Classifier

SignatureClassifier

Classification module using signature-based features.

This simply initialises the SignatureTransformer class which builds the feature extraction pipeline, then creates a new pipeline by appending a classifier after the feature extraction step.

The default parameters are set to best practice parameters found in: “A Generalised Signature Method for Multivariate TimeSeries” [1]

Note that the final classifier used on the UEA datasets involved tuning the hyper-parameters:

depth over [1, 2, 3, 4, 5, 6]
window_depth over [2, 3, 4]
RandomForestClassifier hyper-parameters.

as these were found to be the most dataset dependent hyper-parameters.

Thus, we recommend always tuning at least these parameters to any given dataset.

Schnellstart

python

from sktime.classification.feature_based import SignatureClassifier

estimator = SignatureClassifier(estimator=None, augmentation_list=('basepoint', 'addtime'), window_name='dyadic', window_depth=3, window_length=None, window_step=None, rescaling=None, sig_tfm='signature', depth=4, random_state=None)

Parameter(10)

estimatorsklearn estimator, default=RandomForestClassifier: This should be any sklearn-type estimator. Defaults to RandomForestClassifier.
augmentation_list: list of tuple of strings, default=(“basepoint”, “addtime”): List of augmentations to be applied before the signature transform is applied.
window_name: str, default=”dyadic”: The name of the window transform to apply.
window_depth: int, default=3: The depth of the dyadic window. (Active only if window_name == 'dyadic'].
window_length: int, default=None: The length of the sliding/expanding window. (Active only if `window_name in [‘sliding, ‘expanding’].
window_step: int, default=None: The step of the sliding/expanding window. (Active only if `window_name in [‘sliding, ‘expanding’].
rescaling: str, default=None: The method of signature rescaling.
sig_tfm: str, default=”signature”: String to specify the type of signature transform. One of: [‘signature’, ‘logsignature’]).
depth: int, default=4: Signature truncation depth.
random_state: int, default=None: Random state initialisation.

Referenzen

[1]

Morrill, James, et al. “A generalised signature method for multivariate time series feature extraction.” arXiv preprint arXiv:2006.00873 (2020). https://arxiv.org/pdf/2006.00873.pdf