Feature extraction with tsfresh transformer

tsfresh is a tool for extacting summary features from a collection of time series. It is an unsupervised transformation, and as such can easily be used as a pipeline stage in classification, clustering and regression in conjunction with a scikit-learn compatible estimator.

Preliminaries

You have to install tsfresh if you haven’t already. To install it, uncomment the cell below:

# !pip install --upgrade tsfresh

from sklearn.ensemble import RandomForestClassifier
from sklearn.pipeline import make_pipeline

from aeon.datasets import load_arrow_head, load_basic_motions
from aeon.transformations.collection.feature_based import TSFresh, TSFreshRelevant

Example data set

We use the ArrowHead data from the UCR TSC archive. as an example dataset. See dataset notebook for more details. We only use the first few cases for examples to speed up the notebook.

X, y = load_arrow_head()
n_cases = 24
X_train = X[:n_cases, :, :]
y_train = y[:n_cases]
X_test = X[n_cases : 2 * n_cases, :, :]
y_test = y[n_cases : 2 * n_cases]
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)

(24, 1, 251) (24,) (24, 1, 251) (24,)

Using tsfresh to extract features

There are two versions of TSFresh feature extractors wrapped in aeon. The first is the unsupervised TSFresh which by default extracts all 4662 features. See the documentation for parameter configuration.

t = TSFresh()
Xt = t.fit_transform(X_train)
Xt2 = t.transform(X_test)
print(f"Train shape = {Xt.shape} test shape = {Xt2.shape}")

Train shape = (24, 777) test shape = (24, 777)

The second is TSFreshRelevant which uses y to select the most relevant features.

t = TSFreshRelevant()
t.fit(X_train, y_train)
Xt = t.transform(X_test)
Xt.shape

(24, 75)

Using tsfresh with scikit estimators

You can use the tsfresh transformer with any scikit-learn compatible estimator.

classifier = make_pipeline(
    TSFresh(default_fc_parameters="efficient", show_warnings=False),
    RandomForestClassifier(),
)
classifier.fit(X_train, y_train)
classifier.score(X_test, y_test)

0.625

For convenience and consistency of use we also have hard coded TSFresh classifier, regressor and clusterer.

from aeon.classification.feature_based import TSFreshClassifier
from aeon.clustering.feature_based import TSFreshClusterer

cls = TSFreshClassifier(relevant_feature_extractor=False)
clst = TSFreshClusterer(n_clusters=2)

cls.fit(X_train, y_train)
cls.score(X_test, y_test)
clst.fit(X_train)
print(cls.predict(X_test))
print(clst.predict(X_test))

['0' '1' '0' '1' '1' '2' '0' '1' '1' '0' '1' '1' '0' '2' '0' '0' '0' '2'
 '2' '1' '0' '0' '0' '0']
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0]

By default, the TSFreshClassifier uses the supervised TSFreshRelevant and the scitkit RandomForestClassifier. You can change this through the constructor

from aeon.classification.sklearn import RotationForestClassifier

cls = TSFreshClassifier(estimator=RotationForestClassifier(n_estimators=5))
cls.fit(X_train, y_train)
cls.score(X_test, y_test)

0.5833333333333334

By default, the TSFreshClusterer uses the unsupervised TSFresh and the sklearn clusterer KMeans with default parameters (which fits 8 clusters). You can also configure this through the constructor.

from sklearn.cluster import KMeans

clst = TSFreshClusterer(estimator=KMeans(n_clusters=3))
clst.fit(X_train)
print(clst.predict(X_test))

[1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 0 2 0 0 1]

The TSFreshRegressor uses the supervised TSFreshRelevant and the scitkit RandomForestRegressor.

from aeon.regression.feature_based import TSFreshRegressor

reg = TSFreshRegressor(relevant_feature_extractor=False)
from aeon.datasets import load_covid_3month

X, y = load_covid_3month(split="train")
reg.fit(X, y)

TSFreshRegressor(relevant_feature_extractor=False)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

TSFresh with multivariate time series data

TSFresh transformers and all three estimators can be used with multivariate time series. The transform calculates the features on each channel independently then concatenate the results. The full transform creates 777*n_channels features.

X_train, y_train = load_basic_motions(split="train")
X_test, y_test = load_basic_motions(split="test")
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)

(40, 6, 100) (40,) (40, 6, 100) (40,)

tsfresh = TSFresh()
X = tsfresh.fit_transform(X_train, y_train)
X.shape

(40, 4662)

---

Generated using nbsphinx. The Jupyter notebook can be found here.