Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Multiscale Dubuc: A New Similarity Measure for Time Series

T0 review · 4 major / 4 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 This paper claims that Multiscale Dubuc Distance, a multiscale envelope-overlap measure, is a metric that matches tuned DTW's 1-NN accuracy on 95 UCR datasets while running in linear time.

desk verdict The construction and benchmark are worth a look, but the claim that MDD is a metric fails as written because Eq. 5 computes an unnormalized area, not a convex combination. read the letter →

arxiv 2411.10418 v1 pith:FELSVFJM submitted 2024-11-15 cs.LG

Mahsa Khazaei , Azim Ahmadzadeh , Krishna Rukmini Puthucode This is my paper · ORCID
classification cs.LG
keywords timeseriessimilaritymeasureDubucvariationJaccardindexintersectionoverunionmetricdynamicwarping1-NNclassification
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 introduces Multiscale Dubuc Distance (MDD), a similarity measure for equal-length time series that combines Dubuc's variation envelopes from fractal analysis with a Jaccard-style intersection-over-union ratio, integrated across multiple scales. The authors claim MDD is a metric satisfying the triangle inequality, and that a 1-NN classifier using MDD with a fixed generic set of scales achieves accuracy comparable to DTW with per-dataset optimized windows on 95 UCR datasets, while running in linear time. If the metric claim holds, MDD would be a fast, tuning-free alternative to DTW for large-scale time-series classification and retrieval. The accuracy and complexity claims are supported by the experiments; the metric proof rests on a scaling assumption that the paper's own default scale set violates.

What carries the argument

The core object is the Dubuc variation envelope pair at scale $\varepsilon$: for each time series, define the upper envelope $u_{x,\varepsilon}(t)$ as the supremum of the series' values within distance $\varepsilon$ of $t$, and the lower envelope $l_{x,\varepsilon}(t)$ as the infimum. The intersection $\cap_\varepsilon(x,y)$ and union $\cup_\varepsilon(x,y)$ are computed pointwise from the two envelopes, and their ratio $r(x,y,\varepsilon) = \cap_\varepsilon / \cup_\varepsilon$ is a Jaccard index at that scale. The similarity MDS is the area under the curve obtained by trapezoidal integration of $r$ over the scales in $E$, and $\mathrm{MDD} = 1 - \mathrm{MDS}$. The Jaccard ratio is what carries the claimed metric properties, while the area under the curve is what fuses the scales into a single number; the linear-time complexity comes from computing bounds, intersection, and union each in one pass.

What would settle it

Compute Eq. 5 for a series paired with itself under the paper's generic scale set; with $E=\{1,2,4,\ldots,64\}$ the area under the curve is 63, so $\mathrm{MDD}(x,x)=1-63=-62$, contradicting the reflexivity axiom that the metric proof claims.

Watch

Extended reading notes

Core claim

The paper's central claim is that Multiscale Dubuc Distance (MDD) is a metric—more cautiously, a pseudometric—for time series of equal length. The construction borrows Dubuc's variation envelopes: for each scale $\varepsilon$, each series is replaced by the upper and lower bounds of its values in a sliding neighborhood of radius $\varepsilon$; the overlap of the two resulting strips is normalized by their union, producing a Jaccard-style intersection ratio $r(x,y,\varepsilon)$. The similarity MDS is the trapezoidal area under the $r$-versus-$\varepsilon$ curve, and $\mathrm{MDD} = 1 - \mathrm{MDS}$. The metric property, including the triangle inequality, is argued by inheritance from the Jaccard distance at a single scale. Experimentally, on 95 UCR datasets, 1-NN with MDD using a fixed generic set of scales reaches average accuracy statistically comparable to DTW with a per-dataset tuned window, and the algorithm runs in time linear in the series length. The paper also shows that shrinking the smallest scale in $E$ makes MDD more sensitive to noise, and enlarging it suppresses noise, which on datasets like Worms yields a large accuracy gain.

Load-bearing premise

The metric proof treats the total area under the ratio curve as if it were a single Jaccard distance, which only works if the scale weights sum to 1; with the paper's default scales they sum to 63, so the distance from a series to itself would be negative.

Editorial extensions

If this is right

  • On the 95 UCR datasets tested, 1-NN with MDD using one generic scale set beats or ties DTW with a learned window on 31% of datasets and matches DTW's average accuracy, so a tuning-free MDD could serve as a drop-in DTW replacement in many pipelines.
  • Because MDD runs in $\Theta(d)$ time, similarity search and classification over very long time series become practical where DTW's quadratic cost is prohibitive.
  • The single parameter $E$ controls noise sensitivity: excluding small $\varepsilon$ values suppresses high-frequency noise, which the paper demonstrates can lift accuracy from 45% to 63% on the Worms dataset.
  • If the metric property survives, MDD can be used in metric-space indexing structures and algorithms that require the triangle inequality, not just in brute-force 1-NN.
  • MDD's envelope overlap avoids the pathological point-to-point alignments that plague DTW, because it never aligns individual time steps.

Reading between the lines

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

  • The reflexivity failure is fixable: normalizing the area under the curve by the total measure of $E$ (e.g., dividing by $\sum \Delta\varepsilon_i$) restores $\mathrm{MDD}(x,x)=0$, and the resulting quantity is a convex combination of single-scale Jaccard distances, which would likely inherit the triangle inequality if all scales share the same normalization.
  • The metric property and classification performance are logically independent; even if the metric proof collapses, MDD may remain a strong 1-NN similarity, so the two claims should be evaluated separately.
  • A natural testable extension is an automatic or learned choice of the minimal scale, which could make MDD's already competitive accuracy exceed tuned DTW on noise-heavy datasets without manual inspection.
  • Because the envelopes are defined without an aspect-ratio parameter, the same construction could be applied to other one-dimensional signals or, with a suitable neighborhood, to images and graphs, suggesting MDD-style multiscale overlap is a general dissimilarity template.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes Multiscale Dubuc Similarity (MDS), defined in Eq. (5) as the trapezoidal area under the curve of per-scale intersection ratios r(x,y,ε) from Eq. (4), and defines MDD = 1 − MDS. The authors claim that MDD is a metric (or pseudometric), that it runs in linear time, and that 1-NN classification with MDD using a generic scale set E is competitive with DTW with per-dataset optimized windows on 95 UCR datasets. They also present a case study on the Worms dataset where customizing E improves accuracy from 45% to 63%.

Significance. If the metric claim and the empirical comparisons held, MDD would be an interesting addition to the time-series distance toolbox: it is self-contained (no learned constants enter the per-scale ratios), it has a clear multiscale interpretation via Dubuc envelopes, and the paper reports a broad benchmark against EuD, DTW, LCSS, and TS-MIoU. The authors also provide source code and follow a standard 1-NN evaluation framework. However, the central theoretical contribution—that MDD is a metric—is not supported as written, and the reported Worms gain is obtained by selecting E after seeing test-set results, so the paper's headline claims need substantive repair.

major comments (4)
  1. [Sec. IV-A, Eq. (5)] The reflexivity and non-negativity claims are contradicted by the unnormalized definition of MDS. For x = y, r(x,x,ε)=1 for every ε, so MDS is the trapezoidal integral of the constant function 1 over [ε_1, ε_m]. With the experimental choice E={1,2,4,...,64} (Sec. V-A), the increments Δε_i sum to 63, giving MDS(x,x)=63 and MDD(x,x)=1−63=−62. This violates the required MDD(x,x)=0 and MDD≥0. The proof that positivity follows from non-negativity of ∩ε and ∪ε is therefore invalid: those non-negativities say nothing about whether MDS exceeds 1. The definition in Eq. (5) needs an explicit normalization (e.g., division by ε_m−ε_1), and all metric proofs and experiments must be re-examined under that normalization.
  2. [Sec. IV-A, Eq. (4)] The triangle-inequality argument is not established. The paper states that MDD inherits the triangle inequality from r because the Jaccard distance is a metric [14,17], but r in Eq. (4) is the ratio of two sums over time indices, Σ ∩ε / Σ ∪ε, not the Jaccard similarity of the two envelope regions as sets. The cited results on Jaccard distance do not automatically apply to this ratio-of-sums quantity. Moreover, even if each per-scale r were a Jaccard similarity, MDD = 1 − ∫ r dε is a metric only if the integral is a convex combination of per-scale Jaccard distances, which requires the weights Δε_i to be nonnegative and sum to 1. They do not in the current formulation, so the stated inheritance from [14,17] is unsupported.
  3. [Sec. V-B] The Worms customization result is circular and does not support the claim that MDD's single parameter is a useful noise-sensitivity gauge. The authors report that with the generic E the accuracy was 45%, and that after 'taking a closer look' they redefined E to {16,32,...,256}, boosting accuracy to 63%. This selection is made after observing test-set accuracy on the same dataset, so the 40% improvement over generic E and the 18% improvement over DTW are not out-of-sample findings. A proper evaluation would preselect E on the training set (or via nested cross-validation) and then report test accuracy.
  4. [Sec. IV-B] The linear-time claim is overstated. The pseudocode and Eq. (3) give Θ(|E|·d) per distance computation, and the paper argues that |E| is a user-fixed constant that does not scale with problem size. However, in the experimental setup E is chosen as powers of two up to α·d, so |E| = O(log d). Under that protocol the complexity is Θ(d log d), not Θ(d). If the authors intend |E| to be a true constant independent of d, that should be stated and the experimental protocol should not define E as a function of series length.
minor comments (4)
  1. [Fig. 1 caption] The caption says MDS returns a value between 0 and 1, but Eq. (5) as written is an unnormalized area and can exceed 1; the caption should be updated to match the corrected definition.
  2. [Throughout] There are several typographical errors, including 'Multisclae' in the Fig. 1 caption, 'Mulstiscale' in Sec. IV-A, and the source-code URL 'multiscale dubuc' containing a space.
  3. [Sec. IV-A] The notation in the MDD definition is inconsistent: MDD(x,y,ε) is written with a scalar ε, but the definition in Eq. (5) depends on the whole set E; the authors should write MDD(x,y,E).
  4. [Sec. V-B] The text says LCSS points were removed from the Texas Sharpshooter plot for visibility, but the corresponding accuracy and gain values are not reported; a supplementary table or histogram would make the comparison reproducible.

Circularity Check

1 steps flagged · score 2.0 of 10

MDD definition is self-contained; only the Worms customization example is post hoc parameter fitting, giving a mildly circular demonstration.

  1. fitted input called prediction [Section V-C, Worms case study]
    "Initially, with the generic set of ε values {1, 2, 4, · · ·, 256} the actual accuracy of MDD was 45%. However, once we took a closer look at the dataset, we noticed the impact of noise on 1-NN’s performance. ... Therefore, by excluding the smaller values we redefined E to be {16, 32, ...,256}. This decision boosted the accuracy of 1-NN to 63%, resulting in a 40% improvement compared to the generic E."

    The paper selects the scale set E for Worms after observing the test-set accuracy (45% with generic E), then reports the improved 63% accuracy and a 40% improvement as evidence of MDD's customizable sensitivity. Because E was chosen using the same test set whose accuracy is then reported, the improvement is an in-sample fit rather than a held-out prediction; the reported gain is partly forced by the selection rather than independently validated. The paper does not present this as a prediction of future data, but the abstract highlights datasets where customization improves performance, and this example is the supporting exhibit.

full rationale

The core MDD construction (Eqs. 3-5) is self-contained: no learned constants enter the per-scale intersection ratios, and the generic E uses a fixed α=0.4 chosen before evaluation. The comparison to DTW-with-learned-window uses the reported UCR accuracies, and the TS-MIoU self-citations are used only as baseline code/data, not as load-bearing justification for MDD's metric properties. The metric proof's triangle-inequality inheritance from Jaccard distance [14,17] is not circular, but it is incomplete: Eq. 5 computes an unnormalized trapezoidal area, so MDD(x,x)=1-63=-62 for the experimental E={1,...,64}, and 1-MDS is not a convex combination of per-scale Jaccard distances unless the Δε_i weights sum to 1. That is a mathematical derivation gap, not a reduction of the result to its own inputs. The only genuinely circular element is the Worms customization example, where E is selected after seeing test-set accuracy and the improved result is presented as evidence of MDD's tunable performance. This is a minor post hoc fitting in an illustrative case study, not the paper's central benchmark claim; accordingly the overall circularity score is low.

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

The paper introduces a new distance function, not a new physical or conceptual entity. There is no hidden particle, mediator, force, or new dimension. The central construction adds no fitted physical constants, but the proof relies on an untested normalization of the scale weights and on an algorithmic assumption for linear-time bound computation. The only data-dependent tuning is the generic alpha and the post-hoc Worms E.

free parameters (2)
  • alpha (scale cap fraction) = 0.4
    Hand-chosen in Sec. V-A. It defines the maximum epsilon as 0.4 times the series length for the generic E set.
  • custom E for Worms = {16, 32, ..., 256}
    Selected after observing the generic-E test accuracy of 45% on Worms. The resulting 63% accuracy is a post-hoc fit to test labels, not a held-out prediction.
assumptions (4)
  • standard math Per-scale r(x,y,epsilon) is a Jaccard index on interval bundles, so 1 minus r satisfies the triangle inequality.
    Invoked in Sec. IV-A to transfer the triangle inequality to MDD. It is true for a single epsilon under a product measure, but it does not justify an unnormalized weighted sum over epsilon.
  • domain assumption Time series are compared pointwise at the same time index t and have equal lengths d.
    The intersection and union definitions in Eq. 3 and the complexity claim assume aligned equal-length series. The paper excludes 11 variable-length datasets in Sec. V-A.
  • domain assumption Max and min over a sliding window of radius epsilon can be computed in O(d) per epsilon.
    The claimed Theta(d) running time in Sec. IV-B requires a monotonic-queue or similar sliding-window extremum algorithm. The pseudocode only says 'calculate-bounds' iterates once.
  • ad hoc to paper Choosing E as powers of two up to alpha times d preserves the relevant scales.
    No principled basis is given for powers of two or for alpha = 0.4; it is a heuristic introduced in Sec. V-A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multiscale Dubuc: A New Similarity Measure for Time Series." pith.science (2026). https://pith.science/paper/FELSVFJM

@misc{pith2026241110418,
  author       = {Pith},
  title        = {Pith review of: Multiscale Dubuc: A New Similarity Measure for Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FELSVFJM}},
  note         = {Machine review of arXiv:2411.10418}
}
read the original abstract

Quantifying similarities between time series in a meaningful way remains a challenge in time series analysis, despite many advances in the field. Most real-world solutions still rely on a few popular measures, such as Euclidean Distance (EuD), Longest Common Subsequence (LCSS), and Dynamic Time Warping (DTW). The strengths and weaknesses of these measures have been studied extensively, and incremental improvements have been proposed. In this study, however, we present a different similarity measure that fuses the notion of Dubuc's variation from fractal analysis with the Intersection-over-Union (IoU) measure which is widely used in object recognition (also known as the Jaccard Index). In this proof-of-concept paper, we introduce the Multiscale Dubuc Distance (MDD) measure and prove that it is a metric, possessing desirable properties such as the triangle inequality. We use 95 datasets from the UCR Time Series Classification Archive to compare MDD's performance with EuD, LCSS, and DTW. Our experiments show that MDD's overall success, without any case-specific customization, is comparable to DTW with optimized window sizes per dataset. We also highlight several datasets where MDD's performance improves significantly when its single parameter is customized. This customization serves as a powerful tool for gauging MDD's sensitivity to noise. Lastly, we show that MDD's running time is linear in the length of the time series, which is crucial for real-world applications involving very large datasets.

Figures

Figures reproduced from arXiv: 2411.10418 by the authors.

Figure 1
Figure 1. The graphic showing the idea behind the Multisclae Dubuc Similarity (MDS) measure. Two time series (xt) d t=1 and (yt) d t=1 are compared using their identified envelopes (regions with dashed borders) for ε = 2. MDS quantifies the similarity between the two time series at different granularity levels by changing ε, and computes the intersection ratio (see Eq. 4) at each step. It then aggregates the intersection rati… view at source ↗
Figure 2
Figure 2. The bar plot showing the average expected (blue) and actual (orange) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The Texas Sharpshooter plot showing the accuracy gain of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 16 canonical work pages

  1. [1]

    Ahmadzadeh, A., Chen, Y ., Puthucode, K.R., Ma, R., Angryk, R.A.: Source code for ts-miou: A time series similarity metric without mapping (2023), https://bitbucket.org/gsudmlab/ts miou ecmlpkdd22/

    work page 2023

  2. [2]

    In: Amini, M.R., Canu, S., Fischer, A., Guns, T., Kralj Novak, P., Tsoumakas, G

    Ahmadzadeh, A., Chen, Y ., Puthucode, K.R., Ma, R., Angryk, R.A.: Ts-miou: A time series similarity metric without mapping. In: Amini, M.R., Canu, S., Fischer, A., Guns, T., Kralj Novak, P., Tsoumakas, G. (eds.) Machine Learning and Knowledge Discovery in Databases. pp. 87–102. Springer Nature Switzerland, Cham (2023)

    work page 2023

  3. [3]

    In: 2021 IEEE International Conference on Image Processing (ICIP)

    Ahmadzadeh, A., Kempton, D.J., Chen, Y ., Angryk, R.A.: Multiscale iou: A metric for evaluation of salient object detection with fine structures. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 684–688 (2021). https://doi.org/10.1109/ICIP42928.2021.9506337

    arXiv 2021

  4. [4]

    In: Proceedings of the Eleventh SIAM International Conference on Data Mining, SDM 2011, April 28-30, 2011, Mesa, Arizona, USA

    Batista, G.E.A.P.A., Wang, X., Keogh, E.J.: A complexity-invariant distance measure for time series. In: Proceedings of the Eleventh SIAM International Conference on Data Mining, SDM 2011, April 28-30, 2011, Mesa, Arizona, USA. pp. 699–710. SIAM / Omnipress (2011). https://doi.org/10.1137/1.9781611972818.60

    work page doi:10.1137/1.9781611972818.60 2011

  5. [5]

    In: Proceedings of the 2005 ACM SIGMOD International Conference on Management of Data

    Chen, L., ¨Ozsu, M.T., Oria, V .: Robust and fast similarity search for moving object trajectories. In: Proceedings of the 2005 ACM SIGMOD International Conference on Management of Data. p. 491–502. SIGMOD ’05, Association for Computing Machinery, New York, NY , USA (2005). https://doi.org/10.1145/1066157.1066213

    arXiv 2005

  6. [6]

    Chen, Y ., Nascimento, M.A., Ooi, B.C., Tung, A.K.H.: SpADe: On Shape-based Pattern Detection in Streaming Time Series. pp. 786–795. IEEE Computer Society (Apr 2007). https://doi.org/10.1109/ICDE.2007.367924, iSSN: 1063-6382

    work page arXiv 2007

  7. [7]

    In: Proceedings of the 34th International Conference on Machine Learning - V olume 70

    Cuturi, M., Blondel, M.: Soft-dtw: a differentiable loss function for time-series. In: Proceedings of the 34th International Conference on Machine Learning - V olume 70. p. 894–903. ICML’17, JMLR.org (2017)

    work page 2017

  8. [8]

    IEEE CAA J

    Dau, H.A., Bagnall, A.J., Kamgar, K., Yeh, C.M., Zhu, Y ., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.J.: The UCR time series archive. IEEE CAA J. Autom. Sinica 6(6), 1293–1305 (2019). https://doi.org/10.1109/jas.2019.1911747

    arXiv 2019

Show all 26 references
  1. [9]

    Dau, H.A., Keogh, E., Kamgar, K., Yeh, C.C.M., Zhu, Y ., Gharghabi, S., Ratanamahatana, C.A., Yanping, Hu, B., Begum, N., Bagnall, A., Mueen, A., Batista, G.: The ucr time series classification archive (October 2018), https://www.cs.ucr.edu/∼eamonn/time series data 2018/

    2018

  2. [10]

    Physical review

    Dubuc, Quiniou, Roques-Carmes, Tricot, Zucker: Evaluating the fractal dimension of profiles. Physical review. A, General physics 39(3), 1500–1512 (1989). https://doi.org/10.1103/physreva.39.1500

    1989 doi

  3. [11]

    SIGMOD Rec

    Faloutsos, C., Ranganathan, M., Manolopoulos, Y .: Fast subsequence matching in time-series databases. SIGMOD Rec. 23(2), 419–429 (May 1994). https://doi.org/10.1145/191843.191925

    1994

  4. [12]

    Data Mining and Knowledge Discovery 38(4), 2520–2544 (Jul 2024)

    Foumani, N.M., Tan, C.W., Webb, G.I., Rezatofighi, H., Salehi, M.: Series2vec: similarity-based self-supervised representation learning for time series classification. Data Mining and Knowledge Discovery 38(4), 2520–2544 (Jul 2024). https://doi.org/10.1007/s10618-024-01043-w

    2024 doi

  5. [13]

    In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alch ´e-Buc, F., Fox, E., Garnett, R

    Franceschi, J.Y ., Dieuleveut, A., Jaggi, M.: Unsupervised scalable representation learning for multivariate time series. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alch ´e-Buc, F., Fox, E., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32...

    2019

  6. [14]

    Nature 239, 174–174 (1972)

    Gilbert, G.: Distance between sets. Nature 239, 174–174 (1972)

    1972

  7. [15]

    In: Proceedings of the 2001 SIAM International Conference on Data Mining

    Keogh, E.J., Pazzani, M.J.: Derivative Dynamic Time Warping. In: Proceedings of the 2001 SIAM International Conference on Data Mining. pp. 1–11. Society for Industrial and Applied Mathematics (Apr 2001). https://doi.org/10.1137/1.9781611972719.1

    2001 doi

  8. [16]

    In: Proceedings of the 2002 ACM symposium on Applied computing

    Kim, S.W., Yoon, J., Park, S., Kim, T.H.: Shape-based retrieval of similar subsequences in time-series databases. In: Proceedings of the 2002 ACM symposium on Applied computing. pp. 438–445. ACM, Madrid Spain (Mar 2002). https://doi.org/10.1145/508791.508874

    2002

  9. [17]

    Pattern Recognit

    Kosub, S.: A note on the triangle inequality for the jaccard distance. Pattern Recognit. Lett. 120, 36–38 (2019). https://doi.org/10.1016/j.patrec.2018.12.007

    2019 doi

  10. [18]

    Mandelbrot, B.B.: The fractal geometry of nature. 1982. San Francisco, CA (1982)

    1982

  11. [19]

    https://doi.org/10.5772/intechopen.81958

    Pilgrim, I., Taylor, R.P.: Fractal analysis of time-series data sets: Methods and challenges (2019). https://doi.org/10.5772/intechopen.81958

    2019 doi

  12. [20]

    IEEE Transactions on Acoustics, Speech, and Signal Processing 26(1), 43–49 (1978)

    Sakoe, H., Chiba, S.: Dynamic programming algorithm optimization for spoken word recognition. IEEE Transactions on Acoustics, Speech, and Signal Processing 26(1), 43–49 (1978). https://doi.org/10.1109/TASSP.1978.1163055

    1978

  13. [21]

    In: ˙Zytkow, J.M., Rauch, J

    Struzik, Z.R., Siebes, A.: The Haar Wavelet Transform in the Time Series Similarity Paradigm. In: ˙Zytkow, J.M., Rauch, J. (eds.) Principles of Data Mining and Knowledge Discovery. pp. 12–22. Springer, Berlin, Heidelberg (1999). https://doi.org/10.1007/978-3-540-48247-5 2

    1999 doi

  14. [22]

    Tavenard, R., Faouzi, J., Vandewiele, G., Divo, F., Androz, G., Holtz, C., Payne, M., Yurchak, R., Rußwurm, M., Kolar, K., Woods, E.: Tslearn, a machine learning toolkit for time series data. J. Mach. Learn. Res. 21, 118:1–118:6 (2020)

    2020

  15. [23]

    Journal of The Optical Society of America A-optics Image Science and Vision 7(6), 1055–1073 (1990)

    Theiler, J.: Estimating fractal dimension. Journal of The Optical Society of America A-optics Image Science and Vision 7(6), 1055–1073 (1990). https://doi.org/10.1364/JOSAA.7.001055

    1990 doi

  16. [24]

    In: Proceedings 18th International Conference on Data Engineering

    Vlachos, M., Kollios, G., Gunopulos, D.: Discovering similar multidimensional trajectories. In: Proceedings 18th International Conference on Data Engineering. pp. 673–684 (2002). https://doi.org/10.1109/ICDE.2002.994784

    2002

  17. [25]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yue, Z., Wang, Y ., Duan, J., Yang, T., Huang, C., Tong, Y ., Xu, B.: TS2Vec: Towards Universal Representation of Time Series. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8980–8987 (Jun 2022). https://doi.org/10.1609/aaai.v36i8.20881, iSSN: ...

    2022 doi

  18. [26]

    Pattern Recognition 74, 171–184 (Feb 2018)

    Zhao, J., Itti, L.: shapeDTW: Shape Dynamic Time Warping. Pattern Recognition 74, 171–184 (Feb 2018). https://doi.org/10.1016/j.patcog.2017.09.020

    2018 doi

Pith tools

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