This repository contains a MODIFIED FORK of the tsfresh python package for use with Skyline (https://github.com/earthgecko/skyline).
This modified fork maintains the features extracted at tsfresh==v0.4.0 but moves this tsfresh version forward in line with blue-yonder/tsfresh in terms of tsfresh internals and dependencies, etc. The original methods for features are maintained even if they are changed in blue-yonder/tsfresh.
Required algorithms are maintained in terms of ensuring that any changes to required algorithms in terms
of parameter renames, e.g. unbiased=True changing to adjusted=True and ensuring that any changes made
to the tsfresh versions of the algorithm do not cause regressions in the values of the results of each feature
or the arbitrary names of the features.
This is achieved specifically using modified versions of:
- tsfresh/feature_extraction/feature_calculators.py - to maintain the required algorithms and disable new algorithms added to tsfresh
- tsfresh/feature_extraction/settings.py - to maintain which algorithms are run and disable the running of new algorithms added to tsfresh
The reason for this fork is because tsfresh is constantly adding new algorithms, removing algorithms, changing algorithm run parameters and changing algorithm and features names. For use in Skyline every version of tsfresh must calculate the same values and feature names for every feature (algorithm) on the same test timeseries. Further it must run the same algorithms in every version.
A features profile calculated with earthgecko/tsfresh v0.4.0 should be the same if it is generated earthgecko/tsfresh v0-17.9 or v0.19.1
Version mappings to which blueyonder/tsfresh the earthgecko/tsfresh version is based on: blueyonder/tsfresh-v0.17.1 -> earthgecko/tsfresh-v0-17.9 blueyonder/tsfresh-v0.20.0 -> earthgecko/tsfresh-v0-19.1
It must be noted that the results of the features calculations will change in decimal precision or decimal values
between Python versions, library versions and even machines. These changes are acceptable from a Skyline perspective and are
np.is_close close enough.
NOTE: these branches/versions are tested against the tests/baseline/tsfresh_features_test.py, which was removed from blue-yonder/tsfresh in v0.7.0 but has been readded to this fork. These branches/versions are only tested via the Skyline build tests, they are not tested against the tsfresh tests. Seeing as this fork follows the blue-yonder/tsfresh versions and retrospectively makes backwards compatible changes to the settings and feature_calculators.py which work with the Skyline tests. Therefore these changes are not currently backported to the tsfresh tests themselves and the tsfresh tests will fail if run against any of theses branches.
This repository contains the TSFRESH python package. The abbreviation stands for
"Time Series Feature extraction based on scalable hypothesis tests".
The package provides systematic time-series feature extraction by combining established algorithms from statistics, time-series analysis, signal processing, and nonlinear dynamics with a robust feature selection algorithm. In this context, the term time-series is interpreted in the broadest possible sense, such that any types of sampled data or even event sequences can be characterised.
Data Scientists often spend most of their time either cleaning data or building features. While we cannot change the first thing, the second can be automated. TSFRESH frees your time spent on building features by extracting them automatically. Hence, you have more time to study the newest deep learning paper, read hacker news or build better models.
TSFRESH automatically extracts 100s of features from time series. Those features describe basic characteristics of the time series such as the number of peaks, the average or maximal value or more complex features such as the time reversal symmetry statistic.
The set of features can then be used to construct statistical or machine learning models on the time series to be used for example in regression or classification tasks.
Time series often contain noise, redundancies or irrelevant information. As a result most of the extracted features will not be useful for the machine learning task at hand.
To avoid extracting irrelevant features, the TSFRESH package has a built-in filtering procedure. This filtering procedure evaluates the explaining power and importance of each characteristic for the regression or classification tasks at hand.
It is based on the well developed theory of hypothesis testing and uses a multiple test procedure. As a result the filtering process mathematically controls the percentage of irrelevant extracted features.
The TSFRESH package is described in the following open access paper:
- Christ, M., Braun, N., Neuffer, J., and Kempa-Liehr A.W. (2018). Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests (tsfresh -- A Python package). Neurocomputing 307, p. 72-77, doi: 10.1016/j.neucom.2018.03.067.
The FRESH algorithm is described in the following whitepaper:
- Christ, M., Kempa-Liehr, A.W., and Feindt, M. (2017). Distributed and parallel time series feature extraction for industrial big data applications. ArXiv e-print 1610.07717, https://arxiv.org/abs/1610.07717.
Systematic time-series feature extraction even works for unsupervised problems:
- Teh, H.Y., Wang, K.I-K., Kempa-Liehr, A.W. (2021). Expect the Unexpected: Unsupervised feature selection for automated sensor anomaly detection. IEEE Sensors Journal 15.16, p. 18033-18046, doi: 10.1109/JSEN.2021.3084970.
Due to the fact that tsfresh basically provides time-series feature extraction for free, you can now concentrate on engineering new time-series, like e.g. differences of signals from synchronous measurements, which provide even better time-series features:
-
Kempa-Liehr, A.W., Oram, J., Wong, A., Finch, M., Besier, T. (2020). Feature engineering workflow for activity recognition from synchronized inertial measurement units. In: Pattern Recognition. ACPR 2019. Ed. by M. Cree et al. Vol. 1180. Communications in Computer and Information Science (CCIS). Singapore: Springer, p. 223–231. doi: 10.1007/978-981-15-3651-9_20.
-
Simmons, S., Jarvis, L., Dempsey, D., Kempa-Liehr, A.W. (2021). Data Mining on Extremely Long Time-Series. In: 2021 International Conference on Data Mining Workshops (ICDMW). Ed. by B. Xue et al. Los Alamitos: IEEE, p. 1057-1066. doi: 10.1109/ICDMW53433.2021.00137.
Systematic time-series features engineering allows to work with time-series samples of different lengths, because every time-series is projected into a well-defined feature space. This approach allows the design of robust machine learning algorithms in applications with missing data.
- Kennedy, A., Gemma, N., Rattenbury, N., Kempa-Liehr, A.W. (2021). Modelling the projected separation of microlensing events using systematic time-series feature engineering. Astronomy and Computing 35.100460, p. 1–14, doi: 10.1016/j.ascom.2021.100460
Is your time-series classification problem imbalanced? There is a good chance that undersampling of time-series feature matrices might solve your problem:
- Dempsey, D.E., Cronin, S.J., Mei, S., Kempa-Liehr, A.W. (2020). Automatic precursor recognition and real-time forecasting of sudden explosive volcanic eruptions at Whakaari, New Zealand. Nature Communications 11.3562, p. 1-8, doi: 10.1038/s41467-020-17375-2.
Natural language processing of written texts is an example of applying systematic time-series feature engineering to event sequences, which is described in the following open access paper:
- Tang, Y., Blincoe, K., Kempa-Liehr, A.W. (2020). Enriching Feature Engineering for Short Text Samples by Language Time Series Analysis. EPJ Data Science 9.26, p. 1–59. doi: 10.1140/epjds/s13688-020-00244-9
TSFRESH has several selling points, for example
- it is field tested
- it is unit tested
- the filtering process is statistically/mathematically correct
- it has a comprehensive documentation
- it is compatible with sklearn, pandas and numpy
- it allows anyone to easily add their favorite features
- it both runs on your local machine or even on a cluster
If you are interested in the technical workings, go to see our comprehensive Read-The-Docs documentation at http://tsfresh.readthedocs.io.
The algorithm, especially the filtering part are also described in the paper mentioned above.
We appreciate any contributions, if you are interested in helping us to make TSFRESH the biggest archive of feature extraction methods in python, just head over to our How-To-Contribute instructions.
If you want to try out tsfresh quickly or if you want to integrate it into your workflow, we also have a docker image available:
docker pull nbraun/tsfresh
If you need to reproduce or update time-series features, which were computed with the matrixprofile feature calculators, you need to create a Python 3.8 environment:
conda create --name tsfresh__py_3.8 python=3.8
conda activate tsfresh__py_3.8
pip install tsfresh[matrixprofile]
The research and development of TSFRESH was funded in part by the German Federal Ministry of Education and Research under grant number 01IS14004 (project iPRODICT).