Simple imputation

Marginal imputation replaces every missing entry by a statistic of its own column — the mean, the median, or a fixed constant — computed over the samples for which the view is observed. It ignores the correlations between views, which makes it the natural baseline: a strategy that cannot beat it is not exploiting the multi-view structure at all.

Marginal, per-view imputation of missing views.

class polyview.imputation.simple.SimpleViewImputer(*args: Any, **kwargs: Any)

Bases: BaseViewImputer

Fill missing views with a per-view, per-feature constant.

Each column of each view gets one statistic, computed over the samples for which that view is observed; every NaN in the column is then replaced by it. This is the multi-view counterpart of sklearn.impute.SimpleImputer and the natural baseline any other strategy should be compared against: it ignores the correlations between views, so a missing view contributes nothing beyond its own mean.

Parameters:

  • strategy ({"mean", "median", "constant"}, default="mean") – Statistic used as fill value. "constant" uses fill_value.

  • fill_value (float, default=0.0) – Value used when strategy="constant".

  • n_views (int or None, default=None) – Expected number of views. None accepts any count.

statistics_

Fill value of each column of each view, shape (n_features_i,).

Type:

list of ndarray

missing_view_mask_

True where a whole view was missing at fit time.

Type:

ndarray of bool of shape (n_samples, n_views)

missing_rate_

Fraction of fit samples for which each view was missing.

Type:

ndarray of shape (n_views,)

Notes

A column that is NaN for every fit sample has no statistic to learn; it is filled with 0.0 rather than raising, so that a view which is missing for the whole training set does not break a pipeline.

Examples

>>> import numpy as np
>>> from polyview.imputation import SimpleViewImputer
>>> X1 = np.array([[1.0, 2.0], [3.0, 4.0], [np.nan, np.nan]])
>>> X2 = np.array([[5.0], [6.0], [7.0]])
>>> filled = SimpleViewImputer().fit_transform([X1, X2])
>>> filled[0]
array([[1., 2.],
       [3., 4.],
       [2., 3.]])