A Python implementation of GRAIL, a generic framework to learn compact time series representations.

• Python 3.6+
• numpy
• scipy
• tslearn

Installation using pip:

pip install grailts

To install from the source:

python setup.py install

Karhan Kayan ([email protected])

Here is an example where we load a UCR dataset and run approximate k-nearest neighbors on its GRAIL representations:

from GRAIL.TimeSeries import TimeSeries
from GRAIL.Representation import GRAIL
from GRAIL.kNN import kNN

TRAIN, train_labels = TimeSeries.load("ECG200_TRAIN", "UCR")
TEST, test_labels = TimeSeries.load("ECG200_TEST", "UCR")

representation = GRAIL(kernel="SINK", d = 100, gamma = 5)
repTRAIN, repTEST = representation.get_rep_train_test(TRAIN, TEST, exact=True)
neighbors, _, _ = kNN(repTRAIN, repTEST, method="ED", k=5, representation=None,
                              pq_method='opq')

print(neighbors)

To load UCR type datasets:

TRAIN, train_labels = TimeSeries.load("ECG200_TRAIN", "UCR")
TEST, test_labels = TimeSeries.load("ECG200_TEST", "UCR")

In this package, we assume that each row of the datasets is a time series.

To fetch exact GRAIL representations of a training and a test dataset:

representation = GRAIL(kernel="SINK", d = 100, gamma = 5)
repTRAIN, repTEST = representation.get_rep_train_test(TRAIN, TEST, exact=True)

Here d specifies the number of landmark series, and gamma specifies the hyperparameter used for the SINK kernel. If gamma is not specified, it will be tuned by the algorithm.

If a single dataset is used instead:

repX = representation.get_representation(X)

To get the approximate k-Nearest-Neighbors of TEST in TRAIN use:

neighbors, correlations, return_time = kNN(repTRAIN, repTEST, method="ED", k=5, representation=None,
                              pq_method='opq')

Note that Euclidean Distance in the GRAIL representation space estimates the SINK correlation in the original space.