Time Series Forest

This example illustrates which information is considered important by the algorithm in order to classify time series. The index of the most important window is retrieved via the feature_importance_ and indices_ attributes. The first time series for both classes are plotted and the most important window is highlighted with a larger line width.

It is implemented as pyts.classification.TimeSeriesForest.

# Author: Johann Faouzi <johann.faouzi@gmail.com>
# License: BSD-3-Clause

import matplotlib.pyplot as plt
import numpy as np

from pyts.classification import TimeSeriesForest
from pyts.datasets import load_gunpoint

X_train, X_test, y_train, y_test = load_gunpoint(return_X_y=True)
clf = TimeSeriesForest(random_state=43)
clf.fit(X_train, y_train)

start_idxmax, end_idxmax = clf.indices_[np.argmax(clf.feature_importances_) // 3]

plt.figure(figsize=(12, 5))
plt.plot(X_train[y_train == 1][0], label='First sample in class 1')
plt.plot(
    np.arange(start_idxmax, end_idxmax),
    X_train[y_train == 1][0, start_idxmax:end_idxmax],
    color='C0',
    lw=4,
)

plt.plot(X_train[y_train == 2][0], label='First sample in class 2')
plt.plot(
    np.arange(start_idxmax, end_idxmax),
    X_train[y_train == 2][0, start_idxmax:end_idxmax],
    color='C1',
    lw=4,
)

plt.legend(loc='best', fontsize=14)
plt.title(
    'The most important window according to the feature importance scores',
    fontsize=16,
)
plt.tight_layout()
plt.show()