDA, kNN, SVM, and random forests be applied directly.

What would settle it

Re-run the synthetic and real experiments under a strict protocol in which the modified epigraph and hypograph indexes are recomputed using only the curves in the current training fold, with test curves scored against that same training reference, and compare the resulting accuracy distributions with those reported. A concrete check is to compute test-curve index values once against the training curves only and once against the pooled training-plus-test curves; if the second version systematically raises accuracy, the reported advantage is driven by information leakage.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the ordering information contained in the modified epigraph index and modified hypograph index is sufficient for supervised classification of functional data, without relying on a notion of depth. For two groups with distributions F and G, every curve is mapped to a point such as (MEI_F(x), MEI_G(x)) or (MHI_F(x), MHI_G(x)); because the two indexes are linearly dependent (MHI(x) − MEI(x) = 1/n for a sample of size n), the epigraph and hypograph plots are mirrored versions of each other. The resulting EE-plot turns functional classification into a planar classification problem, and the paper finds that off-the-shelf classifiers achieve high accuracy in controlled experiments using curves with different centerlines, amplitudes, and dispersions, as well as on the Berkeley growth, Tecator, and MCO datasets. The comparison with the DDG-classifier is meant to show that extremality indexes are a practical alternative, matching the benchmark on most tasks and often showing smaller spreads in cross-validation accuracy.

Load-bearing premise

The method assumes that, during cross-validation and testing, each curve's modified epigraph or hypograph index is computed relative only to the training reference sample; the paper does not explicitly state this, and computing the indexes on the pooled train-plus-test sample would leak test information into the features and inflate the reported accuracies.

Editorial extensions

If this is right

  • Functional-data classification can be built on extremality alone: no depth notion is required, and the input to the classifier is just a pair of index values per curve.
  • Because the EE-plot lives in two dimensions, any off-the-shelf classifier can be applied, so the method inherits the flexibility of planar classification without needing functional versions of those algorithms.
  • The linear dependence between the two indexes means the epigraph and hypograph constructions carry the same information up to reflection, and the empirical results support using either, with MHI occasionally giving better accuracy.
  • The experiments give concrete guidance: LDA is especially effective for non-overlapping groups, while QDA is generally the strongest overall performer among the tested classifiers.
  • For the S&P 500 direction task, the best EE-classifier test configuration (random forest on the MEI plot) correctly identifies 10 of the 14 up months in the 2022–2023 test window, indicating that the method can extract some signal from heavily overlapping stock-value curves.

Reading between the lines

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

  • Beyond the paper: the reported EE-plot shape patterns—diagonal, chromosome-like, and L-shaped—suggest the same construction could serve as a simple visual homogeneity diagnostic for two functional samples, complementing existing rank tests.
  • Beyond the paper: because MHI and MEI differ by the constant $1/n$, the two-dimensional EE-plot is essentially a one-dimensional signal plus its reflection; a natural extension is to test whether a single coordinate based on the diagonal or the difference between the indexes suffices for classification.
  • Beyond the paper: the paper develops only the two-class case; a strict multiclass generalization would need a strategy such as one-vs-rest index pairs, and the current results give no direct evidence about how accuracy would degrade in that setting.
  • Beyond the paper: the S&P 500 test window is short, and the paper itself notes that expanding the training history improved metrics; an external check would be to benchmark the EE-classifier against a majority-class or no-change predictor on the same months, since directional stock prediction is difficult to beat.
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 / 6 minor

Summary. The manuscript proposes a supervised classifier for functional data based on the modified epigraph index (MEI) and modified hypograph index (MHI). The method, called the EE-classifier, maps each curve to a two-dimensional space whose coordinates are the curve's extremality indexes with respect to the two class distributions, producing an 'EE-plot', and then applies standard classifiers (LDA, QDA, kNN, SVM, random forest) in that space. The authors compare the EE-classifier with the depth-based DDG-classifier on six simulated experiments and three real functional data sets, and apply the method to predict monthly up/down movements of the S&P 500 index. The central claim is that this extremality-based transformation yields competitive and robust classification accuracy for functional data.

Significance. If the empirical evaluation is valid, the EE-classifier is a simple and interpretable alternative to depth-based functional classifiers: it reuses existing extremality indexes, introduces no additional fitted parameters in the feature construction, and provides a useful visual diagnostic in the EE-plot. The paper also offers a systematic benchmark against the DDG-classifier across multiple synthetic and real settings, which is valuable for practitioners. However, the empirical claims rest on an incompletely specified cross-validation and testing protocol, and there are internal inconsistencies in the reported S&P 500 results; these issues need to be resolved before the central claim can be accepted.

major comments (4)
  1. [Section 3.1, Eq. (10)] The stated relation MHI(x) - MEI(x) = 1/n is incorrect. From Eq. (10), for almost every t, I(xi(t) <= x(t)) + I(xi(t) >= x(t)) = 1, so summing over i gives MHI(x) + MEI(x) = 1 (up to measure-zero tie sets), not a constant difference. The mirrored appearance of the EEE and EEH plots follows from the sum relation. This error should be corrected, and the sentence concluding that the indexes are 'linearly dependent' should be rephrased accordingly.
  2. [Section 4.1 and Eq. (12)] The evaluation protocol does not specify the reference sample used to compute MHI and MEI for test and cross-validation curves. Equation (12) defines the EE-plot using x in X union Y, which is natural for visualization, but if a test curve is included in the reference empirical distribution used to compute its own MHI or MEI, the feature is mechanically biased toward the reference class and the reported accuracies in Figures 4, 6, 8, 10, and 12 and Table 2 are optimistically biased. The authors must state explicitly that, during cross-validation and testing, each curve's indexes are computed with respect to class-specific training reference samples only. Without this clarification, the central empirical claim is not interpretable.
  3. [Section 4.1, cross-validation description] The phrase '100-Fold cross-validation as train control' is not a standard protocol and is undefined. The authors should specify how the folds are constructed, whether the 80% training sample is further split into folds, how the 20% test sample is held out, and how hyperparameters for kNN, SVM, and random forest are selected. This information is necessary for reproducibility and for interpreting the boxplots in Figures 4, 6, 8, 10, and 12.
  4. [Section 5, Table 2] The reported S&P 500 test results are internally inconsistent. For a 14-month test period, an accuracy of 0.71 for RF/MEI corresponds to 10 correct predictions, which matches the statement that 10 of 14 increasing months were correctly predicted; however, the text also describes the accuracy as '~60%', and Table 2 reports 0.64 for LDA/MEI, not 0.60. Additionally, the phrase 'out of the 14 months in which the market increases' should be clarified: is the test set 14 months total, or are there 14 increasing months within a larger test set? These numbers need to be reconciled.
minor comments (6)
  1. [Section 2.2, Eq. (8)] The definitions of hypograph and epigraph in Eq. (8) are reversed relative to the standard conventions and to the text immediately preceding them: the display writes hyp(x) = {y > x(t)} and epi(x) = {y <= x(t)}, while the text and Eq. (9) use the conventional definitions. This display should be corrected.
  2. [Table 1] The centerline expressions 'sin(pi)' and 'x2' need clarification: as written, sin(pi) is identically zero, and x2 should be written as x^2 or x(t)^2. The domain of t should also be stated explicitly.
  3. [Section 4.1] The description of the synthetic data generation is incomplete: 'generate gauss fdata' and 'exp cov' are not fully documented, and the paper does not specify how the 100 data points per curve are obtained from the covariance operator. A brief algorithmic description or a reference to the R package documentation would improve reproducibility.
  4. [Figures 11(b) and 11(c)] The axis labels in Figure 11(b) and 11(c) both read 'MEI G1' and 'MEI G2', although panel (c) is described as the EEH plot. Please check the axis labels in all EEH plots for consistent labeling.
  5. [Section 5] The sentence 'we anticipate a definite market increase for the years 2021-2023' is not supported by the probabilistic classifier outputs and appears to be a forward-looking assertion. It should be replaced with a factual statement about the test-period predictions.
  6. [Throughout] There are several typos and inconsistent notations, e.g., 'coned' for 'coined', 'mithochondiral' for 'mitochondrial', and 'Wavelenghts' in Figure 7. A careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EE-classifier's features are fixed transformations of the functional sample, and the reported accuracies are empirical results benchmarked against the DDG-classifier.

full rationale

The paper's derivation chain is not circular. The modified epigraph and hypograph indexes in Eq. (10) are defined directly from the sample curves and Lebesgue measure; they contain no fitted parameters and no dependence on the class labels. The EE-plot in Eqs. (11)-(12) is a deterministic transformation of those indexes, and the EE-classifier in Section 3.2 applies standard classifiers (LDA, QDA, kNN, SVM, RF) to the resulting R^2 coordinates. The central claim is empirical: the reported accuracies in Figures 4, 6, 8, 10, 12 and Table 2 are measured against held-out folds or a temporal test sample, with the DDG-classifier serving as an external benchmark. Citation [13] is by two co-authors but is used only as background for visualization tools, not as the justification for the EE-classifier's validity, so it is not load-bearing. The S&P 'anticipation' sentence is an overstatement, and Section 4.1's silence on whether the indexes are recomputed on training folds is a potential data-leakage risk, but neither is a circular reduction: no equation in the paper defines the claimed result in terms of itself, and no fitted parameter is later renamed as a prediction. Accordingly, no circular steps are identified.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method introduces no fitted constants or new mathematical objects; its inputs are the MHI/MEI features (from prior literature) and the chosen downstream classifiers. The advertised novelty is the combination, not a new entity.

free parameters (3)
  • k in kNN classifier = not reported
    The paper applies kNN in the EE-plot space but never states the number of neighbors, which affects accuracy results.
  • SVM cost and kernel parameters = not reported
    SVM requires hyperparameters; the paper does not report them.
  • Number of trees in Random Forest = not reported
    Random Forest performance depends on the number of trees; the paper does not report it.
assumptions (3)
  • standard math Sample paths are continuous functions on a compact interval
    Section 2.2 defines MHI and MEI on C(I) with compact I; all datasets are discretized approximations.
  • domain assumption Modified epigraph and hypograph indexes, computed on a reference sample, capture class-discriminative ordering
    Section 3 motivates the EE-plot by assuming these indexes 'provide information about distribution's similarities if one were to compare them'.
  • ad hoc to paper The pattern characterizations of the EE-plot (45-degree line, chromosome-like, L-shape) described from figures are stable across datasets
    Section 3.1 asserts these visual patterns without proof; classification relies on them only informally.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The EE-Classifier: A classification method for functional data based on extremality indexes." pith.science (2026). https://pith.science/paper/UUYBA4QF

@misc{pith2026241114999,
  author       = {Pith},
  title        = {Pith review of: The EE-Classifier: A classification method for functional data based on extremality indexes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UUYBA4QF}},
  note         = {Machine review of arXiv:2411.14999}
}
read the original abstract

Functional data analysis has gained significant attention due to its wide applicability. This research explores the extension of statistical analysis methods for functional data, with a primary focus on supervised classification techniques. It provides a review on the existing depth-based methods used in functional data samples. Building on this foundation, it introduces an extremality-based approach, which takes the modified epigraph and hypograph indexes properties as classification techniques. To demonstrate the effectiveness of the classifier, it is applied to both real-world and synthetic data sets. The results show its efficacy in accurately classifying functional data. Additionally, the classifier is used to analyze the fluctuations in the S\&P 500 stock value. This research contributes to the field of functional data analysis by introducing a new extremality-based classifier. The successful application to various data sets shows its potential for supervised classification tasks and provides valuable insights into financial data analysis.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

  1. [1]

    Psychometrika 47, 379–396 (1982) https://doi.org/10.1007/BF02293704

    Ramsay, J.O.: When the data are functions. Psychometrika 47, 379–396 (1982) https://doi.org/10.1007/BF02293704

    work page doi:10.1007/bf02293704 1982

  2. [2]

    Fang., Wang., J.-L.: Functional data analysis for spar se longitudinal data

    Yao, H.-G.M. Fang., Wang., J.-L.: Functional data analysis for spar se longitudinal data. Journal of the American Statistical Association 100(470), 577–90 (2005)

    work page 2005

  3. [3]

    Journal of the Royal Statistical Society

    James, G.M., Hastie, T.J.: Functional linear discriminant analysis for irregularly sampled curves. Journal of the Royal Statistical Society. Series B (Statistical Methodology) 63, 533–550 (2001)

    work page 2001

  4. [4]

    Li, P.L., Chiou, J.M.: Identifying cluster number for subspace proj ected functional data clustering. Comput. Stat. Data Anal. 55, 2090–2103 (2011) https://doi.org/ 10.1016/j.csda.2011.01.001

    work page doi:10.1016/j.csda.2011.01.001 2011

  5. [5]

    Springer, New York (2012)

    Horv´ ath, L., Kokoszka, P.: Inference for Functional Data wit h Applications. Springer, New York (2012)

    work page 2012

  6. [6]

    Chemo- metrics and Intelligent Laboratory Systems 219 (2021) https://doi.org/10.1016/ 19 j.chemolab.2021.104420

    Calle-Saldarriaga, A., Laniado, H., Zuluaga, F., Leiva, V.: Homogene ity tests for functional data based on depth-depth plots with chemical applicat ions. Chemo- metrics and Intelligent Laboratory Systems 219 (2021) https://doi.org/10.1016/ 19 j.chemolab.2021.104420

    arXiv 2021

  7. [7]

    Pintado, S., Romo, J.: Depth-based classification for functional data 72 (2005)

    work page 2005

  8. [8]

    Journal of the American Statist ical Association 107(498), 737–753 (2012) https://doi.org/10.1080/01621459.2012.688462

    Li, J., Cuesta-Albertos, J.A., Liu, R.Y.: DD-classifier: Nonparamet ric classifica- tion procedure based on dd-plot. Journal of the American Statist ical Association 107(498), 737–753 (2012) https://doi.org/10.1080/01621459.2012.688462

    arXiv 2012

Show all 22 references
  1. [9]

    TEST 26, 119–142 (2017) https://doi.org/10.1007/s11749- 016-0502-6

    Cuesta-Albertos, J.A., Febrero-Bande, M., Fuente, M.: The DDG-classifier in the functional setting. TEST 26, 119–142 (2017) https://doi.org/10.1007/s11749- 016-0502-6

    2017 doi

  2. [10]

    Recent advances in functional data analysis and related topics 14, 651–676 (2011)

    Franco Pereira, A., Lillo, R., Romo, J.: Extremality for functional data. Recent advances in functional data analysis and related topics 14, 651–676 (2011)

    2011

  3. [11]

    Biostatistics 15(4), 603–619 (2014) https://doi.org/ 10.1093/biostatistics/kxu006

    Arribas-Gil, A., Romo, J.: Shape outlier detection and visualization for func- tional data: the outliergram. Biostatistics 15(4), 603–619 (2014) https://doi.org/ 10.1093/biostatistics/kxu006

    2014 doi

  4. [12]

    Journal of Applied Statistics 43(6), 1088–1103 (2016) https://doi.org/10.1080/02664763.2015.1092108

    Martin-Barragan, B., Lillo, R.E., Romo, J.: Functional boxplots ba sed on epigraphs and hypographs. Journal of Applied Statistics 43(6), 1088–1103 (2016) https://doi.org/10.1080/02664763.2015.1092108

    2016

  5. [13]

    Advances in Data Analysis and Classification 14, 651–676 (2020) https://doi.org/10.1007/ s11634-019-00380-9

    Franco-Pereira, A.M., Lillo, R.E.: Rank tests for functional data based on the epigraph, the hypograph and associated graphical representat ions. Advances in Data Analysis and Classification 14, 651–676 (2020) https://doi.org/10.1007/ s11634-019-00380-9

    2020

  6. [14]

    Statistics and Co mputing 33, 33 (2023) https://doi.org/10.1007/s11222-023-10213-7

    Pulido, B., Franco-Pereira, A.M., Lillo, R.E.: A fast epigraph and hyp ograph- based approach for clustering functional data. Statistics and Co mputing 33, 33 (2023) https://doi.org/10.1007/s11222-023-10213-7

    2023 doi

  7. [15]

    The Annals of Statistics 27(3), 783–858 (1999) https://doi.org/10.1214/ aos/1018031260

    Liu, R.Y., J.M., P., Singh, K.: Multivariate analysis by data depth: des criptive statistics, graphics and inference, (with discussion and a rejoinde r by Liu and Singh). The Annals of Statistics 27(3), 783–858 (1999) https://doi.org/10.1214/ aos/1018031260

    1999

  8. [16]

    TEST 10(2), 419– 440 (2001)

    Fraiman, R., Muniz, G.: Trimmed means for functional data. TEST 10(2), 419– 440 (2001)

    2001

  9. [17]

    Cuevas, M. A. Febrero, Fraiman, R.: Robust estimation and clas sification for functional data via projection-based depth notions. Comput. St atist. 22(3), 481– 496 (2007)

    2007

  10. [18]

    Compu- tational Statistics & Data Analysis 55(4), 1679–1695 (2011) https://doi.org/10

    L´ opez-Pintado, S., Romo, J.: A half-region depth for function al data. Compu- tational Statistics & Data Analysis 55(4), 1679–1695 (2011) https://doi.org/10. 1016/j.csda.2010.10.024 20

    2011

  11. [19]

    Publications in child development

    Tuddenham, R.D., Snyder, M.M.: Physical growth of california boy s and girls from birth to eighteen years. Publications in child development. Unive rsity of California, Berkeley 1(2), 183–364 (1954)

    1954

  12. [20]

    Springer, New York (2006)

    Ferraty, F., Vieu, P.: Functional Data Analysis: Theory and Pra ctice. Springer, New York (2006)

    2006

  13. [21]

    American Journal Phys iology Heart Circulatori Physiology

    Ruiz–Meana, M., Garcia-Dorado, D., Pina, P., Inserte, J., Agullo, L., Soler–Soler, J.: Cariporide preserves mitochondrial proton gradient and delays atp depletion in cardiomyocytes during ischemic conditions. American Journal Phys iology Heart Circulatori Physiology. 285 (2003...

    2003

  14. [22]

    Springer, New York (2021) 21 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 MEI G1 MEI G2 Group 0 1

    James, G., Witten, D., Hastie, T., Tibshirani, R.: An Introduction t o Statistical Learning with Applications in R. Springer, New York (2021) 21 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 MEI G1 MEI G2 Group 0 1

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.