mojolearn.KBinsDiscretizer#

class mojolearn.KBinsDiscretizer(n_bins=5, *, encode='onehot', strategy='quantile', quantile_method='averaged_inverted_cdf', dtype=None, subsample=200000, random_state=None)[source]#

Bases: _PrepBase

Bin continuous features into intervals.

Follows scikit-learn’s KBinsDiscretizer. Each column is split into n_bins intervals and each value is replaced by its bin index or a one-hot block. One-hot output is always a dense float32 array. A constant column gets a single bin with edges (-inf, inf).

Parameters:

  • n_bins (int or array-like of int, default 5) – Number of bins, one value for all columns or one per column; each must be >= 2.

  • encode ({‘onehot’, ‘onehot-dense’, ‘ordinal’}, default ‘onehot’) – Output encoding. 'onehot' returns a dense array, the same as 'onehot-dense'.

  • strategy ({‘uniform’, ‘quantile’, ‘kmeans’}, default ‘quantile’) – 'uniform': equal-width bins. 'quantile': edges at the quantiles given by quantile_method. 'kmeans': 1-D k-means started from the uniform bin centers.

  • quantile_method (str, default ‘averaged_inverted_cdf’) – numpy quantile method for strategy='quantile': 'averaged_inverted_cdf', 'inverted_cdf', 'closest_observation', 'interpolated_inverted_cdf', 'hazen', 'weibull', 'linear', 'median_unbiased' or 'normal_unbiased'.

  • dtype (default None) – Accepted for scikit-learn compatibility; the output is always float32.

  • subsample (int or None, default 200000) – Above this many rows, fit uses a resample of subsample rows drawn with replacement. None uses every row.

  • random_state (int or None, default None) – Seed for the resample. None uses seed 0. The draw uses mojolearn’s own generator, so the rows differ from scikit-learn’s for the same seed.

Variables:

  • bin_edges_ (list of arrays) – Edges of each column’s bins.

  • n_bins_ (array of int64) – Number of bins of each column (fewer than requested for a constant column).

  • n_features_in_ (int) – Number of columns seen in fit.

  • numeric_mode_ (str) – Numeric mode the estimator was fitted in; transform runs in the same mode.

Notes

fit accepts sample_weight (nonnegative, not all zero). It weights the quantile edges (only 'averaged_inverted_cdf' and 'inverted_cdf' are allowed with weights, as in scikit-learn; other methods raise ValueError), the uniform range (minimum and maximum over rows of nonzero weight) and the k-means fit. Above subsample rows the weights instead drive a weighted resample.

inverse_transform returns each bin’s center. An all-zero one-hot block or a code that names no bin raises ValueError.

Examples

>>> import numpy as np, mojolearn
>>> X = np.random.default_rng(0).random((1000, 4), dtype=np.float32)
>>> Xt = mojolearn.KBinsDiscretizer(n_bins=4, encode="ordinal").fit_transform(X)
__init__(n_bins=5, *, encode='onehot', strategy='quantile', quantile_method='averaged_inverted_cdf', dtype=None, subsample=200000, random_state=None)[source]#

Methods

fit(X[, y, sample_weight])

fit_transform(X[, y])

Fit to X, then transform X.

get_params([deep])

Return the constructor parameters as a dict.

inverse_transform(X)

set_params(**params)

Set constructor parameters.

transform(X)