The UCR Time Series Archive - introduced in 2002, has become an important resource in the time series data mining community, with at least one thousand published papers making use of at least one data set from the archive. The original incarnation of the archive had sixteen data sets but since that time, it has gone through periodic expansions. The last expansion took place in the summer of 2015 when the archive grew from 45 to 85 data sets. This paper introduces and will focus on the new data expansion from 85 to 128 data sets. Beyond expanding this valuable resource, this paper offers pragmatic advice to anyone who may wish to evaluate a new algorithm on the archive. Finally, this paper makes a novel and yet actionable claim: of the hundreds of papers that show an improvement over the standard baseline (1-nearest neighbor classification), a large fraction may be misattributing the reasons for their improvement. Moreover, they may have been able to achieve the same improvement with a much simpler modification, requiring just a single line of code.

This provides the UCR Time Series Anomaly Detection datasets [1] publicly available on this webpage. This dataset repository is created to ensure the version of the dataset used in the TimeVQVAE-AD paper [2].References[1] Keogh, Eamonn, et al. "Multi-dataset time-series anomaly detection competition." ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. https://compete. hexagonml. com/practice/competition/39. 2021.[2] Lee, Daesoo, Sara Malacarne, and Erlend Aune. "Explainable time series anomaly detection using masked latent generative modeling." Pattern Recognition (2024): 110826.

This is the modified UCR datasets used in the paper Time series classification for varying length series.

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

A Four class face outline problem, we are not sure of the exact provenance

Donator: A. Ratanamahatana, E. Keogh

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

The collection of this data was part of the project titled Personalised Retrofit Decision Support Tools for UK Homes using Smart Home Technology (REFIT). The REFIT dataset includes data from 20 households from Loughborough area over the period 2013-2014 (see [1]). We use data of freezers in House 1 to make two datasets. These two datasets share a same test set and only differ in the number of training instances. There are two classes, one representing the power demand of the fridge freezer in the kitchen, the other representing the power demand of the (less frequently used) freezer in the garage. They are hard to tell apart globally but they differ locally.

[1] Murray, David et al., "A data management platform for personalised real-time energy feedback", Proceedings of the 8th International Conference on Engery Efficiency in Domestic Appliances and Lighting, 2015.

Donator: REFIT project

UCR Time Series Archive is a collection of time series datasets from various domains. It is used for training and testing deep neural networks for time series classification tasks.

The UCR Archive 2018 provides 128 time series datasets with diverse sources. But there are two primary issues to be used to train a time series generative model. Firstly, a majority of the datasets have a larger test set compared to a training set. Secondly, there is clear difference in patterns between training and test sets for some of the datasets.To make the datasets from this archive more suitable for a time series generation task, we merged the existing training and test sets and resplit it using StratifiedShuffleSplit (from sklearn) into 80% and 20% for a training set and test set, respectively.

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

FacesAll is rotationally aligned version of FaceAll with a different train/test split

Donator: X. Xi, E. Keogh

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

The new generation Earth Observation (EO) satellites have been imaging the Earth frequently, completely and in high resolution. This introduces unprecedented opportunities to monitor the dynamics of any regions on our planet over time and revealing the constant flux that underpins the bigger picture of our World.

This dataset is a subset of a bigger dataset [1], which originally comes from 46 geometrically and radiometrically corrected images taken by FORMOSAT-2 satellite. These images are corrected such that every pixel corresponds to the same geographic area on Earth. Each pixel represents an area of 64 square meter; with 1 million pixels per image, this results in an area of 64 square kilometer each. Each geographic area (x, y) (∼pixel) forms a time series of length of 46, showing the temporal evolution of that area.

There are 24 classes corresponding to what the land covers. Class label - class name in French and English translation@

1. mais: corn
2. ble: wheat
3. bati dense: dense building
4. bati indu: built indu
5. bati diffus: diffuse building
6. prairie temporaire: temporary meadow
7. feuillus: hardwood
8. friche: wasteland
9. jachere: jachere
10. soja: soy
11. eau: water
12. pre: pre
13. resineux: softwood
14. tournesol: sunflower
15. sorgho: sorghum
16. eucalyptus: eucalyptus
17. colza: rapeseed
18. mais ensillage: but drilling
19. orge: barley
20. pois: peas
21. peupliers: poplars
22. surface minerale: mineral surface
23. graviere: gravel
24. lac: lake

There is nothing to infer from the order of examples in the train and test set.

Data created by: C.W. Tan, G.I. Webb and F. Petitjean (see [1], [2]). Data edited by Hoang Anh Dau.

[1] Tan, Chang Wei, Geoffrey I. Webb, and François Petitjean. "Indexing and classifying gigabytes of time series under time warping." Proceedings of the 2017 SIAM International Conference on Data Mining. Society for Industrial and Applied Mathematics, 2017.
[2] http://bit.ly/SDM2017

Donator: F. Petitjean

The UCR archive from the original archive. The only difference is that I replaced some datasets with varying length with the provided padded versions of them.

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

Caenorhabditis elegans is a roundworm commonly used as a model organism in the study of genetics. The movement of these worms is known to be a useful indicator for understanding behavioural genetics. Brown {\em et al.} "A dictionary of behavioral motifs reveals clusters of genes affecting Caenorhabditis elegans locomotion" describe a system for recording the motion of worms on an agar plate and measuring a range of human-defined features. It has been shown that the space of shapes Caenorhabditis elegans adopts on an agar plate can be represented by combinations of four base shapes, or eigenworms. Once the worm outline is extracted, each frame of worm motion can be captured by four scalars representing the amplitudes along each dimension when the shape is projected onto the four eigenworms.

The data relates to 258 traces of worms converted into four "eigenworm" series. The eigenworm data are lengths from 17984 to 100674 (sampled at 30 Hz, so from 10 minutes to 1 hour) and in four dimensions (eigwnworm 1 to 4). There are five classes:N2,goa-1,unc-1,unc-38 and un63. N2 is wildtype (i.e. normal) the other 4 are mutant strains. These datasets are the first dimension only (first eigenworm)

The problems Worms.arff and WormsTwoClass.arff are series of first eigenworm1 averaged down so that all series are lengths 900 (the single hour long series is discarded). This smoothing is likely to discard discriminatory information. The Yemini features obtains nearly 100\% accuracy, although we have not independently verified this.

we address the problem of classifying individual worms as wild-type or mutant based on the time series of the first eigenworm, down-sampled to second-long intervals. We have 257 cases, which we split 70\%/30\% into a train and test set. Each series has 900 observations, and each worm is classified as either wild-type (the N2 reference strain - 109 cases) or one of four mutant types: goa-1 (44 cases); unc-1 (35 cases); unc-38 (45 cases) and unc-63 (25 cases). The data were extracted from the {\em C. elegans} behavioural database~\cite{wormWeb}. The formatted classification problems are available from the website associated with this paper~\cite{tscWeb}.

Donator: A. Bagnall

Dataset Card for Time Series Extrinsic Regression

  Dataset Summary

A collection of datasets from Monash, UEA, and UCR supporting research into Time Series Extrinsic Regression (TSER), a regression task of which the aim is to learn the relationship between a time series and a continuous scalar variable. This task is closely related to time series classification, where a single categorical variable is learned. Please read the paper for more. If you use the results or code… See the full description on the dataset page: https://huggingface.co/datasets/foxy-steve/monash_uea_ucr_tser.

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

The data is outlines of four different types of cars extracted from traffic videos using the motion information. The images were mapped onto a 1-D series.
The vehicles were classified into one of the four class: sedan, pickup, minivan or SUV. Further details are available in [1].

[1] Thakoor, Ninad, and Gao, Jean. "Shape classifier based on generalized probabilistic descent method with hidden Markov descriptor." Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1. Vol. 1. IEEE, 2005. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.60.161&rep=rep1&type=pdf90.18%

Donator: J. Gao

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

The earthquake classification problem involves predicting whether a major event is about to occur based on the most recent readings in the surrounding area. The data is taken from Northern California Earthquake Data Center and each data is an averaged reading for one hour, with the first reading taken on Dec 1st 1967, the last in 2003. We transform this single time series into a classification problem by first defining a major event as any reading of over 5 on the Rictor scale. Major events are often followed by aftershocks. The physics of these are well understood and their detection is not the objective of this exercise. Hence we consider a positive case to be one where a major event is not preceded by another major event for at least 512 hours. To construct a negative case, we consider instances where there is a reading below 4 (to avoid blurring of the boundaries between major and non major events) that is preceded by at least 20 readings in the previous 512 hours that are non-zero (to avoid trivial negative cases). None of the cases overlap in time (i.e. we perform a segmentation rather than use a sliding window). Of the 86,066 hourly readings, we produce 368 negative cases and 93 positive.

Donator: A. Bagnall

Time-series datasets for classification tasks

Using the hctsa toolbox v0.97 (link in References below), we computed 7,500+ time-series features on each of the time-series classification tasks contained in the UEA/UCR Time Series Classification Repository. This repository provides the computed hctsa output files (.mat-files) for each classification task.We used the computed feature matrices to select a small subset of 22 hctsa estimators (termed catch22) that were the most useful for the UEA/UCR datasets:C.H. Lubba, S.S. Sethi, P. Knaute, S.R. Schultz, B.D. Fulcher, N.S. Jones. catch22: CAnonical Time-series CHaracteristics. arXiv (2019). https://arxiv.org/abs/1901.10200The matrices can be read in from Python as well using the Matlab_IO interface for which examples can be found in our selection pipeline for catch22 ("op_importance" in References) and in the "hctsaAnalysisPython" GitHub repository.

This contains all the datasets used for time series regression. Please refer to the website http://timeseriesregression.org/ for more details

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

The traffic data are collected with the loop sensor installed on ramp for the 101 North freeway in Los Angeles. This location is close to Dodgers Stadium; therefore the traffic is affected by volume of visitors to the stadium. Missing values are represented with NaN. - Class 1: Normal Day - Class 2: Game Day There is nothing to infer from the order of examples in the train and test set. Missing values are represented with NaN in the text file. Data created by Ihler, Alexander, Jon Hutchins, and Padhraic Smyth (see [1][2][3]). Data edited by Chin-Chia Michael Yeh.

[1] Ihler, Alexander, Jon Hutchins, and Padhraic Smyth. "Adaptive event detection with time-varying poisson processes." Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2006.

[2] “UCI Machine Learning Repository: Dodgers Loop Sensor Data Set.” UCI Machine Learning Repository, archive.ics.uci.edu/ml/datasets/dodgers+loop+sensor.

[3] “Caltrans PeMS.” Caltrans, pems.dot.ca.gov/.

Donator: C. Yeh

This dataset is part of the UCR Archive maintained by University of Southampton researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/.

This dataset is a remake of the famous GunPoint dataset released in 2003. We strive to mimic in every aspect the recording of the original GunPoint. The actors include one male and one female. They are the same actors who created the original GunPoint. We record two scenarios, Gun and Point (also known as Gun and NoGun). In each scenario, the actors aim at a eye-level target. The difference between Gun and Point is that for the Gun scenario, the actors hold a gun, and in the Point scenario, the actors point with just their fingers. A complete Gun action involves the actor moves hand from an initial rest position, points the gun at target, puts gun back to waist holster and then brings free hand to the initial rest position. Each complete action conforms to a five-second cycle. With 30fps, this translates into 150 frames per action. We extract the centroid of the hand from each frame and use its x-axis coordinate to form a time series. We refer to the old GunPoint as GunPoint 2003 and the new GunPoint as Gunpoint 2018. We merged GunPoint 2003 and GunPoint 2018 to make three datasets. Let us denote: - G: Gun - P: Point - M: Male - F: Female - 03: The year 2003 - 18: The year 2018 ## GunPointAgeSpan The task is to classify Gun and Point. There are 4 flavors of each class. - Class 1: Gun (FG03, MG03, FG18, MG18) - Class 2: Point (FP03, MP03, FP18, MP18) ## GunPointMaleVersusFemale The task is to classify Male and Female. There are 4 flavors of each class. - Class 1: Female (FG03, FP03, FG18, FP18) - Class 2: Male (MG03, MP03, MG18, MP18) ## GunPointOldVersusYoung The task is to classify the older and younger version of the actors. There are 4 flavors of each class. - Class 1: Young (FG03, MG03, FP03, MP03) - Class 2: Old (FG18, MG18, FP18, MP18) There is nothing to infer from the order of examples in the train and test set. Data created by Ann Ratanamahatana and Eamonn Keogh. Data edited by Hoang Anh Dau.

Donator: A. Ratanamahatana, E. Keogh

This dataset is part of the TSML dataset archive maintained by University of Southampton and University of Bradford researchers. Please cite a relevant or the latest full archive release if you use the datasets. See http://www.timeseriesclassification.com/ for information on the group and alternative download sources.

This dataset was originally part of the University of California, Riverside (UCR) time series classification archive.

The Automatic Diatom Identification and Classification (ADIAC) project was a pilot study concerning automatic identification of diatoms (unicellular algae) on the basis of images. The data was donated by Andrei Jalba, a PhD student on the project, which finished in the early 2000s. The outlines are extracted from thresholded images. Presumably the time series are generated as distance to a reference point (the centre being the obvious candidate). The data is very sinusoidal.

Donator: A. Jalba