Complete-case analysis

Dropping the incomplete samples is the honest alternative to imputing them: it introduces no synthetic data, at the cost of sample size and of a bias whenever the views are not missing completely at random.

Complete-case handling: discard samples instead of imputing them.

class polyview.imputation.drop.DropIncompleteSamples(*args: Any, **kwargs: Any)

Bases: BaseEstimator

Transformer that drops the samples with too many missing views.

Parameters:

  • max_missing_views (int, default=0) – Maximum number of missing views a sample may have and still be kept.

  • how ({"all", "any"}, default="all") – Passed to missing_view_mask().

support_mask_

True for the samples kept by the last call to transform().

Type:

ndarray of bool of shape (n_samples,)

n_views_in_

Number of views seen during fit().

Type:

int

n_features_in_

Number of features of each view.

Type:

list of int

n_dropped_

Number of samples the last transform() removed.

Type:

int

Notes

This is the one estimator of the module that changes n_samples. It therefore does not belong in the middle of a supervised PolyPipeline, where y would no longer line up with the rows; apply it to the data (and to y) beforehand, or use support_mask_() to subset y yourself. A MultiViewDataset in gives a dataset out, with its labels subset for you.

Examples

>>> import numpy as np
>>> from polyview.imputation import DropIncompleteSamples
>>> X1 = np.array([[1.0], [np.nan], [3.0]])
>>> X2 = np.array([[4.0], [5.0], [6.0]])
>>> dropper = DropIncompleteSamples()
>>> kept = dropper.fit_transform([X1, X2])
>>> kept[0].ravel()
array([1., 3.])
>>> dropper.n_dropped_
1
fit(views, y=None) → DropIncompleteSamples

Record the shape of the views. Nothing is learned from the data.

Parameters:
Return type:

self

fit_transform(views, y=None)

Fit and immediately drop.

Parameters:
Return type:

list of ndarray, or MultiViewDataset

transform(views)

Return views without the samples that miss too many views.

Parameters:

views (list of array-like or MultiViewDataset)

Return type:

list of ndarray, or MultiViewDataset

polyview.imputation.drop.drop_missing_views(views, max_missing_views: int = 0, how: str = 'all', return_mask: bool = False) → List[numpy.ndarray] | Tuple[List[numpy.ndarray], numpy.ndarray]

Keep only the samples that have enough observed views.

Complete-case analysis is the honest baseline for missing views: it introduces no synthetic data, at the cost of sample size and of a bias whenever the views are not missing completely at random.

Parameters:

  • views (list of array-like or MultiViewDataset)

  • max_missing_views (int, default=0) – Maximum number of missing views a sample may have and still be kept. 0 keeps only fully observed samples.

  • how ({"all", "any"}, default="all") – Passed to missing_view_mask(): whether a view counts as missing only when the whole row is NaN, or as soon as one entry is.

  • return_mask (bool, default=False) – Also return the boolean mask of the kept samples.

Returns:

  • views_out (list of ndarray) – The views restricted to the kept samples.

  • support (ndarray of bool of shape (n_samples,)) – Only when return_mask=True. True for kept samples.

Examples

>>> import numpy as np
>>> from polyview.imputation import drop_missing_views
>>> X1 = np.array([[1.0], [np.nan], [3.0]])
>>> X2 = np.array([[4.0], [5.0], [6.0]])
>>> kept, support = drop_missing_views([X1, X2], return_mask=True)
>>> support
array([ True, False,  True])
>>> kept[0].ravel()
array([1., 3.])