Transformer

SplitterBootstrapTransformer

Multivariate Missing values Unequal length

Splitter based Bootstrapping method for synthetic time series generation.

A generalized form of bootstrap based on an sktime splitter.

Any sktime splitter can be passed as a component to this transformer, which will then produce for each series in the input of transform a panel of time series with the train and/or test sub-series.

The output of transform will have additional levels:

all levels of the transform input
an additional integer indexed top level, indicating the number of the sample note: this is in general the number of the sample and corresponds to the number of the fold only in the deterministic, exhaustive case
if split="train" or split="test", no further levels
if split="both", the second top level contains strings "train" and

"test": to indicate train or test fold from the split

For instance, if split="train", and there is a single original series X, the output of transform will have a top level (level 0) with integer index ranging from 0 to splitter.get_n_splits(X)-1.

The splitter can be exhaustive and deterministic, or random. By default, exhaustive ordered samples are returned (deterministic). Randomness is controlled by the following parameters:

shuffle (by default off) applies random uniform shuffling to the instances
subsample (by default off) applies sub-sampling with or without replacement
replace (by default False) selects sub-sampling with or without

replacement

Caution: the instance index of the transform output will correspond to the split index only if shuffle=False and subsample=None (unless by coincidence)

Schnellstart

python

from sktime.transformations.bootstrap import SplitterBootstrapTransformer

estimator = SplitterBootstrapTransformer(splitter=None, fold='train', shuffle=False, subsample=None, replace=True, random_state=None)

Parameter(6)

splitteroptional, sktime splitter, BaseSplitter descendant: default = SlidingWindowSplitter(window_length=3, step_length=1) The splitter used for the bootstrap splitting.
foldstr, one of “train” (default), “test”, and “both”: Determines which fold is returned as new instances in the panel. “train” - the training folds; “test” - the test folds; “both” - both training and test folds, and an additional string level with possible values "train" and "test" is present
shufflebool, default=False: whether to shuffle the order of folds uniformly at random before returning if not, folds will be returned in the ordering defined by the splitter
subsampleoptional, int or float, default = None: if provided, subsamples the folds returned uniformly at random int = subsample of that size will be returned (or full sample if smaller) float, must be between 0 and 1 = subsample of that fraction is returned Note: integer 1 selects one series; float 1 selects number in splitter many
replacebool, default=True; only used if subsample=True: whether sampling, if subsample is provided is with or without replacement True = with replacement, False = without replacement
random_stateint, np.random.RandomState or None (default): Random seed for the estimator if None, numpy environment random seed is used if int, passed to numpy RandomState as seed if RandomState, will be used as random generator

Beispiele

>>> from sktime.transformations.bootstrap import SplitterBootstrapTransformer
>>> from sktime.datasets import load_airline
>>> y = load_airline ()
>>> transformer = SplitterBootstrapTransformer (fold = "both")
>>> y_hat = transformer. fit_transform (y)