Learning Shapelets

This example illustrates how the “Learning Shapelets” method, presented in [1], can quicklyfind a set of shapelets that results in excellent predictive performance when used for a shapelet transform.

References

[1]

J. Grabocka et al. Learning Time-Series Shapelets. SIGKDD 2014.

• 
• 

Correct classification rate: 1.0

# Author: Romain Tavenard
# License: BSD 3 clause

import os

# Should be set before importing keras
os.environ["KERAS_BACKEND"] = "torch"

from keras.optimizers import Adam
import numpy
from sklearn.metrics import accuracy_score
import matplotlib.pyplot as plt

from tslearn.datasets import CachedDatasets
from tslearn.preprocessing import TimeSeriesScalerMinMax
from tslearn.shapelets import LearningShapelets, \
    grabocka_params_to_shapelet_size_dict
from tslearn.utils import ts_size

# Load the Trace dataset
X_train, y_train, X_test, y_test = CachedDatasets().load_dataset("Trace")

# Normalize each of the timeseries in the Trace dataset
X_train = TimeSeriesScalerMinMax().fit_transform(X_train)
X_test = TimeSeriesScalerMinMax().fit_transform(X_test)

# Get statistics of the dataset
n_ts, ts_sz = X_train.shape[:2]
n_classes = len(set(y_train))

# Set the number of shapelets per size as done in the original paper
shapelet_sizes = grabocka_params_to_shapelet_size_dict(n_ts=n_ts,
                                                       ts_sz=ts_sz,
                                                       n_classes=n_classes,
                                                       l=0.1,
                                                       r=1)

# Define the model using parameters provided by the authors (except that we
# use fewer iterations here)
shp_clf = LearningShapelets(n_shapelets_per_size=shapelet_sizes,
                            optimizer=Adam(0.01),
                            batch_size=16,
                            weight_regularizer=0.01,
                            max_iter=200,
                            random_state=0,
                            verbose=0)
shp_clf.fit(X_train, y_train)

# Make predictions and calculate accuracy score
pred_labels = shp_clf.predict(X_test)
print("Correct classification rate:", accuracy_score(y_test, pred_labels))

# Plot the different discovered shapelets
plt.figure()
for i, sz in enumerate(shapelet_sizes.keys()):
    plt.subplot(len(shapelet_sizes), 1, i + 1)
    plt.title("%d shapelets of size %d" % (shapelet_sizes[sz], sz))
    for shp in shp_clf.shapelets_:
        if ts_size(shp) == sz:
            plt.plot(shp.ravel())
    plt.xlim([0, max(shapelet_sizes.keys()) - 1])

plt.tight_layout()
plt.show()

# The loss history is accessible via the `model_` that is a keras model
plt.figure()
plt.plot(numpy.arange(1, shp_clf.n_iter_ + 1), shp_clf.history_["loss"])
plt.title("Evolution of cross-entropy loss during training")
plt.xlabel("Epochs")
plt.show()