Selected Publications

Virtually all my research between 2003 and 2008 was funded by my NSF Career Award, for which I am extremely grateful. My research (2008 to 2012) was funded by NSF awards 0803410  and 0808770. My current research is funded by IIS-1161997 II, Vodafone's Wireless Innovation Project, several Bill & Melinda Gates Foundation Grand Challenge Awards, a Smarter Planet Award from IBM and gifts from Google, Microsoft, MERL Labs, Samsung and Siemens.

My publications on DBLP are here, Google scholar has this on me. My H-index is 89 (according to Google scholar)

My favorite venues: I have 28 SIGKDD papers, 39 ICDM papers, 23 SDM papers, 25 Data Mining and Knowledge Discovery papers, 11 (P)VLDB papers,  12 KAIS papers,..

Best paper awards:   IEEE ICDM 17, ACM SIGKDD 12, IEEE ICDM 07, ACM SIGMOD 01,  SIAM SDM 10 (Best student paper with Bilson), JCDL 09 (Best student paper r-up with Xiaoyue). KDD 97 r-up.

  • Matrix Profile I: All Pairs Similarity Joins for Time Series: A Unifying View that Includes Motifs, Discords and Shapelets. Chin-Chia Michael Yeh, Yan Zhu, Liudmila Ulanova, Nurjahan Begum, Yifei Ding, Hoang Anh Dau, Diego Furtado Silva, Abdullah Mueen, Eamonn Keogh (2016). IEEE ICDM 2016. [pdf] [slides]
  • Matrix Profile II: Exploiting a Novel Algorithm and GPUs to break the one Hundred Million Barrier for Time Series Motifs and Joins.  Yan Zhu, Zachary Zimmerman, Nader Shakibay Senobari, Chin-Chia Michael Yeh, Gareth Funning, Abdullah Mueen, Philip Berisk and Eamonn Keogh (2016). EEE ICDM 2016. [pdf] [slides] Shortlisted for best paper award.
  • Matrix Profile III: The Matrix Profile allows Visualization of Salient Subsequences in Massive Time Series. Chin-Chia Michael Yeh, Helga Van Herle, Eamonn Keogh (2016). IEEE ICDM 2016.  [pdf] [slides] Supporting Page.
  • Matrix Profile IIII: Using Weakly Labeled Time Series to Predict Outcomes. Chin-Chia Michael Yeh, Nickolas Kavantzas and; Eamonn Keogh. VLDB 2017 [pdf] Munich Germany.
  • Matrix Profile V: A Generic Technique to Incorporate Domain Knowledge into Motif Discovery. Hoang Anh Dau and Eamonn Keogh. [pdf KDD'17, Halifax, Canada.
  • Matrix Profile VI: Meaningful Multidimensional Motif Discovery. Chin-Chia Michael Yeh, Nickolas Kavantzas, Eamonn Keogh. [pdf] ICDM 2017.
  • Matrix Profile VII: Time Series Chains: A New Primitive for Time Series Data Mining. Yan Zhu, Makoto Imamura, Daniel Nikovski, and Eamonn Keogh. [pdf] ICDM 2017. Winner best paper award.  [slides]
  • Matrix Profile VIII: Domain Agnostic Online Semantic Segmentation at Superhuman Performance Levels. Shaghayegh Gharghabi, Yifei Ding, Chin-Chia Michael Yeh, Kaveh Kamgar, Liudmila Ulanova, and Eamonn Keogh. [pdf] ICDM 2017.
  • Matrix Profile IX: Admissible Time Series Motif Discovery with Missing Data [temp link]. Yan Zhu, Abdullah Mueen and Eamonn Keogh.TKDE 2020
  • Matrix Profile X: VALMOD - Scalable Discovery of Variable-Length Motifs in Data Series. Michele Linardi ,Yan Zhu ,Themis Palpanas and Eamonn Keogh. SIGMOD 2018.
  • Matrix Profile XI: SCRIMP++: Time Series Motif Discovery at Interactive Speed. Yan Zhu, Chin-Chia Michael Yeh, Zachary Zimmerman, Kaveh Kamgar and Eamonn Keogh, ICDM 2018. [PDF]
  • Matrix Profile XII: MPdist: A Novel Time Series Distance Measure to Allow Data Mining in More Challenging Scenarios. Shaghayegh Gharghabi, Shima Imani, Anthony Bagnall, Amirali Darvishzadeh, Eamonn Keogh. ICDM 2018. [expanded version PDF]
  • Matrix Profile XIII: Time Series Snippets: A New Primitive for Time Series Data Mining. Shima Imani, Frank Madrid, Wei Ding, Scott Crouter, Eamonn Keogh. IEEE Big Knowledge 2018. [expanded version PDF].
  • Matrix Profile XIV: Scaling Time Series Motif Discovery with GPUs: Breaking the Quintillion Pairwise Comparisons a Day Barrier. SoCC 2018. [pdf] [This paper had an interesting history]
  • Matrix Profile XV: Time Series Consensus Motifs: A New Primitive for Finding Repeated Structure in Time Series Sets. Kaveh Kamgar, Shaghayegh Gharghabi, and Eamonn Keogh (2019).  IEEE ICDM 2019. [pdf]
  • Matrix Profile XVI: Time Series Semantic Motifs: A New Primitive for Finding Higher-Level Structure in Time Series. Shima Imani and Eamonn Keogh (2019).  IEEE ICDM 2019. [pdf]
  • Matrix Profile XVII: (range) (work in progress)
  • Matrix Profile XVIII: Time Series Mining in the Face of Fast Moving Streams using a Learned Approximate Matrix Profile. Zachary Zimmerman, Nader Shakibay Senobari, Gareth Funning, Evangelos Papalexakis, Samet Oymak, Philip Brisk, and Eamonn Keogh (2019). IEEE ICDM 2019. [pdf]
  • Matrix Profile XIX: Efficient and Effective Labeling of Massive Time Series Archives. Frank Madrid, Shailendra Singh, Quentin Chesnais, Kerry Mauck and Eamonn Keogh. DSAA 2019: International Conference on Data Science and Advanced Analytics [pdf].
  • Matrix Profile XX: Finding and Visualizing Time Series Motifs of All Lengths using the Matrix Profile. Frank Madrid, Shima Imani, Ryan Mercer, Zacharay Zimmerman, Nader Shakibay, Eamonn Keogh. IEEE Big Knowledge 2019 [pdf]

Shima Imani, Sara Alaee, Eamonn J. Keogh: Putting the Human in the Time Series Analytics Loop. WWW (Companion Volume) 2019: 635-644
Chin-Chia Michael Yeh, Yan Zhu, Hoang Anh Dau, Amirali Darvishzadeh, Mikhail Noskov, Eamonn J. Keogh: Online Amnestic DTW to allow Real-Time Golden Batch Monitoring. KDD 2019: 2604-2612 [pdf]
Rodica Neamtu, Ramoza Ahsan, Elke A. Rundensteiner, Gábor N. Sárközy, Eamonn J. Keogh, Hoang Anh Dau, Cuong Nguyen, Charles Lovering: Generalized Dynamic Time Warping: Unleashing the Warping Power Hidden in Point-Wise Distances. ICDE 2018: 521-532
Yilin Shen, Yanping Chen, Eamonn J. Keogh, Hongxia Jin: Accelerating Time Series Searching with Large Uniform Scaling. SDM 2018: 234-242

Alireza Abdoli, Amy C. Murillo , Chin-Chia M. Yeh, Alec C. Gerry , Eamonn J. Keogh (2018) Time Series Classification to Improve Poultry Welfare. ICMLA 2018 [pdf]

Germain Forestier, Francois Petitjean, Hoang Anh Dau, Geoffrey Webb, and Eamonn Keogh. Generating synthetic time series to augment sparse datasets.  ICDM 2017.

Yifei Ding and Eamonn Keogh. Query Suggestion to allow Intuitive Search in Multidimensional Time Series. SSDBM 2017.

Diego Furtado Silva, Gustavo Batista, Eamonn Keogh (2016). Prefix and Suffix Invariant Dynamic Time Warping. IEEE ICDM 2016. [pdf]

Hoang Anh Dau, Nurjahan Begum and Eamonn Keogh. Semi-Supervision Dramatically Improves Time Series Clustering under Dynamic Time Warping. ACM CIKM 2016 [pdf]

Yan Zhu and Eamonn Keogh (2016). Irrevocable-choice algorithms for sampling from a stream. Data Mining and Knowledge Discovery, (), 1-26. [pdf]

Liudmila Ulanova, Nurjahan Begum, Mohammad Shokoohi-Yekta and Eamonn Keogh (2016). Clustering in the Face of Fast Changing Streams. SDM (2016)

Mohammad Shokoohi-Yekta, Yanping Chen, Bilson Campana, Bing Hu, Jesin Zakaria, Eamonn Keogh (2015). Discovery of Meaningful Rules in Time Series. SIGKDD 2015. [pdf]

Nurjahan Begum, Liudmila Ulanova, Jun Wang, Eamonn Keogh (2015). Accelerating Dynamic Time Warping Clustering with a Novel Admissible Pruning Strategy  SIGKDD 2015 [pdf]

Liudmila Ulanova, Tan Yan, Haifeng Chen, Guofei Jiang, Eamonn Keogh, Kai Zhang. Efficient Long-Term Degradation Profiling in Time Series for Complex Physical Systems. SIGKDD 2015 [pdf]

Liudmila Ulanova, Nurjahan Begum and Eamonn Keogh (2015). Scalable Clustering of Time Series with U-Shapelets. SDM 2015 [pdf]

Mohammad Shokoohi-Yekta, Jun Wang and Eamonn Keogh (2015). On the Non-Trivial Generalization of Dynamic Time Warping to the Multi-Dimensional Case. SDM 2015. [pdf]

Francois Petitjean, Germain Forestier, Geoffrey Webb, Ann E. Nicholson, Yanping Chen, and Eamonn Keogh (2014) Dynamic Time Warping Averaging of Time Series allows Faster and more Accurate Classification. ICDM 2014 [pdf]

Nurjahan Begum and Eamonn Keogh. Rare Time Series Motif Discovery from Unbounded Streams. VLDB 2015 [pdf]

Yuan Hao, Mohammad Shokoohi-Yekta, George Papageorgiou, and Eamonn Keogh. Parameter-Free Audio Motif Discovery in Large Data Archives. ICDM 2013 [pdf]

Bing Hu, Yanping Chen, Jesin Zakaria, Liudmila Ulanova, and Eamonn Keogh. Classification of Multi-Dimensional Streaming Time Series by Weighting each Classifiers Track Record. ICDM 2013. [pdf]

Yanping Chen, Bing Hu, Eamonn Keogh, Gustavo E.A.P.A Batista. (2013) DTW-D: Time Series Semi-Supervised Learning from a Single Example.  SIGKDD 2013 [pdf]

Yuan Hao, Yanping Chen, Jesin Zakaria, Bing Hu, Thanawin Rakthanmanon, Eamonn Keogh (2013). Towards Never-Ending Learning from Time Series Streams. SIGKDD 2013. [pdf]

Oben Tataw, Thanawin Rakthanmanon and Eamonn Keogh (2013) Clustering of Symbols using Minimal Description Length. ICDAR 2013

Thanawin Rakthanmanon and Eamonn Keogh. Fast-Shapelets: A Scalable Algorithm for Discovering Time Series Shapelets. SDM 2013 [pdf]

Bing Hu, Yanping Chen and Eamonn Keogh. Time Series Classification under More Realistic Assumptions. SDM 2013 [pdf] [slides]

Jesin Zakaria, Abdullah Mueen and Eamonn Keogh: Clustering Time Series using Unsupervised-Shapelets. ICDM 2012  [pdf].
Thanawin Rakthanmanon, Bilson Campana, Abdullah Mueen, Gustavo Batista, Brandon Westover, Qiang Zhu, Jesin Zakaria, Eamonn Keogh (2012). Searching and Mining Trillions of Time Series Subsequences under Dynamic Time Warping SIGKDD 2012. Best paper award [pdf].[data/code/video].

  • Recent optimizations on DTW similarity search can make this entire operation feasible in real time. The optimizations used by this paper are ..the UCR Suite. Stuart Russell et al CHI 2013

  • We observe that UCR-Suite wins in exact query answering and on hard queries. Echihabi et al, VLDB 2019.


Bing Hu, Thanawin Rakthanmanon, Bilson Campana, Abdullah Mueen, Eamonn Keogh. Image Mining of Historical Manuscripts to Establish Provenance. SDM 2012 [pdf].
Qiang Zhu, Gustavo Batista, Thanawin Rakthanmanon, Eamonn Keogh. A Novel Approximation to Dynamic Time Warping allows Anytime Clustering of Massive Time Series Datasets. SDM 2012
Yuan Hao, Bilson Campana, Eamonn Keogh. Monitoring and Mining Insect Sounds in Visual Space. SDM 2012 [pdf].
Jesin Zakaria, Sarah Rotschafer, Abdullah Mueen, Khaleel Razak, Eamonn Keogh. Mining Massive Archives of Mice Sounds with Symbolized Representations. SDM 2012 [pdf].

Bing Hu, Thanawin Rakthanmanon, Yuan Hao, Scott Evans, Stefano Lonardi, and Eamonn Keogh (2011). Discovering the Intrinsic Cardinality and Dimensionality of Time Series using MDL. ICDM 2011 [pdf].
Thanawin Rakthanmanon, Qiang Zhu, and Eamonn Keogh (2011). Mining Historical Documents for Near-Duplicate Figures.
ICDM 2011 [pdf].
Thanawin Rakthanmanon, Eamonn Keogh, Stefano Lonardi, and Scott Evans (2011). Time Series Epenthesis: Clustering Time Series Streams Requires Ignoring Some Data.
ICDM 2011 [pdf].

Abdullah Mueen, Eamonn Keogh, Neal Young  (2011). Logical-Shapelets: An Expressive Primitive for Time Series Classification.  SIGKDD 2011. [pdf].

Gustavo  Batista, Eamonn J. Keogh, Agenor Mafra-Neto, Edgar Rowton. Sensors and Software to allow Computational Entomology, an Emerging Application of Data Mining. SIGKDD 2011

Gustavo Batista, Xiaoyue Wang and Eamonn J. Keogh (2011) A Complexity-Invariant Distance Measure for Time Series. SDM 2011 [pdf] [Supplemental Material].

  • we can observe that more discrimination is made between sensors inside the house with the CID method.  Gouy-Pailler et al CIB 2011

  • we adopt the complexity invariant distance measure CID  which uses complexity differences between two time series as a correction factor for existing distance measures. Empirically, we found that it is robust against noise introduced by video transformations.  Zhang et al. 2012
  • to take into account the various complexities in time series matching, we use complexity-invariant distance measure CID ..Empirically, we have found that (CID) is robust against noise. Ren et al. ICMR 2012


Qiang Zhu and Eamonn Keogh (2010) Mother Fugger: Mining Historical Manuscripts with Local Color Patches. ICDM 2010 [pdf].

Vit Niennattrakul, Chotirat Ann Ratanamahatana and Eamonn Keogh (2010). Data Editing Techniques to Allow the Application of Distance-Based Outlier Detection to Streams. ICDM 2010. [pdf]

Doruk Sart, Abdullah Mueen, Walid Najjar, Vit Niennattrakul, and Eamonn Keogh (2010). Accelerating Dynamic Time Warping Subsequnce Search with GPUs and FPGAs. ICDM 2010 [pdf]
Jin Shieh and Eamonn Keogh (2010). Polishing the Right Apple: Anytime Classification Also Benefits Data Streams with Constant Arrival Times. ICDM 2010 [pdf]

Alessandro Camerra, Themis Palpanas,  Jin Shieh and Eamonn Keogh (2010) iSAX 2.0: Indexing and Mining One Billion Time Series. ICDM 2010 [pdf]

Abdullah Mueen and Eamonn Keogh (2010) Online Discovery and Maintenance of Time Series Motifs. SIGKDD 2010. [pdf] [Youtube demo of online motifs].

Qiang Zhu and Eamonn Keogh (2010) Using CAPTCHAs to Index Cultural Artifacts. The Ninth International Symposium on Intelligent Data Analysis [pdf].

Bilson Campana and Eamonn Keogh (2010). A Compression Based Distance Measure for Texture. SDM 2010. [pdf] Best Student Paper.

  • We introduce the use of the Campana-Keogh distance, a video compression-based measure, to retrieve music informationbased in its structure. Silva et al. ISMIR 2013

  • we consider the work of Campana and Keogh (for) measuring and summarizing movement in microblog postings. Ruiz et al. ICWSM 2013

  • (We use a minor variant of CK measures for) Cytoplasm Image Classification. Zhang et al. BMEI 2012

  • Taking (Campana-Keogh ) as a reference, we classify stem cell differentiation Images. Ming Li et al. ECML 2012

  • We show that our augmentation of the CK method works extremely well for content-based commercial produce image retrieval. Chai et al. IEEE ICIP 2012.


Abdullah Mueen, Eamonn Keogh and  Nima Bigdely-Shamlo (2009). Finding Time Series Motifs in Disk-Resident Data. ICDM 2009. [pdf] [slides] [Youtube demo of motifs for MoCap data].  

Xiaoyue Wang, and Eamonn Keogh (2009) Finding Centuries-Old Hyperlinks with a Novel Semi-Supervised Learning Technique. 9th ACM/IEEE-CS joint conference on Digital libraries. JCDL 2009.

Lexiang Ye and Eamonn Keogh (2009) Time Series Shapelets: A New Primitive for Data Mining. SIGKDD 2009  [pdf] [datasets]

  • Our idea of making use of T-Motifs is motivated by a recent work (on time series shapelets). Li, Lin, Ding and Han 2011.

  • Specifically, we advocate local shapelets as features.. Xing, Pei, Yu and Wang 2011.

  • Our approach for gesture recognition is similar to time series shapelets. Hartman and Link 2010.

  • Multivariate Shapelets Detection allows for early and patient-specific classification of multivariate time series... Ghalwash and Obradovic. Bioinformatics 2012

  • the most promising approaches proposed for TSC is time series shapelets.. Lines and Bagnall IDEAL 2012

  • Two key advantages of classification with shapelets are its high accuracy and the interpretability of the classification model learnt. Gordon, Hendler and Rokach 2012

  • (to learn) a dictionary of behavioral motifs revealing clusters of genes affecting Caenorhabditis elegans locomotion we used the rmsd as the distance between a motif and a time series (shapelet)...  Brown et al. PNAS 2012


Qiang Zhu, Xiaoyue Wang, Eamonn Keogh, Sang-Hee Lee (2009). Augmenting the Generalized Hough Transform to Enable the Mining of Petroglyphs. SIGKDD 2009.  [pdf] [datasets]

Abdullah Mueen, Eamonn Keogh, Qiang Zhu, Sydney Cash,  Brandon Westover (2009). Exact Discovery of Time Series Motifs. SDM 2009.[pdf]. See also this page.

  • However, all of these temporal motifs have been identified by subjective visual inspection of EEG traces and not by a principled and automated search. Mueen and associates (2009) extend the notion of timeseries motifs to EEG timeseries with a motif being an EEG snippet that contains a recurrent (and potentially meaningful) pattern in a set of longer timeseries. Petra Ritter  BRAIN CONNECTIVITY 2013

  • The algorithm we used for discovering motifs in the music data is called the Mueen-Keogh (MK) algorithm.. Cabredo et al. 2011

  • Our solution adopts (Mueen and Keoghs) motifs to compress data streams... Danieletto et al 2012.

  • we apply the MK algorithm to amplitude time series retrieved from seismic signals recorded during episodic eruptive activity of Mt Etna in 2011. . This is only one example of the potential application of this fairly novel MK technique in seismology Cassisi et al 2012


Lexiang Ye, Xiaoyue Wang, Eamonn Keogh and Agenor Mafra-Neto (2009). Autocannibalistic and Anyspace Indexing Algorithms with Applications to Sensor Data Mining. SDM 2009. [pdf] See also this page.

Shashwati Kasetty, Candice Stafford, Gregory P. Walker, Xiaoyue Wang and, Eamonn Keogh (2008). Real-Time Classification of Streaming Sensor Data. 20th IEEE Int'l Conference on Tools with Artificial Intelligence. [pdf]

Jin Shieh and Eamonn Keogh (2008) iSAX: Indexing and Mining Terabyte Sized Time Series. SIGKDD 2008.  [pdf] The iSAX page is here. [slides].

  • The times needed to run an EGL search on a standard laptop were 8 hours for HMMs, 10 - 15 minutes for 1-NN, and 22 seconds with iSAX. Kohlsdorf, Starner and Ashbrook. FG 2011

  • Our approach is based on the state of art time series representation, iSAX. Castro & Azevedo 2010

  • we found iSAX to perform extremely well. Sorokin et al. Journal of Pharmaceutical Sciences.

  • Leveraging iSAX, we organize the...  Gu and Wang 2013

Hui Ding, Goce Trajcevski, Peter Scheuermann, Xiaoyue Wang and Eamonn Keogh (2008) Querying and Mining of Time Series Data: Experimental Comparison of Representations and Distance Measures VLDB 2008. [pdf][slides]

Xiaoyue Wang, Lexiang Ye, Eamonn Keogh and Christian Shelton. (2008). Annotating Historical Archives of Images. JCDL 2008 [pdf]
Lexiang Ye, Xiaoyue Wang, Dragomir Yankov and Eamonn Keogh (2008). The Asymmetric Approximate Anytime Join: A New Primitive with Applications to Data Mining. SDM 2008. [pdf] Datasets are here.

Dragomir Yankov, Eamonn Keogh, and Umaa Rebbapragada (2007). Disk Aware Discord Discovery: Finding Unusual Time Series in Terabyte Sized Datasets. ICDM 2007. [pdf] Best paper award.. Data and code is here: The powerpoint slides are here, but you should also download the movie and put it it the same folder as the talk.  [pdf]

Dragomir Yankov, Eamonn Keogh, and Kin Kan (2007). Locally Constrained Support Vector Clustering. ICDM 2007. [pdf]

Dragomir Yankov, Eamonn Keogh, Jose Medina, Bill Chiu, and Victor Zordan (2007). Detecting Motifs Under Uniform Scaling. SIGKDD 2007. [pdf] [Powerpoint] Supporting webpage with video and datasets.

Xiaopeng Xi, Eamonn Keogh, Li Wei, Agenor Mafra-Neto (2007). Finding Motifs in Database of Shapes. SIAM International Conference on Data Mining (SDM'07). [pdf]
Dragomir Yankov, Eamonn Keogh, Li Wei, Xiaopeng Xi and Wendy Hodges (2007). Fast Best-Match Shape Searching in Rotation Invariant Metric Spaces. SIAM International Conference on Data Mining (SDM'07). [pdf]

Ken Ueno, Xiaopeng Xi, Eamonn Keogh, Dah-Jye Lee (2006). Anytime Classification Using the Nearest Neighbor Algorithm with Applications to Stream Mining. ICDM  2006. [pdf] [Powerpoint]

Dragomir Yankov and Eamonn Keogh. (2006). Manifold Clustering of Shapes ICDM  2006. [pdf] [Powerpoint]

Eamonn Keogh, Li Wei, Xiaopeng Xi, Stefano Lonardi, Jin Shieh, Scott Sirowy  (2006). Intelligent Icons: Integrating Lite-Weight Data Mining and Visualization into GUI Operating Systems. ICDM  2006. [pdf]  [Powerpoint]

1.      We've had some success in using your intelligent icon concept combined with MDL learning for intrusion detection - detection of zero day attacks. Scott C. Evans, PhD, GE Research

2.      In the on-line mode of the DSS system,  intelligent icons give short-term support to members of planning departments as well as production operators to know which type of product is better...  Holzknecht et al. European Commission Report.

3.      The solution implemented use Intelligent Icons to represent these specific sub populations... Lefait and Kechadi

Li Wei, Eamonn Keogh and Xiaopeng Xi (2006) SAXually Explict Images: Finding Unusual Shapes. ICDM 2006. [Expanded version pdf] [Powerpoint] See also  HOT SAX  in ICDM 2005 below.

Eamonn Keogh, Li Wei, Xiaopeng Xi, Sang-Hee Lee and Michail Vlachos  (2006) LB_Keogh Supports Exact Indexing of Shapes under Rotation Invariance with Arbitrary Representations and Distance Measures. VLDB 2006. [pdf] [Powerpoint] For more information, see here.

Li Wei and Eamonn Keogh  (2006) Semi-Supervised Time Series Classification. SIGKDD 2006. [pdf]

Xiaopeng Xi, Eamonn Keogh, Christian Shelton, Li Wei & Chotirat Ann Ratanamahatana (2006). Fast Time Series Classification Using Numerosity Reduction. ICML. [pdf]

E. Keogh, J. Lin and A. Fu (2005). HOT SAX: Efficiently Finding the Most Unusual Time Series Subsequence. In Proc. of the 5th IEEE International Conference on Data Mining (ICDM 2005), pp. 226 - 233., Houston, Texas, Nov 27-30, 2005.  [pdf]. More info on discords and HOT SAX is here 

1.      ..on 19 different publicly available data sets, comparing 9 different techniques (time series discords) is the best overall technique among all techniques. Chandola, Cheboli, and Vipin Kumar 2009

2.      This (severe weather phenomena) rule-finding algorithm draws from the data structures for efficient detection of Discords described in Keogh et al  Amy McGovern et al. 2007

3.      Discords have great implications for fast and intelligent data mining  Ameen et al 2006.

4.      ..has been improved by designing the more intelligent Discord algorithms that can bypass most of the seemingly unnecessary parameters calculated  Basha et al 2007

5.      a seminal paper..the authors introduce the new problem of finding time series discords we confirm it can detect heartbeats anomalies  Mooi Choo 2007

6.      Requiring the analyst to provide only one parameter is a significant improvement over past methods that require several unintuitive parameters. Preston, Protopapas and Carla Brodley 2009

A SIGKDD paper claims to be more accurate that discords, but see this.

L. Wei, E. Keogh, H. Van Herle, and A. Mafra-Neto (2005). Atomic Wedgie: Efficient Query Filtering for Streaming Time Series. In Proc. of the 5th IEEE International Conference on Data Mining (ICDM 2005), pp. 490-497, Houston, Texas, Nov 27-30, 2005. [pdf] [slides].

Ada Wai-chee Fu, Eamonn Keogh, Leo Yung Hang Lau and Chotirat Ann Ratanamahatana (2005). Scaling and Time Warping in Time Series Querying.  VLDB 2005. [pdfNow a VLDB Journal.

Ratanamahatana, C., Keogh, E., Bagnall, T.  and Lonardi, S. (2005). A Novel Bit Level Time Series Representation with Implications for Similarity Search and Clustering. PAKDD 05. [pdf 

Kumar, N.,  Lolla  N.,  Keogh, E.,  Lonardi, S. , Ratanamahatana, C. A. and Wei, L. (2005). Time-series Bitmaps: A Practical Visualization Tool for working with Large Time Series Databases . In proceedings of SIAM International Conference on Data Mining (SDM '05), Newport Beach, CA, April 21-23. pp. 531-535 [pdf

Ratanamahatana, C. A. and Keogh. E. (2005). Three Myths about  Dynamic Time Warping. In proceedings of SIAM International Conference on Data Mining (SDM '05), Newport Beach, CA, April 21-23,  pp. 506-510  [pdf , slides].   Also appeared as a workshop paper with the following unlikely title...

Ratanamahatana, C. A. and Keogh. E. (2004). Everything you know about Dynamic Time Warping is Wrong. Third Workshop on Mining Temporal and Sequential Data, in conjunction with the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), August 22-25, 2004 - Seattle, WA. [pdf , slides

Keogh, E., Lonardi, S. and Ratanamahatana, C. (2004). Towards Parameter-Free Data Mining. In proceedings of the tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Seattle, WA, Aug 22-25, 2004. [pdf, slides ].

1.      A feature-free approach to spam filtering, offered by CDM, has several advantages over a feature-based approach...its remarkable accuracy. A second advantage is its low set-up costs: feature extraction, selection and weighting are all unnecessary.  Delany and Bridge

2.      ...CDM compares favorably with the results from the conventional methods of Fetal heart rate (FHR) monitoring reported in International guidelines.  Santos et al.

3.      (for) forensic authorship attribution we follow the Compression Distance Metric (CDM) defined in Keogh et al. Lambers and Veenman 2009.

4.      ...Keogh et. al  suggested a clever way of approximating the Kolmogorov complexity with the Lempel-Ziv compression length. .  Faloutsos 2007.

5.      ..inspired by (CDM) found to be reliable indicator of predictability  for a deployment of enterprise services.. Andrzejak et al DSOM 08

6.      a powerful, parameter-free, data mining paradigm. Giancarlo et al. Bioinformatics 2009

7.      Through the use of the Compression-based Dissimilarity Measure CDM.. we have identified cases where .. Rost, Edsberg, Grimsmo and Nytro, 2008

Lin, J., Keogh, E., Lonardi, S., Lankford, J. P. & Nystrom, D. M. (2004). Visually Mining and Monitoring Massive Time Series. In proceedings of the tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Seattle, WA, Aug 22-25, 2004.
(This work also appears as a VLDB 2004 demo paper, under the title  VizTree: a Tool for Visually Mining and Monitoring Massive Time Series. ) [pdf ,slides]

1.      (VizTree).. a way to do such analysis more systematically  Edward Tufte (website)

2.      Your Viztree work has inspired our thinking around visualizing health care data using state machines.  Melanie Rosenthal,

Keogh, E., Palpanas, T., Zordan, V., Gunopulos, D. and Cardle, M. (2004) Indexing Large Human-Motion Databases. In proceedings of the 30th International Conference on Very Large Data Bases, Toronto, Canada. [pdf, slides ]

Lin, J., Keogh, E., Lonardi, S. & Chiu, B. (2003) A Symbolic Representation of Time Series, with Implications for Streaming Algorithms. In proceedings of the 8th ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery. San Diego, CA. June 13. [pdf, slides] An expanded version of this paper appears in The Data Mining and Knowledge Discovery Journal.

1.      the performance SAX enables is amazing, and I think a real breakthrough. As an example, we can find similarity searches using edit distance over 10,000 time series in 50 milliseconds. Ray Cromwell,

2.      SAX represents the state-of-the-art in time series data streams analysis due to its generality   Gaber and Gama, Tutorial PKDD07

3.      In order to characterize the expression waveforms we follow the basic SAX  formalism for time-series analysis presented by Keogh and Lin... Androulakis et al

4.      (SAX based VizTree is).. a way to do such analysis more systematically Edward Tufte.

5.      The method is based on the notion of so called time signature of the clusters, introduced in Lin & Keogh and obtained using a recent time series analysis method called the Symbolic Aggregate approximation (SAX). Pouget, M. Dacier, J. Zimmerman, A. Clark, and G. Mohay. Journal of Information Assurance and Security 1 (2006) 21-32

6.      "Our goal is to identify those transcripts that share significant components of their expression patterns. In order to do so, we explore the SAX idea of Lin and Keogh... Yang et al..

7.      we take SAX Motif developed by Keogh in order to support a medical expert in discovering interesting knowledge. Kitaguchi, S.

8.      In order to symbolize a street data, we utilize the SAX approach. Jalili and Alipour. 

9.      ...we examine another interesting query, the Time Relaxed Spatiotemporal Trajectory Join (TRSTJ)... we address the TRSTJ problem using SAX... Bakalov, Hadjieleftheriou and Tsotras. 

10.  We have decided to use SAX (to detect sophisticated attack tools ).. SAX is a recent and popular method with interesting proven properties. F. Pouget, G. Urvoy-Keller, and M. Dacier 

11.  ..we are currently using (Lin and Keoghs SAX) approach to creating discrete data from continuous data. Amy McGovern et al.

12.  (to find repeated patterns in protein unfolding data) ...we adopted a two step approach called SAX Ferreira et al.

13.  ..we use SAX and Keogh's Tarzan algorithm to do anomaly detection in network traffic. Kyoji Umemura et. al.

14.  SAX has already prove efficient in a large variety of domains Fabian Pouget, Telecom Paris.

15.  SAX representation of abstracted data makes analysis (of anterior-posterior center of pressure) more easy and accurate.   Bhatkar et al.

16.  Our Symbolic Transformation (based on SAX method) can be use to discover novel gene relations by mining similar subsequences in time-series microarray data. Vincent Shin-Mu Tseng

17.  ..we use SAX bitmap matrices to compute an anomaly score for acoustic signals, enabling the extraction of bird vocalizations and other acoustic events Kasten, McKinley and Gage. 2007

18.  Using the time-series data as an input, it takes too much computation amount to extract motifs from the human motion information. Therefore, we use Symbolic Aggregate approXimation (SAX) .Araki , Arita and Taniguchi 2006

19.  SAX has the advantage of dimensionality and noise reductions. It also allows real valued data to remain the original characteristics with only an infinitesimal time and space overhead...we therefore use it for to determine behavior of system... Lavangnananda and Wongwattanakarn. SMCai07.

20.  SAX demonstrates some promising properties for the field of anomaly detection in a marine engine. Morgan, Liu, Turnbull, and Brown 2007.

21.  ..motivated by recent advances in the symbolic representation of streaming data (SAX), effectively reduces the dimensionality of.. Annu. Rev. Biomed. Eng. 2007.

22.  (we use SAX to create a) multiparty protocol for the privacy preserving pattern discovery problem. Costa da Silva and  Klusch 2007.

23.  By using SAX with the sensor network data, we are able to detect such complex patterns with good accuracy .. SAX is a very mature and robust solution for mining time-series data. Zoumboulakis and Roussos 2007

24.  we apply the (SAX based) motif discovery approach the analysis of responses obtained by tactile stimulation of different body areas. Fabri et al. IJCNN07

25.  We extend the symbolic aggregate approximation (SAX) approach.. to support Query-by-Singing/Humming. Duda, Nurnberger and Stober (2007).

26.  We use an algorithm based on SAX (Symbolic Aggregate approXimation) to discover human skills..  Makio, Tanaka, and Uehara 2007

27.  symbolic aggregate approximation (SAX) outperform other dimensionality reduction techniques like singular value decomposition or discrete fourier transform (SVD, DFT) for time series data.. Assent, Krieger, Afschari and Seidl EDBT 2008

28.  We will often apply this symbolization approach, using the methodology of SAX, with a primary goal of reducing the number of tunable parameters to increase the robustness of the approach Bollt and Skufca 2009

1.         More than 300 references, see the SAX page.

Ratanamahatana, C. A. and Keogh. E.  (2004). Making Time-series Classification More Accurate Using Learned Constraints. In proceedings of SIAM International Conference on Data Mining (SDM '04), Lake Buena Vista, Florida, April 22-24, 2004. pp. 11-22. [pdf, slides]

1.      ..our recognition algorithm can also be viewed as ...their (RK-Band) algorithm. The average performance is better than all the other algorithms.  Veeraraghavan et. al


Lin, J., Vlachos, M., Keogh, E., & Gunopulos, D (2004). Iterative Incremental Clustering of Time Series. In proceedings of the IX Conference on Extending Database Technology. Crete, Greece. March 14-18, 2004. [pdf]

T. Palpanas, M. Vlachos, E. Keogh, D. Gunopulos, W. Truppel (2004). Online Amnesic Approximation of Streaming Time Series. In ICDE . Boston, MA, USA, March 2004. [ pdf] Now a TKDE paper

1.      Amnesic approximation of streaming time series was set forth.. offering the intuitive idea that data can be approximated with a precision proportional to its age.  Timos Sellis 2006.

E. Keogh, J. Lin, and W. Truppel. (2003). Clustering of Time Series Subsequences is Meaningless: Implications for Past and Future Research. In proceedings of the 3rd IEEE International Conference on Data Mining . Melbourne, FL. Nov 19-22. pp 115-122. [ pdf]. Appears as a journal paper in Knowl. Inf. Syst. 8(2): 154-177 (2005)

  1. A deeper analysis... Hoppner and Klawonn 2009.

  2. The research field of subsequence clustering, which was already a widely applied and studied technique, took a dramatic turn after Keogh (published this paper) Robards and Sunehag 2009.

  3. Eamonn Keogh challenged the data mining community by showing that both the k-means and hierarchical clustering algorithms return meaningless results. Denton, Besemann and Dorr 2009

  4. Clustering of Time Series was a widely used technique in the Data Mining Community and beyond until the discovery by Keogh et. al. in 2003 that it is meaningless.  Chen 2006.

  5. Data mining and machine leaning communities were surprised when Keogh et al. (2003) pointed out that the k-means cluster centers in subsequence time-series clustering become sinusoidal pseudopatterns for almost all kinds of input time-series data.  Ide 2006.

  6. Keogh et al. considered clustering of streaming time series data, and claim that clustering such data is meaningless. We also found... (this to be true).  Singhal and Seborg 2006

  7. (after reading the "meaningless" paper) I checked the experiments with random walks and heart beat data and they seem to show that Das's algorithm does have unexpected and unexplained false positives and false negatives. Yasser F. O. Mohammad 2010

  8. after spending a couple of weeks in frustration trying to get k-means on subsequences to work and getting nothing but arbitrary clusterings, followed by a paper of yours and a tremendous forehead slap...Martin Mladenov 2010

Chiu, B. Keogh, E., & Lonardi, S. (2003). Probabilistic Discovery of Time Series Motifs. In the 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. August 24 - 27, 2003. Washington, DC, USA. pp 493-498. [Expanded Version pdf]

  1. Motif extraction makes it easier for system developers to enumerate human actions without omission since they can choose human actions from extracted motifs . Arita et al

  2. We use Keoghs time series motifs to find patterns in water levels.. Li and Nallela

  3. we take SAX Motif developed by Keogh in order to support a medical expert in discovering interesting knowledge . Kitaguchi, S.

  4. To analyze these ill-formed time-series data, we take another time-series data analysis method such .. and finding Motif based on SAX  Abe & Yamaguchi

  5. In our approach, the video database is pre-processed to classify the motion of the human figures and identify the movements of repeated sequences using motifsCelly & Zordan

  6. Our proposed algorithm identifies genes that maximally contribute to changes in the overall transcriptional state of the system... The characteristic attribute will be defined using established methods of SAX based motifs..  Yang et al.

  7. Our notion of KNTN was inspired by (the Motif definition) who define the concept of trivial matches in data mining for univariate time-series  Maja J Mataric.

  8. We use Keogh's motifs to compress data found in large system call data sets.. Wilson, Feyereisl and Aickelin 2007

  9. we apply the SAX based motif discovery approach the analysis of responses obtained by tactile stimulation of different body areas. Fabri et al. IJCNN07

  10. (for the problem of Estimation of Micro-drilled Hole Wall of PWBs) we take the Motif method developed by Keogh in order to express through hole wall quality more accurately. Toshiki et al.

  11. We use time series motifs to discover skills from music performances..  Makio, Tanaka, and Uehara 2007

  12. Our multi dimensional motif discovery and optimization method, which is an extension of the well-known approach proposed by Keogh. Majid Sarrafzadeh et al 2009.

  13. we explicitly focus on the SAX representation, which also provides some significant advantages for mining motifs. Patnaiky, Marwah, Sharma, and Ramakrishnan SIGKDD 2009

  14. In this case switching from motif 8 to motif 5 gives us a nearly $40,000 in annual savings! Debprakash Patnaik et al SIGKDD 2009

  15. We adopt a technique from data mining called motif discovery.. Ahmadi et al 2009

  16. (we use) finding time series motifs to forecast some short-term stock price tendencies and values. Jiang, Li and Han 2009

  17. we use time series motifs to find gesture patterns with applications to robot-human interactions. Okada, Izukura and Nishida 2011

  18. We use (time series) motifs and pattern cluster to handle sequential interaction processes and the meaning of occurring elemental events. Mase et al 2009

  19. We use Keogh's Motifs for unsupervised discovery of abnormal occurrences of activities in multi-dimensional time series data... Vahdatpour SDM 2010.

  20.  promising work in recent years on the discovery of motifs directly in time-series signals .. using piecewise aggregate approximation.. Zeeshan Syed TKDD 2010

  21. The key to our method is the automatic discovery of repetitive and informative subsequences,(Keoghs) motifs, in the noisy IMU data. ICPR 2010 Zhao et al

  22. the most efficient motif provided a power savings of 41 KW over the least efficient motif . This translates to an annual reduction of 287 tons of CO2.  Brian Watson InterPACK09.

  23. We adopt the algorithm of Mueen, et al. for finding motifs in a time series in music... Cabredo et all 2012

  24. (to learn) a dictionary of behavioral motifs revealing clusters of genes affecting Caenorhabditis elegans locomotion we used  the MK motif discovery algorithm... Brown et al. PNAS 2012

  25. First, the input command history is examined for recurring subsequence patterns using (Keoghs) motif discovery S. Dong AAAI 2010


Vlachos, M., Hadjieleftheriou, M., Gunopulos, D. & Keogh. E. (2003). Indexing Multi-Dimensional Time-Series with Support for Multiple Distance Measures. In the 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. August 24 - 27, 2003. Washington, DC, USA. pp 216-225. [ pdf].

Keogh, E. (2002). Exact indexing of dynamic time warping. In 28th International Conference on Very Large Data Bases. Hong Kong. pp 406-417. [Conference version pdf , Journal version pdf , slides ]

Visit the LB_Keogh Page

  • LB_Keogh makes retrieval of time-warped time series feasible even for large data sets. Muller et. al. SIGGRAPH 05.

  •  LB_Keogh is fast, because it cleverly exploits global constraints that appear in dynamic programming. Christos Faloutsos et al. PODS 2005

  • LB_Keogh, the best-so-far known method to lower bound DTW.. Capitani and Ciaccia SEBD05.

  • (LB_Keogh) makes DTW indexable by approximating time series with bounding envelopes. Lian et. al. TKDE 2009

  • Although DTW has a time complexity of O(n2), (LB_Keogh) bounding can effectively make DTW run in O(n). Smith and Craven csb2008

  • the standard DTW algorithm has a quadratic time and space complexity. For performance reason we implemented a FastDTW based on the ideas of LB_Keogh....and therefore the algorithm performs nearly in linear time. Juffinger, Granitzer and Lex WWW 2009

  • LB_Keogh  can significantly speed up DTW. Suzuki et al. ICML 2003.

  • (for) the problem of searching for similar acoustic data over unstructured decentralised P2P networks LB_Keogh can be effectively used for pruning, resulting in considerably less number of DTW computations. Karydis et. al. ICEIS05

  • by exploiting recent results on metric access structures and LB_Keogh, we can still guarantee the indexability of DFT coefficients extracted from large data sets. Bartolini et. al. IEEE PAMI.

  • LB_Keogh is preferred (for motion capture) since it offers the tightest lower bounds, is readily indexible with MBRs. Marc Cardle University of Cambridge Thesis

  • Keogh has pioneered many of the recent ideas in the indexing of dynamic time warping". Cole, Shasha, and Zhao SIGKDD 2005.

  • LB_Keogh performs uniformly the best. Chen and Ng. VLDB 2005.

  • Among the existing lower bound functions, LB_Keogh is the best in terms of the tightness of the lower bound. Zhou, M. and Wong, Information Sciences 2005

  • Despite the fact that DTW distance still does not satisfy the triangle inequality, LB_Keogh exactly index(s) the DTW distance. Wang, Z.

  • LB_Keogh has provided a convincing lower bound that can be tuned to provide high tightness to the actual DTW distance. Toni Rath

  • We exploit (LB_Keogh) in the context of reverse engineering of dynamic feature behavior to detect similarities between feature traces. This technique is ideally suited to handling large amounts of data. Kuhn and Greevy

  •  Recently, Keogh et al. presented an algorithm, based on the LB_Keogh function, which dramatically reduced the time complexity of the calculation of the Euclidean Distance measure. This speed up was further achieved by allowing indexing. Frentzos et al. ICDE 2007

  •  Exact indexing of DTW has been proposed in the literature,.. the LB_Keogh. Employing the indexing method allows us to reduce the computation (needed for query by humming). Roger Jang, PCM 2006.

  • To reduce the computational time and (quadratic) complexity inherent in dynamic time warping, we use (LB_)Keogh minimum bounds to quickly determine candidates for the set of nearest neighbors. Dalal and Olson. SPECTS07

  • (we use) LB_Keogh for efficient similarity range query processing in music databases. Ruxanda & Jensen 2006

  •  LB_Keogh.. has become a popular solution for indexing DTW because of its performance. Zhou &Wong ICDE 2007

  •  To speed up computations we could utilize spatiotemporal access methods similar to LB_Keogh" Zeinalipour Yazti , Lin and Gunopulos CIKM 2006

  • EDR incurs huge computational cost due to the lack of pruning techniques such as LB_Keogh lower bound Chen et al. ICDE07.

  • Since DTW is relatively slow to calculate... We therefore use (LB_)Keogh minimum bounds. Dalal, Musicant, Olson, McMenamy, Benzaid, Kazez and Bolan. icc2007.

  • Because DTW is not a metric, we use LB_Keogh to make indexing possible (for our query by humming system) Leung Tat-wan. 2005.

  • In order to use the SBR representation in a multidimensional index, we must have a distance function that lower bounds the distance between a query object and a group of time series data. Therefore we can use LB_Keogh.. Li, Lopez and Moon TKDE 2004

  • Keogh proposed LB Keogh and LB PAA, tight lower bounds under time-warping ... our solutions exploit these two lower bounds.. Han et al VLDB 2007

  • (LB_Keogh) significantly speeds up the DTW (for music similarity). Joan Serra 2007. Master Thesis UPF. Barcelona

  • For fault diagnosis (LB_keogh) substantially reduce the computational expense required. Rajshekhar, et. al 2007

  •  (we use) the tightest existing lower bound.. (for) exact indexing of Dynamic Time Warping, LBKeogh. Assent, Krieger, Afschari and Seidl EDBT 2008

  • (LB_Keogh) has significantly increased the accuracy of time series classification while reducing the computational expense required. Kumar, Gupta, Jayaraman & Kulkarni 2008

  • Exact indexing of DTW has been proposed in the literature.. using LB_Keogh.. Employing the indexing method allows us to reduce the computation (of query by humming). Jang and Lee 2008

  • the best lowerbound function in terms of tightness is the LB_Keogh... It cleverly exploits the warping window.. Zhou and Wong ICDE08

  • we describe how to further reduce matching time using a lower bound function based on LB_Keogh. Yoon-Sik Tak and Eenjun Hwang 2008

  • we use LB Keogh to obtain of two orders of magnitude over the brute force algorithm when doing fast correlation analysis.. Nguyen and Shiri CIKM 08

  • The tightest lower bound of DTW, between a query envelope E(Q) and a data sequence S, is known as LBKeogh Han et al SIGMOD 2011

  • We used the Piecewise Aggregate Approximation (PAA) descriptor  to get equidistant time series representations.. .J. Bernard (2011)  Int J Digit Libr.

  • showed that their tight lower bound LB Keogh could also be used in rotation-invariant image matching and provided a novel solution for the DTW distance. Their solution is excellent.. Kim et all DEXA08


Keogh, E. and Kasetty, S. (2002). On the Need for Time Series Data Mining Benchmarks: A Survey and Empirical Demonstration. In the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. July 23 - 26, 2002. Edmonton, Alberta, Canada. pp 102-111. [pdf]  [Journal Version pdf]

Keogh, E., Lonardi, S and Chiu, W. (2002). Finding Surprising Patterns in a Time Series Database In Linear Time and Space. In the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. July 23 - 26, 2002. Edmonton, Alberta, Canada. pp 550-556. [Expanded Version pdf]

1.      ..we use SAX and Keogh's Tarzan algorithm to do anomaly detection in network traffic . Kyoji Umemura et. al.

2.      To discover the patterns in the behavior of state statistics, we use the Tarzan algorithm for analyzing time series.. Emre Kıcıman  

Keogh, E. & Pazzani, M. (1999). Learning augmented Bayesian classifiers: A comparison of distribution-based and classification-based approaches. In Uncertainty 99, 7th. Int'l Workshop on AI and Statistics, Ft. Lauderdale, FL, pp. 225--230. [ps]. 

1.      superparent provides considerable decrease in prediction error . Geoff Webb.

2.      superparent-one-dependence estimators have received a lot of attention because they offer a combination of high training efficiency, high classification efficiency and high classification accuracy. Y.Yang, et al. IEEE TKDE 2007.

3.      (in our extensive empirical comparisons) Superparent for structure learning produces the best performing network structures. Pernkopf, Pham and Bilmes. Speech Communication 2008.

4.      an ensemble of SuperParent one-dependence estimators  can deliver very high classification accuracy. Ying Yang et al Machine Leaning 2007.

5.      we propose two different approaches in order to deal directly with numeric attributes (one of them) keeps the superparent on each model discrete. Flores et al ICML 09

6.      to classify an instance, AAPE first computes NB and then utilizes any additional time to refine the probability estimates by invoking superparent-one-dependence estimators. Bei Hui et. al. Machine Learning

7.      similar to SuperParent, our algorithm finds one.. Zhang and Ling 2001

Keogh, E., Chu, S., Hart, D. & Pazzani, M. (2001). An Online Algorithm for Segmenting Time Series. In Proceedings of IEEE International Conference on Data Mining. pp 289-296. [ps, pdf]

  1. This amazing result is achieved due to two factors. The SWAB can reduce the size of patterns to about 20% of the original size  Tang et al. SIGMOD 07

  2. SWAB is incorporated into NAP's motion artifact detection.. T Fekete, et al. The NIRS Analysis Package: noise reduction and statistical inference. PLoS ONE 2012.

  3. The SWAB algorithm [Keogh et al. 2001] was used to partition the time series given a user-specified threshold.. It is attractive for solving our problem because of its simplicity. Yin et al. ACM Transactions on Sensor Networks. 2008

  4. in these results, we present the cost of performing model fitting, via an online segmentation-based algorithm (SWAB) to find a piecewise linear model to the input data .. is indeed able to support high-throughput stream processing. Ahmad et al ICDE'08

  5. For the construction of approximate rank synopses we adopt the SWAB algorithm that we found to perform well in practice. SPIRE 2007 Berberich, Bedathur, and Weikum

  6. For the motion segmentation the SWAB (Sliding window and bottom up) algorithm proposed by Keogh et al.  was used due to its excellent performance . Oliver Amft.

  7. (SWAB) was first proposed by Keogh.. we evaluate its benefits in the context of our BSearch algorithm.. (it allows) query processing that is 10X to 100X faster.  Arvind Thiagarajan MIT 2007

  8. Keoghs ICDM paper  provides a good overview of the process of time series segmentation. . Mark Moss

  9. To best serve our purpose without any modification, the bottom-up algorithm or SWAB is appropriate.  Jai-Jin Lim and Kang G. Shin

  10. The SWAB algorithm given by Keogh et al using sliding window, is particularly popular with the medical community, since patient monitoring is inherently an online task...with the help of SWAB , we separate the data stream..  Yu Fan et al

  11. especially Keogh, which presents a good survey of segmentation methods appearing in the data mining literature  Kehagias et al.

  12. For the gesture recognition of the parking problem, we used the SWAB  Bannach et al. Embedded Systems Lab, University of Passau

  13. SumTime-Mousam determines content using linear segmentation using Keoghs algorithm  Reiter et. al.

  14. We used the Sliding-Window And Bottom-up (SWAB) algorithm to detect and classify of normal swallowing in humans.. . Oliver Amft and Gerhard Troster 2006

  15. to capture the evolutionary history of the webs structure and content we use (Keoghs) close-to-optimal rank synopsis... Furthermore, (using SWAB) the close-to-optimal rank synopses can be maintained incrementally as new observations of the evolving Web graph become available.  Berberich, Bedathur and Weikum, SIGIR 2007.

  16. we use (SWAB) for generating textual summary of Neonatal intensive care data  Portet, Reiter, Hunter and Sripada AIME07

  17. We experimented with the (SWAB) online algorithm of Keogh et al.. to use them to detect epidemics.  Heino and Toivonen 2003.

  18. To begin with, dealing with large amounts of data requires a representation that would allow efficient computation. As a robust and efficient method for representing time series, Keogh's Piece-Wise Linear Segmentation (PWLS) was found to be a suitable method.. for a 600 day mission to Mars. Aydogan, Pekny and Orcun 2007

  19. For intensive care monitoring .. in the first step of the procedure, trend detection was performed on separate periods relying on a piecewise linear segmentation of the time series which was carried out by SWAB. Verduijn, Sacchi, Peek, Bellazzi, de Jonge, de Mol  Artificial Intelligence in Medicine 2007.

  20.  We adapt (SWAB).. to develop a linear-time algorithm that generates nearly-optimal temporal coalescing for the given error threshold. Berberich, Bedathur, Neumann and Weikum VLDB 2007

  21. The algorithm utilized for (recognizing gestures) is Sliding window and bottom-up (SWAB). Bannach, Amft, Kunze, Heinz, Troster and Lukowicz. CIG07

  22. Accordingly we have used a well-known segmentation algorithm from the KDD community known as SWAB (Keogh et al 2002).. Sripada, Reiter, Hunter and Yu. Proc. Corpus Linguistics 2003

  23. Using a user-defined error tolerance for approximation and an online-algorithm SWAB. the curve is represented by piecewise linear segments. This approximation significantly reduces the number of extracted features search.. Chen, Yau, Hansen, and Estrin 2007.

  24.  (to create) textual summaries of spatio-temporal data .. We use the linear segmentation algorithm (SWAB) developed by Keogh2001 and successfully applied in the SumTime system. Turner et al. 2007

  25. To adapt the characteristics of streaming data, we need an online piecewise linear approximation algorithm. The proposed SWAB algorithm not only takes the advantage of linear time and online fashion of sliding window method, but also.. Yeh, Dai, and Chen, TKDE 2008.

  26.  We use a online segmentation-based algorithm (SWAB) ..  for processing continuous queries over models of continuous-time data.  Ahmad, Papaemmanouil, Cetintemel, Rogers ICDE08

  27.  ..since all financial time series contain high levels of noise, we need to smoothen the time series by using SWAB.. Wu et al. APWeb 2008

  28.  The gestures potting procedure uses an explicit time series segmentation algorithm (SWAB) for activity recognition.  Bannach, D. Lukowicz, P. Amft, O.  Pervasive Computing, IEEE

  29. The (SWAB) segmentation algorithm which was developed by Keogh et al. was used to segment the time series signals for fatigue analysis.. Nopiah et, a. SIP-08

  30.  we used (SWAB) introduced by Keogh ..based on the evaluation of typical test data, we found the algorithm to be well suited for our application. Junker et al. Pattern Recognition 2008

  31. After experimenting with a vector quantization method, we switched to a shifting-window piecewise linear segmentation method (SWAB). Parncutt and Hair 2008

  32. While taking the cost and accuracy into consideration, we implement SWAB to simplify the data points in each section. Zhou et al 2008

  33. For mechanical fatigue data.. SWAB segmentation algorithm has proven to be the best at performing batch segmentation with the least amount of error. Nopoah et al 2008

  34.  Finally, the module looks for long-term trends in the data.. this is done using Keoghs SWAB algorithm. Portet,  Reiter, Gatt, Hunter,  Sripada, Freer, Sykes Artificial Intelligence. 2009.

  35. After experimenting with a vector quantization method we switched to SWAB. Parncutt and Hair 2008.

  36. For trend abstractions, we segment and label the series by using the sliding window (SWAB) method (Keogh et al. 2003) Batal et al 2009

  37. Key to our approach is the approximation of time series the approximation algorithm is based on SWAB. Laerhoven 2009

  38.  first divides a sequence into S segments (S is a user specified parameter) using bottom-up segmentation algorithm, SWAB. Wu et al 2009

  39. The segmentation of the resultant acceleration (of gait) was performed with (SWAB) as described by Keogh et al. A. Sant'Anna, et. al. Gerontechnology 2008

  40. (for our robotics problem) the raw data is read at 100 Hertz and is approximated by linear segments using a version of SWAB. Berlin et al. 2010

  41. we applied the SWAB method (to allow) Top-k Queries on Temporal Data. Li,  Yi, Le VLDBJ 2001

  42. The resultant acceleration was segmented according to a bottom-up piecewise linear segmentation algorithm (SWAB). Anita Sant' Anna and Nicholas Wickstrom 2010

  43. we discover change points with SWAB, as a subroutine in finding gesture patterns with applications to robot-human interactions. Okada, Izukura and Nishida 2011

  44. After extracting representative motion flows, we adopt SWAB to segment a motion flow into multiple segments Zhao et. al. Multimedi Tools 2011

  45. we use SWAB for an application of knowledge discovery methods in diagnostics of standard radio frequency generator .. Miczulski, W. Szulim, R. 2009

  46. To deal with data noise and avoid the edge-blurring problem from filtering (for Spatiotemporal Segmentation) we use (BU) algorithm (from Keogh) Wang and Yu IEEE Geoscience

  47. The algorithm proposed in this paper is a modification of SWAB..  Van Laerhovem, SenseApp 2010 ,

  48. we discuss how to design an online segmentation algorithm with manually fixed features. We extend the SWAB...  Aberer 2012


Keogh, E. & Pazzani, M. (1999).  Scaling up dynamic time warping to massive datasets. In Proceedings of the 3rd European Conference on Principles and Practice of Knowledge Discovery in Databases. pp 1-11. [ps, pdf].

Keogh, E. & Pazzani,M (1999). Relevance feedback retrieval of time series data . In Proceedings of the 22th Annual International ACM-SIGIR Conference on Research and Development in Information Retrieval. pp 183-190. [ps, pdf]

Keogh, E. & Pazzani, M. (2000). A simple dimensionality reduction technique for fast similarity search in large time series databases.  In Proceedings of Pacific- Asia Conf. on Knowledge Discovery and Data Mining, pp 122-133. [ps, pdf]

Keogh, E.,  Chakrabarti, K., Pazzani, M. & Mehrotra, S. (2001). Locally adaptive dimensionality reduction for indexing large time series databases. In proceedings of ACM SIGMOD Conference on Management of Data, May. pp 151-162  [PDF] According to iFuice (Prof. Dr. Erhard Rahm) This is one of the top 5 most referenced papers of SIGMOD 2001.

1. In our work, we use the APCA approximation with the distance measures proposed by Keogh et al. Our results indicated APCA it increases the precision of the index without any adverse affect on the search performance which translates to fewer disk I/Os.  Kadiyala and Shiri, KAIS 2007

2. coupled with the state of the art dimensionality reduction techniques such as Adaptive Piecewise Constant Approximation (APCA) and improve its performance by up to a factor of 3. Li, Lopez and Moon TKDE 2004

3. Our (indexing) technique is based on the segmentation of each sequence into an Adaptive Piecewise Constant Approximation (APCA). SHOU et al, ML Journal 2005

Keogh, E.,  Chakrabarti, K., Pazzani, M. & Mehrotra, S. (2000). Dimensionality Reduction for Fast Similarity Search in Large Time Series Databases. Knowledge and Information Systems 3(3): 263-286. [ps, pdf]

  1. More specifically, to be able to support all types of search queries listed above, we use the Piecewise Aggregate Approximation (PAA) . A.H.R. Albers et al

  2.  The technique proposed in Piecewise Aggregate Approximation (or PAA) overcomes this drawback while bearing same or better time complexity compared to existing methods. Das, Bhaduri and Kargupta 2009

  3. Furthermore, we propose an efficient indexing method to retrieve similar trajectories for a query by combining a spatial indexing technique and PAA Yanagisawa et al

  4. (our systems uses) the Piecewise Aggregate Approximation (PAA). This results in a more compact description of the trajectory and enables fast search-queries.  Jaspers  et. al (Bosch Security Systems)

  5. we observe that the accuracy of the PAA method is quite similar to the other method that stores the actual vectors, while the storage of PAA is much smaller. Georgoulas and Kotidis 2012

  6. inspired by the Piecewise Aggregate Approximation (PAA) technique introduced .. Pelekis et al. 2009

  7. By taking advantage of Piecewise Aggregate Approximation (PAA) we significantly speed up our algorithm both in the preprocessing and segmentation stage  Zifan et al. 2007

  8. PAA results in a more compact description of the trajectory...   E. Jaspers

  9. Using PAA based approach the search time is reduced by approx an order of magnitude.  Austin (2006) AURA technology project.

  10. Motivated by Keogh and Pazzanis PAA we have present a Karaoke music retrieval system that allows users to locate their desired music by singing to the system  Hung-Ming Yu et al

  11. (we use) LB_Keogh with PAA for efficient similarity range query processing in music databases  Ruxanda & Jensen 2006

  12. We propose to use the piecewise aggregate approximation (PAA) to solve this problem.  Ting Wang 2006:

  13. PAA is a powerful compression tool  Wilson, Birkin, and Aickelin 2007

  14. many linear transformations have been tested but PAA has been shown to be the best dimensionality reduction technique DTW queries. Thus we use PAA to..   Assent, Krieger, Afschari and Seidl EDBT 2008

  15. Keoghs PAA method was used in this study (on) a telemedicine application using an internet ECG database for risk stratification of patients with various cardiovascular disease state. Khoor et al. 2007

  16. (to support our system for Target Classification) each reference signal is compressed to the length of input signal using the PAA technique. Kim, Kim, Kim, Sung, and Yoo 2007.

  17. telemedicine application using an internet ECG database for risk stratification of patients with various cardiovascular disease state.. the PAA method was used in this study... Khoor et al. 2008

  18. We propose to accomplish this by using Piecewise Aggregate Approximation (PAA) .. PAA is simple, fast, and can still be applied in an online setting. Quanz and Tsatsoulis 2008

  19. our solution will make use of PAA.. Nguyen and Shiri CIKM 08

  20. In order to use the learning algorithms for anomaly detection, it is first necessary to extract some type of features from the streaming sensor data. We use the simple but effective approach of piecewise aggregate approximation (PAA). Quanz et al ICCCN 2007

  21. To accelerate the process .. it is possible to compress the data by a method called piecewise aggregate approximation (PAA). Lipowsky et al 2009

  22. To reduce the size of the data-set, the time-series (per ship) are averaged with contiguous, non-overlapping windows of length w. This method is called piecewise aggregate approximation.. de Vriesa et al 2009.

  23. promising work in recent years on the discovery of motifs directly in time-series signals .. using piecewise aggregate approximation.. Zeeshan Syed TKDD 2010

  24. We use this PAA definition to obtain the index-level lower bound. Yang-Sae Moon et al 2010

  25. To eliminate non-relevant cyclic patterns, we use a technique called Piecewise Aggregate Approximation (PAA) .  Rossi, Russo, and Succi

  26.  In this research, which aims at scalability, we consider PAA with higher scalability.. Lee, Roh, Hwang and Kim. FSE 2010

  27. Numerical Time-Series Pattern Extraction Based on Irregular Piecewise Aggregate Approximation (PAA) and Gradient Specification, M.Ohsaki, H. Abe and T.Yamaguchi: (2007)

  28. leaf entries are sorted in one-dimensional distance space using LBPAA. Han et al. SIGMOD 2011

  29. The simplest, but powerful segmentation technique for univariate time series is Piecewise Aggregate Approximation. Dobos et al 2008

  30. A Piecewise Aggregate Approximation (PAA) Lower-Bound Estimate for Posteriorgram-based Dynamic Time Warping., Zhang and Glass, Interspeech 2011.

  31.  Piecewise cloud approximation for time series mining. Hailin Li, Chonghui Guo:Knowl.-Based Syst. 24(4): 492-500 (2011)

  32. An Improved Piecewise Aggregate Approximation (PAA) Based on Statistical Features for Time Series Mining. Chonghui Guo, Hailin Li, Donghua Pan: KSEM 2010: 234-244

  33. First, each window was transformed in a piecewise aggregate approximation (PAA) representaion (for context predictions) .. Voigtmann et al CoMoRea 2011

  34. piecewise aggregate approximation (PAA) will be used to smooth the original accelerometer values and fingertip travel distances .. Liping Wu 2011

  35. For the purpose of preserving distance orders, we use the noise averaging effect of piecewise aggregate approximation (PAA)    Moon, Kim, Kim and Bertino

  36. In this work, we opted to use PAA as it could be directly applied without further processing of the real world equipment data available for this project. Yang and Letourneau 2011

  37. First, we convert the pattern detection problem to a pattern matching problem. A mature symbolic representation algorithm -SAX  - used for numerous data mining tasks is employed to convert sequences of numeric sensor observations to character string.  Zoumboulakis and Roussos 2011

  38. in this paper, we propose an improved lower-bound estimate using piecewise aggregate approximation (PAA)..  Zhang and Glass, INTERSPEECH 2011

  39. The PAA dimensionality reduction is intuitive and simple, yet has been shown to rival more sophisticated dimensionality reduction techniques. Kamam and Thakore 2012

  40. A Piecewise Aggregate Approximation (PAA) was used to abstract the data..  Bhatkar et al. BioMedical Engineering OnLine 2010

  41. To support large-scale image databases.. we present a concept of image-PAA features.. Lee et al WASET 2011. 

  42. Using the distribution of angles of the adjacent piecewise aggregate approximation (PAA) coefficients...Hiroki Ashida et al. Nucl. Acids Res. (2012)

  43. The data is compressed using the Piece-wise Aggregate Approximation (PAA) technique  that is simple enough to compute even on a microcontroller. Ganu et al. e-Energy 2012

Chu, S., Keogh, E., Hart, D. & Pazzani, M.  (2002). Iterative Deepening Dynamic Time Warping.  In Second SIAM International Conference on Data Mining. [ps, pdf]

Keogh, E. & Pazzani, M. (2001). Dynamic Time Warping with Higher Order Features. In First SIAM International Conference on Data Mining (SDM'2001), Chicago, USA. [pdf]

  1. Using Keoghs  DDTW on Electrocardiogram segmentation is robust enough to give measures comparable to those given by experts.  Zifan et al.

  2. the results obtained by the DDTW measure are very promising when compared with the classical feature extraction.. Miroslav Bursa and Lenka Lhotska 2007

  3. The results on DDTW, obtained by using only one sensor channel on the shin showed an average recognition score of 95%, higher than the values obtained with DTW. Rossana Muscillo, IEEE EMBS 2007.

  4. The similarity between two faces is then be computed using a modification of Derivative Dynamic Time Warping (Keogh & Pazzani, 2001) between the two dynamic signals. Our preliminary results provide evidence that the dynamics of even very short facial actions contain sufficient information for identity recognition. 2009 Nicholas Costen, Psychological and computational perspectives on recognition of moving faces Symposium

  5. The right distance to use is the Derivative Dynamic Time Warping. DDTW correctly found the six high-level groups, but Euclidean distance did not. Ferraretti et al 2009

  6. The approach we take is similar to DTW, particularly the Derivative DTW (DDTW) . Chan, James Bailey and Leckie 2009

  7. we have extended Keogh and Pazzani's (derivative dynamic time warping) measure.  Marek Kulbacki

  8. we use Keogh's DDTW  for predicting the quality of a batch process using an inferential sensor... G. Gins, I.Y. Smets and J.F. Van Impe

  9. We use the Derivative Dynamic Time Warping (DDTW) method of Keogh for matching pairs of points between two subsequences..  Walter P. Schiefele 

  10. To help make training series more consistent and to ultimately merge these training series into one (generalize), derivative dynamic time warping (DDTW) is employed..  Sophia Antipolis

  11. The DTW cost function was based on derivative spectral features, as advocated generally by Keogh   Simon Dixon.

  12.  {Translated from Polish} with Derivative Dynamic Time Warping we got the best accuracy  Khalid Saeed.

  13. This approach especially requires to define a similarity measure appropriate to the comparison of multivariate time-series of different lengths, so that events and states are properly identified and clustered. We use the Derivative Dynamic Time Warping (DDTW) method . Alberto Avanzi et al.

  14.  In this work, characteristic voltage has been used to apply the DDTW algorithm in order to find similarity criteria among of a set of sag registers... Llanos et al.

  15.  One of them, derivative dynamic time warping, is used in our program which verificates signatures by computing distance between curves or shapes of signatures.  Algirdas Bastys

  16.  we use the DDTW to establish the correspondence of crucial points.  Quan and Ji.

  17. We use the Derivative Dynamic Time Warping (DDTW) method for.. ...decision-making in surveillance systems  Monique Thonnat.

  18. For Multi-Dimensional Dynamic TimeWarping for Gesture Recognition Derivative DTW gave the better performance.  G.A. ten Holt , M.J.T. Reinders , E.A. Hendriks (2007)

  19.  Derivative Dynamic Time Warping (DDTW) has been proved to be useful in time series alignment to avoid singularity points and reduce the bias of alignment results...  Zhang, Yang and Edgar, Thomas 2008

  20.  The correspondence among similar shapes of two signals may be more robustly captured using the derivative DTW Grisan, Tiso and Ruggeri 2009

  21. To deal with varying movement velocity profiles, we apply derivative dynamic time warping (DDTW). Eppner et al 2009

  22. The approach we take is similar to DTW, particularly the Derivative DTW (DDTW). Chan, Bailey and  Leckie 2009

  23. The novelty of DDTW is that local derivatives of the data points are estimated to capture information on the trends in the sequences and to find a warping more robust to singularities. Gullo,et al 2009

  24. .. uses Derivative Dynamic Time Warping (DDTW) for time series pattern matching. .. has inherent advantages for satellite image classification as land use practices such as field crops/vegetation exhibit temporal shifts from pixel to pixel based on the geographic location. Gupta, K.S.Rajan 2009.

  25.  In contrast, DDTW, (is) able to effectively align points of interest in gait data trajectories, regardless of intensity differences. Helwig et al. 2009

  26.  The similarity between two faces is then be computed using a modification of Derivative Dynamic Time Warping between the two dynamic signals.  Paul Rosin and David Marshall 2009

  27.   To make the alignments more robust, we compute (DDTW) effectively incorporating information about the shape of the trajectories being aligned.  Daniel Meyer-Delius et al  IROS'09

  28. In (our experiments on Persian Signature Verification) it was observed that Derivative DTW method leads to higher accuracy and better performance. Zoghi1 and  Abolghasemi 2009.

  29. The figure shows that the speed estimated by DDTW matches the actual speed very well... Chandrasekaran et al PERCOM 2011

  30. However, in this paper, only DDTW is described because it is used as the classifier of our systems..Mokhtar, Arof and Iwahash 2010

  31. We prefer Derivative Dynamic Time Warping (DDTW) to DTW.... Ganeshapillai MIT Masters Thesis 2011

  32.  Derivative Dynamic Time Warping (DDTW) is chosen as the classifier for this experiment since it can..   Mokhtar, Arof and Iwahashi 2010. Sci. Res. Essays

  33. DDTW produces superior alignment than the classic dynamic time warping (DTW) algorithm. Therefore, only DDTW is considered in this study.  He. et al. Cancer Inform. 2011; 10: 65-82.

  34. In order to be independent of translations along the y axis, we used a variant of DTW called Derivative Dynamic Time Warping.  Gasser, Flexer and Grill SMC 2011.

  35. To increase awareness and promote driver  safety, we are proposing a novel system that uses Derivative Dynamic Time Warping (DDTW) .. D Johnson 2011

  36. An alignment technique .. commonly known as Derivative Dynamic Time Warping has been used in this paper as time-alignment method and a segmentation method. Avitia et al. PAHCE 2010

  37. We compared the average alignment precision of the top 50 peaks for seven different methods... DDTW performed the best.  He, Wang, Mobley, Richman and Grizzle.  Cancer Informatics 2011.

  38. chromatographic peak alignment using Derivative Dynamic Time Warping. Bork IFPAC 2011

  39. In this paper, we present a novel Derivative Dynamic Time Warping (DDTW) based method for querying desired songs in Hindi.  Prakhar K. Jai, al 2011 ALP

  40. For alignment, the derivative dynamic time warping algorithm was employed..   Zdenek Hanzlicek.  URE AVCR, pp. 90-97 (2006)

  41. (DDTW is) a very robust distance measure for... Automatic sleep stage scoring and apnea-hypopnea detection.  Schluter and Conrad 2012

  42. DDTW has been shown to yield superior results compared to DTW when the evolution of a profile is more important that the actual measurement value.  Gins et al Ind. Eng. Chem. Res 2012


Keogh, E. & Pazzani,M.(2000). Scaling Up Dynamic Time Warping for Data Mining Applications. In proceedings of the 6th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, August 20-23. Boston, MA, USA. pp 285-289. [pdf]

Keogh, E. and Pazzani, M. (1998). An enhanced representation of time series which allows fast and accurate classification clustering and relevance feedback. In 4th International Conference on Knowledge Discovery and Data Mining. New York, NY, Aug 27-31. pp 239-243. [ps, pdf]

Keogh, E. & Smyth, P. (1997). A probabilistic approach to fast pattern matching in time series databases. In Proceedings of the 3rd International Conference of Knowledge Discovery and Data Mining. pp 24-20. Runner up, best paper award.  Also appears as: Clustering and Mode Classification of Engineering Time Series Data. JPL Technical report 960621. [ps]

Keogh, E. An Overview of the Science of Fingerprints. Anil Aggrawal's Journal of Forensic Medicine and Toxicology, 2001; Vol. 2, No. 1 (January-June 2001): Published January 8, 2001