A self-organizing, living library of time-series data

Fulcher, Ben D.; Lubba, Carl Henning; Sethi, Sarab S.; Jones, Nick

doi:10.1038/s41597-020-0553-0

Cited by 11 publications

(10 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…While some applications may be able to justify the computational expense of searching across a large feature library such as hctsa Jones 2014, 2017), the availability of an efficient, reduced set of features, as catch22, will make the advantages of feature-based time-series classification and clustering more widely accessible. As an example application catch22 is being used in the self-organizing time-series database for data-driven interdisciplinary collaboration CompEngine to assess the similarity of recordings (Fulcher et al 2019). Unlike the Matlab-based hctsa, catch22 does not require a commercial license to run, computes efficiently, and scales approximately linearly with time-series length in the cases we tested.…”

Section: Discussionmentioning

confidence: 99%

catch22: CAnonical Time-series CHaracteristics

Lubba

Sethi

Knaute

et al. 2019

Data Min Knowl Disc

Self Cite

238

148

View full text Add to dashboard Cite

Capturing the dynamical properties of time series concisely as interpretable feature vectors can enable efficient clustering and classification for time-series applications across science and industry. Selecting an appropriate feature-based representation of time series for a given application can be achieved through systematic comparison across a comprehensive time-series feature library, such as those in the hctsa toolbox. However, this approach is computationally expensive and involves evaluating many similar features, limiting the widespread adoption of feature-based representations of time series for real-world applications. In this work, we introduce a method to infer small sets of time-series features that (i) exhibit strong classification performance across a given collection of time-series problems, and (ii) are minimally redundant. Applying our method to a set of 93 time-series classification datasets (containing over 147,000 time series) and using a filtered version of the hctsa feature library (4791 features), we introduce a set of 22 CAnonical Time-series CHaracteristics, catch22, tailored to the dynamics typically encountered in time-series data-mining tasks. This dimensionality reduction, from 4791 to 22, is associated with an approximately 1000fold reduction in computation time and near linear scaling with time-series length, despite an average reduction in classification accuracy of just 7%. catch22 captures a diverse and interpretable signature of time series in terms of their properties, including linear and non-linear autocorrelation, successive differences, value distributions and outliers, and fluctuation scaling properties. We provide an efficient implementation of catch22, accessible from many programming environments, that facilitates feature-Responsible editor: Eamonn Keogh. Ben D. Fulcher and Nick S. Jones have contributed equally to this study.

show abstract

Section: Discussionmentioning

confidence: 99%

catch22: CAnonical Time-series CHaracteristics

Lubba

Sethi

Knaute

et al. 2019

Data Min Knowl Disc

Self Cite

238

148

View full text Add to dashboard Cite

show abstract

“…2. This diversity of dynamical patterns reflects the corresponding diversity in the types of real-world and generative models studied in science and industry [17].…”

Section: Evaluation Datasetmentioning

confidence: 91%

“…Having evaluated the differences in computation time of different feature sets, we next aimed to compare their behavior on real data, in order to assess the similarity of the constituent features, both within and between feature sets. To do this, we use the 'Empirical 1000' dataset (version 10) [16], which consists of a diverse set of 1000 real-world and modelgenerated time series [17]. Model-generated data spans a broad range of generative processes, including chaotic and nonchaotic deterministic dynamical systems (both discrete and continuous), and a range of stochastic processes.…”

Section: Evaluation Datasetmentioning

confidence: 99%

“…We then aimed to capture how similar features in the second set, labeled test, T , were to the features in the We assess time-series features based on their behavior on a dataset of 1000 diverse real-world and model-generated time series, a selection of which are plotted here [16], [17]. Six example time series from the Empirical 1000 dataset [16], [17] are plotted (after truncation to the first 1000 samples, as described in text): A audio of the Black-bellied Plover (AS_s2.1_f4_b8_l8280_135971), B the Trans Polar Index normalized pressure difference between Hobart and Stanley (CM_tpi_hob), C a simulation from the dissipative forced Brusselator ODE [18] (FL_fbruss_L300_N10000_IC_0.3_2_y), D a gait time series from the Parkinson's disease gait dataset [19], [20] (MD_gaitpdb_GaCo05_01l_SNIP_12562955), E river flow (ME_winter_rf), and F a simulation from the dissipative predator-prey map [18] (MP_predprey_L5000_IC_0.5_0.55_x). Time-series names are provided in parentheses for reference to their corresponding data entries on www.comp-engine.org [17].…”

Section: Between-set Overlap and Redundancymentioning

confidence: 99%

See 1 more Smart Citation

An Empirical Evaluation of Time-Series Feature Sets

Henderson¹,

Fulcher²

2021

Preprint

Self Cite

View full text Add to dashboard Cite

Solving time-series problems using informative features has been rising in popularity due to the availability of numerous software packages for time-series feature extraction. Feature-based time-series analysis can now be performed using any one of a range of time-series feature sets, including hctsa (7730 features: Matlab), feasts (42 features: R), tsfeatures (63 features: R), Kats (40 features: Python), tsfresh (up to 1558 features: Python), TSFEL (390 features: Python), and the C-coded catch22 (22 features, able to be run from Matlab, R, Python, and Julia). There is substantial overlap in the types of time-series analysis methods included in these feature sets (including properties of the autocorrelation function and Fourier power spectrum, and distributional shape statistics), but they are yet to be systematically compared. Here we compare these seven feature sets on their computational speed, assess the redundancy of features contained in each set, and evaluate the overlap and redundancy across different feature sets. We take an empirical approach to measuring feature similarity, based on the similarity of their outputs across a diverse set of real-world and modelsimulated time series. We find that feature sets vary across approximately three orders of magnitude in their computation time per feature on a laptop for a 1000-sample time series, from the fastest feature sets catch22 and TSFEL (∼ 0.1 ms per feature) to tsfeatures (∼ 3 s per feature). Using PCA to evaluate feature redundancy within each set, we find the highest within-set redundancy for TSFEL and tsfresh. For example, in TSFEL, 90% of the variance across 390 features can be captured with just four principal components. Finally, we introduce a metric for quantifying overlap between pairs of feature sets, which indicates substantial overlap between the feature sets. We found that the largest feature set, hctsa, is the most comprehensive, and that tsfresh is the most distinctive, due to its incorporation of large numbers of Fourier coefficients that are summarized at higher levels in the other sets. Our results provide empirical understanding of the differences between existing feature sets, information that can be used to better understand and tailor feature sets to their applications.Index Terms-time-series analysis, time-series features.

show abstract

“…That is, if we used a very large number of algorithms designed for general use in time-series, would we discover some that were more effective than our tailored design? This approach has been described by Fulcher and coworkers, who called it highly comparative time-series analysis [10][11][12] . The fundamental idea is to extract features from many time series, using many algorithms, most operating with many sets of parameter values.…”

Section: Introductionmentioning

confidence: 99%

Discovery of signatures of fatal neonatal illness in vital signs using highly comparative time-series analysis

Niestroy

et al. 2021

Preprint

View full text Add to dashboard Cite

Objective: Signatures of illness in vital signs of Neonatal Intensive Care Unit (NICU) patients can inform on future adverse events and outcomes. We implemented highly comparative time-series analysis to discover features and predictive analytics tools for all-cause mortality in the next 7 days, using the ubiquitous HR and SpO2 vital sign data from bedside monitors. Design: We populated a Time Series Commons with the complete HR and SpO2 data from all infants in the University of Virginia NICU from 2009 to 2019. We calculated the results of applying over 80 members of 11 mathematical families to random ten-minute segments of 0.5Hz data each day for each infant, with varying parameter sets, resulting in 4998 algorithmic operations on each infant. We used an unsupervised mutual information-based method to cluster the results, and we selected a single representative operation from each cluster. We used our FAIRSCAPE framework to compute a detailed provenance of all computations, and we constructed a complete software library with links to the analyzed data for reproducibility and reuse. We made multivariable logistic regression models using the lasso to assay the usefulness of the algorithms. Setting: Neonatal ICU. Patients: 5957 NICU infants, of whom 206 died. Measurements and main results: 3555 algorithmic operations returned usable results. Twenty representative operations, selected from each of 20 unsupervised clusters, held more than 81% of the information predicting death. A multivariable model had an AUC of 0.81 for predicting death in the next 7 days. In addition, five algorithms outperformed others: moving threshold, successive increases, surprise, and a random walk model. Conclusions: Highly comparative time-series analysis revealed new vital sign metrics to identify NICU patients at the highest risk of death in the next week. This approach can facilitate the discovery of signatures of impending, potentially actionable, clinical decompensation in monitored patients.

show abstract

A self-organizing, living library of time-series data

Cited by 11 publications

References 17 publications

catch22: CAnonical Time-series CHaracteristics

catch22: CAnonical Time-series CHaracteristics

An Empirical Evaluation of Time-Series Feature Sets

Discovery of signatures of fatal neonatal illness in vital signs using highly comparative time-series analysis

Contact Info

Product

Resources

About