Pith. sign in

REVIEW 3 major objections 3 minor 46 references

Panache: One-Pass Motif Discovery at Every Window Length

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read The paper claims that one streaming pass over a time series can maintain a pan matrix profile over all window lengths, writing only exact z-normalized distances to valid neighbors, so motif discovery no longer requires a separate self-join

desk verdict The one-pass streaming PMP construction is real and mostly well proved, but the headline recall claim is softer than it looks because the anchor-persistence heuristic is validated only in-sample and no code is released. read the letter →

arxiv 2607.17481 v1 pith:TQJKWTWZ submitted 2026-07-20 cs.AI cs.DBcs.LG

classification cs.AIcs.DBcs.LG
keywords timeseriesmotifdiscoverypanmatrixprofilestreamingalgorithmlocality-sensitivehashingslidingDFTz-normalizationParsevallowerbound
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 introduces Panache, an algorithm that maintains, in one pass over a streaming time series, a pan matrix profile — the nearest-neighbor distance for every window start at every window length in a chosen interval. Existing methods run one quadratic self-join per length, so L lengths cost L full rescans. Panache replaces this with a near-linear streaming scan by keeping, for each anchor length, the non-DC Fourier spectrum of the sliding window, which serves both as a locality-sensitive hash key that collides similar subsequences and as a Parseval lower bound that rejects most candidate pairs before exact comparison. The paper proves that every cell the algorithm writes is an exact distance to a valid neighbor (a sound upper pan-profile), and shows empirically that at default settings it recovers all top-20 motifs against exact ground truth on 17 configurations, faster than the CPU and GPU baselines benchmarked. If the empirical coverage claim holds in general, motif discovery over an unknown duration becomes a single pass, not a search over lengths.

What carries the argument

The central object is the normalized AC spectrum of a window — the leading non-DC Fourier coefficients divided by the window standard deviation, with the DC coefficient omitted. Because centering changes only the DC term, this vector can be updated per sample in O(K) by sliding-DFT removals, insertions, and rotations. It is used as the SimHash key (a locality-sensitive hash) under which windows of similar normalized shape collide, and its truncated squared difference gives, via Parseval's theorem, a lower bound on the exact z-normalized distance that rejects most colliding candidates. Around this object the algorithm builds anchor lengths (a stride over the length interval), a global-depth p

What would settle it

Take a time series engineered so a strong recurring shape has z-normalized distance that is minimal at exactly one non-anchor window length and much larger at every neighboring length; with the default anchor stride s=2, the algorithm may never probe that length, so a miss would confirm the coverage limitation rather than contradict it. The sharper test is to measure, on a held-out set of natural series, the distribution of the number of consecutive lengths over which each top motif remains a best match — the empirical premise on which anchoring rests.

Watch

Extended reading notes

Core claim

The central discovery is that z-normalized motif discovery over a whole interval of window lengths can be formulated as a one-pass streaming problem rather than as L independent self-joins. The load-bearing identity is that mean-centering a subsequence changes only its DC Fourier coefficient; therefore the non-DC spectrum of every z-normalized window can be maintained online by sliding-DFT recurrences and running statistics. This spectral vector serves triple duty: as the streaming state, as a SimHash key under which windows of similar normalized shape collide, and as a Parseval lower bound that rejects most colliding pairs before any exact computation. With the addition of anchor lengths, o

Load-bearing premise

The method's recall depends on the empirical assumption that low-distance motif regions are contiguous across neighboring window lengths and span at least the anchor stride, so that every motif intersects an anchor length; a motif that is best at exactly one isolated length can be silently missed.

Editorial extensions

If this is right

  • Analysts no longer need to choose a window length: the same one-pass run returns motifs across the whole interval, so unknown-duration patterns are found without a length sweep.
  • Every reported motif carries an exact distance to a valid neighbor, so downstream use of the distances is as reliable as with exact matrix-profile methods despite the approximate search.
  • Runtime is near-linear in the series length and independent of the number of window lengths (up to anchor-stride and emit constants), so doubling the length interval costs little.
  • The method is deterministic: outputs are bit-identical across thread counts because of exclusion-zone batching and a fixed merge order.
  • On the benchmarked configurations, the paper reports full top-20 recovery against exact ground truth, so the method can return the same motifs as exact methods in minutes where those baselines take hours.

Reading between the lines

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

  • Editorial inference — if the anchor-persistence assumption generalizes, the same one-pass machinery could serve as an online detector for motif drift, where the best window length of a recurring shape shifts over time; the contiguity of low regions would let a single scan track it.
  • Editorial inference — the normalized AC spectrum is a compact sketch of each window; a natural extension is to reuse that streaming state for other shape-based tasks such as anomaly detection or cross-length clustering, avoiding recomputation of windows.
  • Editorial inference — because recall at skipped lengths rests on an empirical persistence property, a stress test on synthetic series engineered to hide a motif at a single isolated length would reveal the practical boundary of the coverage guarantee; the paper itself concedes that such a low region can escape every anchor.
  • Editorial inference — the main reported limitation is memory, which grows linearly in the stream (features, signatures, profile rows, raw samples); a plausible next step is a compressed or disk-backed feature cache, which the paper mentions as future work.
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

3 major / 3 minor

Summary. This paper introduces Panache, a one-pass streaming algorithm for z-normalized pan matrix profile (PMP) motif discovery over a length interval [m_min, m_max]. Instead of computing one quadratic self-join per length, Panache maintains sliding-DFT-based spectral features for a subset of anchor lengths, uses SimHash and an occupancy-controlled extendible directory to propose candidate pairs, applies a Parseval lower bound to prune most candidates, and performs an emit stage with a hot-class sieve and local completion. The paper proves that every written profile cell holds an exact z-normalized distance to a valid neighbor (Proposition 1), that the AC spectrum is centering-invariant (Lemma 1), that the truncation bound is safe (Lemma 2), that occupancy is controlled (Proposition 4), and that the output is deterministic across thread counts (Proposition 5). Empirically, it reports recovery of all top-20 pan-motifs on 17 UCR configurations with exact ground truth, and speedups over CPU and GPU baselines, including a 5M-sample Wafer run where the pass takes 2.9 minutes and the emit stage 6.0 minutes, against 7.95 hours for the fastest exact CPU baseline.

Significance. If the claims hold, Panache is a significant algorithmic contribution: it avoids the L-fold repetition of self-joins that characterizes existing PMP methods, maintains a sound upper pan-profile with exact-valued cells, and ships a thorough parameter study showing that recovery is insensitive to design constants (Appendix E). The theoretical core is well executed: Proposition 1 cleanly separates soundness from coverage, Lemma 2's disjoint-bin argument is correct, and the determinism proof is convincing. The main weakness is that the headline recall guarantee is conditional on an empirical persistence assumption that is validated only in-sample; this limits the generality of the 'recovers all top-20' claim.

major comments (3)
  1. [§3.3, Proposition 3] The load-bearing recall guarantee is conditional on the anchor-persistence heuristic. Proposition 3 bounds the probability that no anchor discovers a pair only over M_A, the set of anchors at which the pair is already matchable; it does not bound the probability that a true motif is matchable at any anchor. The paper itself concedes: 'A low region that spans fewer than s consecutive lengths can still escape every anchor.' Since s=2 is the default, a motif whose best length is isolated at a single value can be silently missed. The persistence assumption is validated only empirically, on the same UCR datasets used for the benchmark (Figure 3, median 8 consecutive lengths), and the cited cross-length bounds [24] do not imply that the minimum over all windows stays at the same occurrence. To make the central claim robust, the paper should either state a precise persistence condition under wh
  2. [Appendix C, Lemma 4] The proof of Lemma 4 does not establish the stated threshold. The text says that the condition D ≥ 2εS/λ 'becomes' (2/m)D² ≥ (√2 m/600)² after substituting λ=10^-4, ε=2^-24, and S≤√2 m. But 2ε/λ ≈ 1.19×10^-3, so D ≥ S/839, whereas the stated condition is dimensionally inconsistent and yields a different numerical relation. Since Lemma 4 is the basis for the claim that shrinking the computed bound by (1-λ) preserves Lemma 2 'at every magnitude where the bound can prune,' this is a technical gap in a load-bearing pruning step. It does not invalidate Proposition 1 (written values remain exact), but an incorrect shrink factor could cause the algorithm to skip a pair that would have improved a profile cell, affecting recall. Please correct the derivation or present a conservative numerical bound with a complete proof.
  3. [§4, Table 1] The evaluation's recovery metric includes value matches with a tolerance of 5% relative or 0.05 absolute. This is loose enough to obscure positional misses. The paper partly addresses this in Appendix G by identifying 320 'determinate' motifs and showing 272/320 recovered at the identical start, with the rest equal-distance value matches. However, the abstract and §4.1 state 'recovers all top-20' without emphasizing that on several rows a substantial fraction of the recovered motifs are value matches rather than exact positional matches (e.g., StarLightCurves 1M: 1+19; Wafer 152k: 0+20 in Table 1). Including the determinate-motif breakdown in the main text would make the strength of the empirical claim clearer and prevent readers from inferring exact positional recovery where it does not occur.
minor comments (3)
  1. [Algorithm 2] The refinement loop uses 'A ← top cells of the current profiles' but does not define the number of cells M. Please state that M is the number of seeds and how it is chosen.
  2. [Figure 3] The caption says 'all 234 exact top-20 motifs of the benchmark (9 datasets, two scales)' while the text and Table 1 describe ten UCR datasets. Clarify whether one dataset (likely Wafer) is omitted because its top-20 motifs are all distance-zero duplicates, and reconcile the counts.
  3. [§3.2] The phrase 'the bound is therefore multiplied by 1−λ before the comparison' and the claim that Lemma 4 shows λ=10^-4 preserves validity at every magnitude where the bound can prune should be revisited in light of the Lemma 4 proof issue; at minimum, add a pointer to the corrected argument.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained, and the empirical claims are benchmarked externally.

full rationale

Panache's central formal guarantee, Proposition 1, is a definitional statement about the induced profile P^S: every finite cell is an exact distance to a valid neighbor and P^S_m[i] >= P_m[i] because the minimum is taken over a subset of the admissible comparison universe. The paper explicitly separates this soundness guarantee from coverage, and no prediction is derived from a fitted value or from the evaluated set S. The algorithm's data-dependent parameters (K, directory depth, probe count) are computed by the algorithm itself, and Appendix E's exhaustive sweeps show recovery is insensitive to every design constant, which is evidence against benchmark-tuned overfitting. The anchor-persistence assumption in §3.3 is presented as a testable heuristic with an explicit limitation ('A low region that spans fewer than s consecutive lengths can still escape every anchor'), and it is validated empirically against brute-force ground truth rather than used as an input that forces the output. Recovery claims are scored against external exact implementations (Stumpy and Scamp-GPU) on held-out ground truth, and the paper contains no load-bearing self-citations, no imported uniqueness theorem, and no renaming of known results as a new derivation. The only concerns raised in the manuscript are correctness risks about recall at skipped lengths, not circularity.

Assumptions & free parameters 9 free parameters · 8 assumptions · 0 invented entities

The paper's core guarantee — every written cell is an exact distance to a valid neighbor — is parameter-independent and does not rest on ad hoc assumptions; the listed free parameters tune cost and recall, not soundness. The one genuinely load-bearing data assumption is the persistence/contiguity of motif regions across neighboring lengths, which the anchor mechanism relies on; it is validated only in-sample. All mathematical building blocks (Parseval, sliding DFT, SimHash collision probability, extendible-hashing amortization) are standard or cited. No invented entities are introduced.

free parameters (9)
  • occupancy target α (with verification multiplier γ=8, budget R=128) = 16
    The single analyst-exposed knob; sets mean bucket occupancy and verification budget R=γα. Sweep α=4..64 recovers 740/740.
  • anchor stride s = 2
    Indexes every s-th length; skipped lengths are covered by emit-time local completion. Sweep s=1..8 recovers all motifs.
  • signature width B (SimHash bits) = 24
    12 address bits + 12 tail bits; fixed by design.
  • energy threshold η for spectral-width calibration = 0.85
    Part of the K-calibration rule that reads sampled windows of the series being searched; recovery shown insensitive to fixed K=4..16.
  • spectral width K = 8–16 (calibrated)
    Data-dependent width selected by calibration; soundness is width-independent (Prop. 2), so it tunes cost and recall, never validity.
  • sieve width w = 8
    Hot-class sorted-neighbor width in the emit-time sieve; sweep w=2..32 recovers 740/740.
  • refinement radii (Δm, Δp) = (6, 4)
    Local-completion box around seeds; radii do not affect recovery in the swept range (2,1)–(12,8).
  • exclusion radius excl (= batch size β) = 8
    Validity rule |i−j|≥excl fixed at 8 samples regardless of window length; applied identically to Panache and exact ground truth, so comparisons are internally consistent.
  • zero-snap ε0, bound-shrink λ, probe reference depth d0, refinement round cap Pmax = 1e-6, 1e-4, 12, 8
    Remaining design constants; all shown non-critical in the parameter sweeps.
assumptions (8)
  • standard math Parseval's theorem for the DFT: ||Z_i − Z_j||² = (1/m)Σ_k |Â_k − B̂_k|² (Lemma 2, Appendix C)
    Standard harmonic analysis; load-bearing for the candidate-rejection bound.
  • standard math Mean-centering changes only the DC Fourier coefficient; the geometric sum Σ e^{-2πikr/m} vanishes for k not divisible by m (Lemma 1)
    Correct as stated; verified directly.
  • standard math SimHash collision probability Pr[sign⟨g,u⟩ ≠ sign⟨g,v⟩] = θ/π for Gaussian g (Fact 1, cited to [6])
    Standard LSH fact; load-bearing for candidate generation and for Proposition 3's independence argument.
  • standard math Sliding-DFT recurrence X_{i+1,k} = e^{2πik/m}(X_{i,k} − t_i + t_{i+m}) [13,14]
    Standard; keeps the per-sample update at O(K).
  • domain assumption Persistence assumption: low pan-profile cells occupy contiguous regions of the length–position plane, spanning at least the anchor stride, so anchors intersect every motif region (Figure 3)
    The load-bearing heuristic for recall. Soundness of written cells is independent of it, but the recovered set is not. Validated only in-sample on the benchmark's 10 UCR datasets.
  • domain assumption Fixed exclusion zone excl=8 for all lengths in the interval
    A modeling choice for what counts as a trivial-match pair; applied identically to ground truth, but differs from the m/4 rule common elsewhere in the field.
  • standard math Hot-class sieve adjacency: under a Gaussian projection, the closest pair of a class lands within w sorted positions with probability bounded by Lemma 3
    Proof given in Appendix C via convexity of arcsin; the constants are plausible.
  • ad hoc to paper K-calibration rule: minimal K satisfying the energy-capture condition (η=0.85) and the collision-count expectation over sampled windows (§3.2)
    Paper-specific data-dependent rule; soundness is width-independent (Prop. 2), so it tunes cost and recall rather than validity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Panache: One-Pass Motif Discovery at Every Window Length." pith.science (2026). https://pith.science/paper/TQJKWTWZ

@misc{pith2026260717481,
  author       = {Pith},
  title        = {Pith review of: Panache: One-Pass Motif Discovery at Every Window Length},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQJKWTWZ}},
  note         = {Machine review of arXiv:2607.17481}
}
abstract

Motif discovery, the search for recurring patterns within a time series, is a core primitive of exploratory data analysis. A pattern, however, is defined by its duration, which analysts rarely know in advance. To resolve this unknown duration, an interval of window lengths is defined, and the accepted method is to try every length in that interval. Existing pan matrix profile (PMP) methods compute one z-normalized matrix profile per length, so $L$ lengths cost $L$ quadratic self-joins over the same series. We introduce Panache, to our knowledge the first one-pass streaming algorithm for z-normalized PMP motif discovery. It replaces the repeated self-joins with a single scan whose runtime is near-linear in the series length. The key observation is that mean-centering a subsequence changes only its DC Fourier coefficient, so the non-DC spectrum of every z-normalized subsequence can be maintained online by sliding-DFT recurrences and running statistics. This spectral state is the key under which similar subsequences collide in an occupancy-controlled hash directory and, through Parseval's theorem, yields a lower bound that rejects most colliding pairs before any exact computation. Panache computes every data-dependent parameter itself, leaving only a resource budget to tune. At the default budget, it recovers all top-20 pan-motifs against exact fixed-exclusion ground truth on 17 UCR configurations, and is faster than every CPU and GPU baseline benchmarked in this paper. On Wafer at five million samples over 51 lengths, Panache completes one pass in 2.9 minutes and emits the exact motifs in 6.0 minutes, against 7.95 hours for the fastest exact CPU baseline and 38.3 minutes for SCAMP on an H100 GPU.

Figures

Figures reproduced from arXiv: 2607.17481 by the authors.

Figure 1
Figure 1. Existing PMP representation on a sample of ECG200. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Panache end to end on an ECG5000 motif pair (𝑖 = 3925, 𝑗 = 7564, 𝑚 = 20). All values are computed from the data, and the left spine is the stream itself. (A) Each sample updates the retained coefficients in 𝑂(𝐾): drop 𝑡𝑖 , add 𝑡𝑖+𝑚, one fixed rotation (worked example in each lane); a completed window is normalized to 𝐹 and hashed into a 12-bit directory address (§3.1–3.4). (B) 𝐹𝑗 ≈ 𝐹𝑖 : same prefix, same bucket. (C)… view at source ↗
Figure 3
Figure 3. Empirical validation of the persistence assumption: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Runtime response to each design constant, aggre [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 10 canonical work pages

  1. [24]

    Abdullah Mueen. 2013. Enumeration of Time Series Motifs of All Lengths. In2013 IEEE 13th International Conference on Data Mining, Dallas, TX, USA, December 7- 10, 2013, Hui Xiong, George Karypis, Bhavani Thuraisingham, Diane J. Cook, and Xindong Wu (Eds.). IEEE Computer Society, 547–556. doi:10.1109/ICDM.2013.27

  2. [1]

    Rakesh Agrawal, Christos Faloutsos, and Arun N. Swami. 1993. Efficient Simi- larity Search In Sequence Databases. InFoundations of Data Organization and Algorithms, 4th International Conference, FODO’93, Chicago, Illinois, USA, October 13-15, 1993, Proceedings (Lecture Notes in Computer Science, Vol. 730), David B. Lomet (Ed.). Springer, 69–84. doi:10.100...

  3. [2]

    Nurjahan Begum and Eamonn J. Keogh. 2014. Rare Time Series Motif Discovery from Unbounded Streams.Proc. VLDB Endow.8, 2 (2014), 149–160. doi:10.14778/ 2735471.2735476

  4. [3]

    Jeremy Buhler and Martin Tompa. 2002. Finding Motifs Using Random Projec- tions.J. Comput. Biol.9, 2 (2002), 225–242. doi:10.1089/10665270252935430

  5. [4]

    Matteo Ceccarello and Johann Gamper. 2022. Fast and Scalable Mining of Time Series Motifs with Probabilistic Guarantees.Proc. VLDB Endow.15, 13 (2022), 3841–3853. doi:10.14778/3565838.3565840

  6. [5]

    Matteo Ceccarello, Francesco Pio Monaco, and Francesco Silvestri. 2025. MO- MENTI: Scalable Motif Mining in Multidimensional Time Series.Proc. VLDB Endow.18, 10 (2025), 3463–3476. doi:10.14778/3748191.3748208

  7. [6]

    Moses Charikar. 2002. Similarity estimation techniques from rounding algo- rithms. InProceedings on 34th Annual ACM Symposium on Theory of Computing, May 19-21, 2002, Montréal, Québec, Canada, John H. Reif (Ed.). ACM, 380–388. doi:10.1145/509907.509965

  8. [7]

    Keogh, and Stefano Lonardi

    Bill Yuan-chi Chiu, Eamonn J. Keogh, and Stefano Lonardi. 2003. Probabilistic discovery of time series motifs. InProceedings of the Ninth ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 24 - 27, 2003, Lise Getoor, Ted E. Senator, Pedro M. Domingos, and Christos Faloutsos (Eds.). ACM, 493–498. doi:1...

Show all 46 references
  1. [8]

    Shasha, and Xiaojian Zhao

    Richard Cole, Dennis E. Shasha, and Xiaojian Zhao. 2005. Fast window correla- tions over uncooperative time series. InProceedings of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Chicago, Illi- nois, USA, August 21-24, 2005, Robert Gr...

  2. [9]

    Bagnall, Kaveh Kamgar, Chin-Chia Michael Yeh, Yan Zhu, Shaghayegh Gharghabi, Chotirat Ann Ratanamahatana, and Eamonn J

    Hoang Anh Dau, Anthony J. Bagnall, Kaveh Kamgar, Chin-Chia Michael Yeh, Yan Zhu, Shaghayegh Gharghabi, Chotirat Ann Ratanamahatana, and Eamonn J. Keogh. 2019. The UCR time series archive.IEEE CAA J. Autom. Sinica6, 6 (2019), 1293–1305. doi:10.1109/JAS.2019.1911747

  3. [10]

    Raymond Strong

    Ronald Fagin, Jürg Nievergelt, Nicholas Pippenger, and H. Raymond Strong. 1979. Extendible Hashing - A Fast Access Method for Dynamic Files.ACM Trans. Database Syst.4, 3 (1979), 315–344. doi:10.1145/320083.320092

  4. [11]

    Ranganathan, and Yannis Manolopoulos

    Christos Faloutsos, M. Ranganathan, and Yannis Manolopoulos. 1994. Fast Subsequence Matching in Time-Series Databases. InProceedings of the 1994 ACM SIGMOD International Conference on Management of Data, Minneapolis, Minnesota, USA, May 24-27, 1994, Richard T. Snodgrass and Ma...

  5. [12]

    Yifeng Gao and Jessica Lin. 2019. HIME: discovering variable-length motifs in large-scale time series.Knowl. Inf. Syst.61, 1 (2019), 513–542. doi:10.1007/S10115- 018-1279-6

  6. [13]

    Eric Jacobsen and Richard G. Lyons. 2003. The sliding DFT.IEEE Signal Process. Mag.20, 2 (2003), 74–80. doi:10.1109/MSP.2003.1184347

  7. [14]

    Eric Jacobsen and Richard G. Lyons. 2004. An update to the sliding DFT.IEEE Signal Process. Mag.21, 1 (2004), 110–111. doi:10.1109/MSP.2004.1516381

  8. [15]

    Sean Law. 2019. STUMPY: A Powerful and Scalable Python Library for Time Series Data Mining.J. Open Source Softw.4, 39 (2019), 1504. doi:10.21105/JOSS. 01504

  9. [16]

    Keogh, Stefano Lonardi, and Pranav Patel

    Jessica Lin, Eamonn J. Keogh, Stefano Lonardi, and Pranav Patel. 2002. Finding Motifs in Time Series. InProceedings of the Second Workshop on Temporal Data Mining, at the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Edmonton, Alberta, Canada, 53–68

  10. [17]

    Michele Linardi and Themis Palpanas. 2018. Scalable, Variable-Length Similarity Search in Data Series: The ULISSE Approach.Proc. VLDB Endow.11, 13 (2018), 2236–2248. doi:10.14778/3275366.3275372

  11. [18]

    Michele Linardi, Yan Zhu, Themis Palpanas, and Eamonn J. Keogh. 2018. Matrix Profile X: VALMOD - Scalable Discovery of Variable-Length Motifs in Data Series. InProceedings of the 2018 International Conference on Management of Data, SIGMOD Conference 2018, Houston, TX, USA, Jun...

  12. [19]

    Michele Linardi, Yan Zhu, Themis Palpanas, and Eamonn J. Keogh. 2020. Matrix profile goes MAD: variable-length motif and discord discovery in data series.Data Min. Knowl. Discov.34, 4 (2020), 1022–1071. doi:10.1007/S10618-020-00685-W

  13. [20]

    Yue Lu, Thirumalai Vinjamoor Akhil Srinivas, Takaaki Nakamura, Makoto Imamura, and Eamonn J. Keogh. 2023. Matrix Profile XXX: MADRID: A Hyper-Anytime and Parameter-Free Algorithm to Find Time Series Anom- alies of all Lengths. InIEEE International Conference on Data Mining, IC...

  14. [21]

    Zuluaga, and Eamonn J

    Yue Lu, Renjie Wu, Abdullah Mueen, Maria A. Zuluaga, and Eamonn J. Keogh

  15. [22]

    Qin Lv, William Josephson, Zhe Wang, Moses Charikar, and Kai Li. 2007. Multi- Probe LSH: Efficient Indexing for High-Dimensional Similarity Search. InPro- ceedings of the 33rd International Conference on Very Large Data Bases, Uni- versity of Vienna, Austria, September 23-27, ...

  16. [23]

    Frank Madrid, Shima Imani, Ryan Mercer, Zachary Zimmerman, Nader Shakibay Senobari, and Eamonn J. Keogh. 2019. Matrix Profile XX: Finding and Visual- izing Time Series Motifs of All Lengths using the Matrix Profile. In2019 IEEE International Conference on Big Knowledge, ICBK 2...

  17. [25]

    Abdullah Mueen and Eamonn J. Keogh. 2010. Online discovery and maintenance of time series motifs. InProceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, July 25-28, 2010, Bharat Rao, Balaji Krishnapuram, Andre...

  18. [26]

    Keogh, Jessica Lin, and Stefano Lonardi

    Pranav Patel, Eamonn J. Keogh, Jessica Lin, and Stefano Lonardi. 2002. Mining Motifs in Massive Time Series Databases. InProceedings of the 2002 IEEE Inter- national Conference on Data Mining (ICDM 2002), 9-12 December 2002, Maebashi City, Japan. IEEE Computer Society, 370–377...

  19. [27]

    Mendelzon

    Davood Rafiei and Alberto O. Mendelzon. 1998. Efficient Retrieval of Similar Time Sequences Using DFT. InThe 5th International Conference of Foundations of Data Organization (FODO’98), Kobe, Japan, November 12-13, 1998, Katsumi Tanaka and Shahram Ghandeharizadeh (Eds.). 249–257

  20. [28]

    Patrick Schäfer and Ulf Leser. 2022. Motiflets - Simple and Accurate Detection of Motifs in Time Series.Proc. VLDB Endow.16, 4 (2022), 725–737. doi:10.14778/ 3574245.3574257

  21. [29]

    Boedi- hardjo, Crystal Chen, and Susan Frankenstein

    Pavel Senin, Jessica Lin, Xing Wang, Tim Oates, Sunil Gandhi, Arnold P. Boedi- hardjo, Crystal Chen, and Susan Frankenstein. 2018. GrammarViz 3.0: Interactive Discovery of Variable-Length Time Series Patterns.ACM Trans. Knowl. Discov. Data12, 1 (2018), 10:1–10:28. doi:10.1145/3051126

  22. [30]

    Papalexakis, Eamonn J

    Maryam Shahcheraghi, Trevor Cappon, Samet Oymak, Evangelos E. Papalexakis, Eamonn J. Keogh, Zachary Zimmerman, and Philip Brisk. 2021. Matrix Profile Index Approximation for Streaming Time Series. In2021 IEEE International Conference on Big Data (Big Data), Orlando, FL, USA, D...

  23. [31]

    Avery Wang. 2003. An Industrial Strength Audio Search Algorithm. InISMIR 2003, 4th International Conference on Music Information Retrieval, Baltimore, Maryland, USA, October 27-30, 2003, Proceedings

  24. [32]

    Daan Van Wesenbeeck, Aras Yurtman, Wannes Meert, and Hendrik Blockeel

  25. [33]

    Chin-Chia Michael Yeh, Yan Zhu, Liudmila Ulanova, Nurjahan Begum, Yifei Ding, Hoang Anh Dau, Diego Furtado Silva, Abdullah Mueen, and Eamonn J. Keogh

  26. [34]

    Chin-Chia Michael Yeh, Yan Zhu, Liudmila Ulanova, Nurjahan Begum, Yifei Ding, Hoang Anh Dau, Zachary Zimmerman, Diego Furtado Silva, Abdullah Mueen, and Eamonn J. Keogh. 2018. Time series joins, motifs, discords and shapelets: a unifying view that exploits the matrix profile.D...

  27. [35]

    Yoon, Ossian O’Reilly, Karianne J

    Clara E. Yoon, Ossian O’Reilly, Karianne J. Bergen, and Gregory C. Beroza. 2015. Earthquake detection through computationally efficient similarity search.Science Advances1, 11 (2015). doi:10.1126/sciadv.1501057

  28. [36]

    Nikovski

    Jing Zhang and Daniel N. Nikovski. 2022. Algorithms for Fast Computation of Pan Matrix Profiles of Time Series Under Unnormalized Euclidean Distances. In International Conference on Applied Statistics and Data Analytics. Also available as MERL Technical Report TR2022-041. http...

  29. [37]

    Yunyue Zhu and Dennis E. Shasha. 2002. StatStream: Statistical Monitoring of Thousands of Data Streams in Real Time. InProceedings of 28th International Conference on Very Large Data Bases, VLDB 2002, Hong Kong, August 20-23, 2002. Morgan Kaufmann, 358–369. doi:10.1016/B978-15...

  30. [38]

    Yan Zhu, Chin-Chia Michael Yeh, Zachary Zimmerman, Kaveh Kamgar, and Eamonn J. Keogh. 2018. Matrix Profile XI: SCRIMP++: Time Series Motif Dis- covery at Interactive Speeds. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. IEEE Comput...

  31. [39]

    Funning, Abdullah Mueen, Philip Brisk, and Eamonn J

    Yan Zhu, Zachary Zimmerman, Nader Shakibay Senobari, Chin-Chia Michael Yeh, Gareth J. Funning, Abdullah Mueen, Philip Brisk, and Eamonn J. Keogh

  32. [40]

    Funning, Philip Brisk, and Eamonn J

    Zachary Zimmerman, Kaveh Kamgar, Nader Shakibay Senobari, Brian Crites, Gareth J. Funning, Philip Brisk, and Eamonn J. Keogh. 2019. Matrix Profile XIV: Scaling Time Series Motif Discovery with GPUs to Break a Quintillion Pairwise Comparisons a Day and Beyond. InProceedings of ...

  33. [41]

    Funning, Evangelos E

    Zachary Zimmerman, Nader Shakibay Senobari, Gareth J. Funning, Evangelos E. Papalexakis, Samet Oymak, Philip Brisk, and Eamonn J. Keogh. 2019. Matrix Profile XVIII: Time Series Mining in the Face of Fast Moving Streams using a Learned Approximate Matrix Profile. In2019 IEEE In...

  34. [43]

    Matrix Profile II: Exploiting a Novel Algorithm and GPUs to Break the One Hundred Million Barrier for Time Series Motifs and Joins. InIEEE 16th International Conference on Data Mining, ICDM 2016, December 12-15, 2016, Barcelona, Spain, Francesco Bonchi, Josep Domingo-Ferrer, R...

  35. [340]

    Its outputs are deterministic (Prop

    This configuration was measured in its own session on a fresh instance of the same machine type from a byte-identical binary. Its outputs are deterministic (Prop. 5), and the control totals reproduced across sessions within half a percent. Table 4 condenses the study. The swee...

  36. [2016]

    Matrix Profile I: All Pairs Similarity Joins for Time Series: A Unifying View That Includes Motifs, Discords and Shapelets. InIEEE 16th International Conference on Data Mining, ICDM 2016, December 12-15, 2016, Barcelona, Spain, Francesco Bonchi, Josep Domingo-Ferrer, Ricardo B...

  37. [2022]

    InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022, Aidong Zhang and Huzefa Rangwala (Eds.)

    Matrix Profile XXIV: Scaling Time Series Anomaly Detection to Trillions of Datapoints and Ultra-fast Arriving Data Streams. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022, Aidong Zhang and Huzefa Rang...

  38. [2024]

    LoCoMotif: discovering time-warped motifs in time series.Data Min. Knowl. Discov.38, 4 (2024), 2276–2305. doi:10.1007/S10618-024-01032-Z

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.