scikit-learn/sklearn/feature_selection/univariate_selection.py

570 lines
18 KiB
Python

# -*- coding: utf-8 -*-
"""Univariate features selection."""
# Authors: V. Michel, B. Thirion, G. Varoquaux, A. Gramfort, E. Duchesnay.
# L. Buitinck
# License: BSD 3 clause
from abc import ABCMeta, abstractmethod
from warnings import warn
import numpy as np
from scipy import stats
from scipy.sparse import issparse
from ..base import BaseEstimator, TransformerMixin
from ..preprocessing import LabelBinarizer
from ..utils import (array2d, as_float_array, atleast2d_or_csr, check_arrays,
safe_asarray, safe_sqr, safe_mask)
from ..utils.extmath import safe_sparse_dot
def _clean_nans(scores):
"""
Fixes Issue #1240: NaNs can't be properly compared, so change them to the
smallest value of scores's dtype. -inf seems to be unreliable.
"""
# XXX where should this function be called? fit? scoring functions
# themselves?
scores = as_float_array(scores, copy=True)
scores[np.isnan(scores)] = np.finfo(scores.dtype).min
return scores
######################################################################
# Scoring functions
# The following function is a rewriting of scipy.stats.f_oneway
# Contrary to the scipy.stats.f_oneway implementation it does not
# copy the data while keeping the inputs unchanged.
def f_oneway(*args):
"""Performs a 1-way ANOVA.
The one-way ANOVA tests the null hypothesis that 2 or more groups have
the same population mean. The test is applied to samples from two or
more groups, possibly with differing sizes.
Parameters
----------
sample1, sample2, ... : array_like, sparse matrices
The sample measurements should be given as arguments.
Returns
-------
F-value : float
The computed F-value of the test
p-value : float
The associated p-value from the F-distribution
Notes
-----
The ANOVA test has important assumptions that must be satisfied in order
for the associated p-value to be valid.
1. The samples are independent
2. Each sample is from a normally distributed population
3. The population standard deviations of the groups are all equal. This
property is known as homoscedasticity.
If these assumptions are not true for a given set of data, it may still be
possible to use the Kruskal-Wallis H-test (`scipy.stats.kruskal`_) although
with some loss of power.
The algorithm is from Heiman[2], pp.394-7.
See ``scipy.stats.f_oneway`` that should give the same results while
being less efficient.
References
----------
.. [1] Lowry, Richard. "Concepts and Applications of Inferential
Statistics". Chapter 14.
http://faculty.vassar.edu/lowry/ch14pt1.html
.. [2] Heiman, G.W. Research Methods in Statistics. 2002.
"""
n_classes = len(args)
args = [safe_asarray(a) for a in args]
n_samples_per_class = np.array([a.shape[0] for a in args])
n_samples = np.sum(n_samples_per_class)
ss_alldata = reduce(lambda x, y: x + y,
[safe_sqr(a).sum(axis=0) for a in args])
sums_args = [a.sum(axis=0) for a in args]
square_of_sums_alldata = safe_sqr(reduce(lambda x, y: x + y, sums_args))
square_of_sums_args = [safe_sqr(s) for s in sums_args]
sstot = ss_alldata - square_of_sums_alldata / float(n_samples)
ssbn = 0.
for k, _ in enumerate(args):
ssbn += square_of_sums_args[k] / n_samples_per_class[k]
ssbn -= square_of_sums_alldata / float(n_samples)
sswn = sstot - ssbn
dfbn = n_classes - 1
dfwn = n_samples - n_classes
msb = ssbn / float(dfbn)
msw = sswn / float(dfwn)
f = msb / msw
# flatten matrix to vector in sparse case
f = np.asarray(f).ravel()
prob = stats.fprob(dfbn, dfwn, f)
return f, prob
def f_classif(X, y):
"""Compute the Anova F-value for the provided sample
Parameters
----------
X : {array-like, sparse matrix} shape = [n_samples, n_features]
The set of regressors that will tested sequentially
y : array of shape(n_samples)
The data matrix
Returns
-------
F : array, shape = [n_features,]
The set of F values
pval : array, shape = [n_features,]
The set of p-values
"""
X, y = check_arrays(X, y)
args = [X[safe_mask(X, y == k)] for k in np.unique(y)]
return f_oneway(*args)
def chi2(X, y):
"""Compute χ² (chi-squared) statistic for each class/feature combination.
This score can be used to select the n_features features with the
highest values for the χ² (chi-square) statistic from either boolean or
multinomially distributed data (e.g., term counts in document
classification) relative to the classes.
Recall that the χ² statistic measures dependence between stochastic
variables, so using this function "weeds out" the features that are the
most likely to be independent of class and therefore irrelevant for
classification.
Parameters
----------
X : {array-like, sparse matrix}, shape = (n_samples, n_features_in)
Sample vectors.
y : array-like, shape = (n_samples,)
Target vector (class labels).
Returns
-------
chi2 : array, shape = (n_features,)
chi2 statistics of each feature.
pval : array, shape = (n_features,)
p-values of each feature.
Notes
-----
Complexity of this algorithm is O(n_classes * n_features).
"""
# XXX: we might want to do some of the following in logspace instead for
# numerical stability.
X = atleast2d_or_csr(X)
Y = LabelBinarizer().fit_transform(y)
if Y.shape[1] == 1:
Y = np.append(1 - Y, Y, axis=1)
observed = safe_sparse_dot(Y.T, X) # n_classes * n_features
feature_count = array2d(X.sum(axis=0))
class_prob = array2d(Y.mean(axis=0))
expected = np.dot(class_prob.T, feature_count)
return stats.chisquare(observed, expected)
def f_regression(X, y, center=True):
"""Univariate linear regression tests
Quick linear model for testing the effect of a single regressor,
sequentially for many regressors.
This is done in 3 steps:
1. the regressor of interest and the data are orthogonalized
wrt constant regressors
2. the cross correlation between data and regressors is computed
3. it is converted to an F score then to a p-value
Parameters
----------
X : {array-like, sparse matrix} shape = (n_samples, n_features)
The set of regressors that will tested sequentially.
y : array of shape(n_samples).
The data matrix
center : True, bool,
If true, X and y will be centered.
Returns
-------
F : array, shape=(n_features,)
F values of features.
pval : array, shape=(n_features,)
p-values of F-scores.
"""
if issparse(X) and center:
raise ValueError("center=True only allowed for dense data")
X, y = check_arrays(X, y, dtype=np.float)
y = y.ravel()
if center:
y = y - np.mean(y)
X = X.copy('F') # faster in fortran
X -= X.mean(axis=0)
# compute the correlation
corr = safe_sparse_dot(y, X)
corr /= np.asarray(np.sqrt(safe_sqr(X).sum(axis=0))).ravel()
corr /= np.asarray(np.sqrt(safe_sqr(y).sum())).ravel()
# convert to p-value
dof = y.size - 2
F = corr ** 2 / (1 - corr ** 2) * dof
pv = stats.f.sf(F, 1, dof)
return F, pv
######################################################################
# General class for filter univariate selection
class _AbstractUnivariateFilter(BaseEstimator, TransformerMixin):
__metaclass__ = ABCMeta
def __init__(self, score_func):
""" Initialize the univariate feature selection.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
"""
if not callable(score_func):
raise TypeError(
"The score function should be a callable, %r "
"was passed." % (score_func, type(score_func)))
self.score_func = score_func
def fit(self, X, y):
"""
Evaluate the function
"""
self.scores_, self.pvalues_ = self.score_func(X, y)
if len(np.unique(self.pvalues_)) < len(self.pvalues_):
warn("Duplicate p-values. Result may depend on feature ordering."
"There are probably duplicate features, or you used a "
"classification score for a regression task.")
return self
def get_support(self, indices=False):
"""
Return a mask, or list, of the features/indices selected.
"""
mask = self._get_support_mask()
return mask if not indices else np.where(mask)[0]
@abstractmethod
def _get_support_mask(self):
"""
Must return a boolean mask indicating which features are selected.
"""
def transform(self, X):
"""
Transform a new matrix using the selected features
"""
X = atleast2d_or_csr(X)
mask = self._get_support_mask()
if len(mask) != X.shape[1]:
raise ValueError("X has a different shape than during fitting.")
return atleast2d_or_csr(X)[:, safe_mask(X, mask)]
def inverse_transform(self, X):
"""
Transform a new matrix using the selected features
"""
support_ = self.get_support()
if X.ndim == 1:
X = X[None, :]
Xt = np.zeros((X.shape[0], support_.size))
Xt[:, support_] = X
return Xt
######################################################################
# Specific filters
######################################################################
class SelectPercentile(_AbstractUnivariateFilter):
"""Select features according to a percentile of the highest scores.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
percentile : int, optional, default=10
Percent of features to keep.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
Notes
-----
Ties between features with equal p-values will be broken in an unspecified
way.
"""
def __init__(self, score_func=f_classif, percentile=10):
if not 0 <= percentile <= 100:
raise ValueError("percentile should be >=0, <=100; got %r"
% percentile)
self.percentile = percentile
super(SelectPercentile, self).__init__(score_func)
def _get_support_mask(self):
percentile = self.percentile
if percentile > 100:
raise ValueError("percentile should be between 0 and 100"
" (%f given)" % (percentile))
# Cater for NaNs
if percentile == 100:
return np.ones(len(self.scores_), dtype=np.bool)
elif percentile == 0:
return np.zeros(len(self.scores_), dtype=np.bool)
scores = _clean_nans(self.scores_)
alpha = stats.scoreatpercentile(scores, 100 - percentile)
# XXX refactor the indices -> mask -> indices -> mask thing
inds = np.where(scores >= alpha)[0]
# if we selected too many features because of equal scores,
# we throw them away now
inds = inds[:len(scores) * percentile // 100]
mask = np.zeros(scores.shape, dtype=np.bool)
mask[inds] = True
return mask
class SelectKBest(_AbstractUnivariateFilter):
"""Select features according to the k highest scores.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
k : int, optional, default=10
Number of top features to select.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
Notes
-----
Ties between features with equal scores will be broken in an unspecified
way.
"""
def __init__(self, score_func=f_classif, k=10):
self.k = k
super(SelectKBest, self).__init__(score_func)
def _get_support_mask(self):
k = self.k
if k > len(self.scores_):
raise ValueError("cannot select %d features among %d"
% (k, len(self.scores_)))
scores = _clean_nans(self.scores_)
# XXX This should be refactored; we're getting an array of indices
# from argsort, which we transform to a mask, which we probably
# transform back to indices later.
mask = np.zeros(scores.shape, dtype=bool)
mask[np.argsort(scores)[-k:]] = 1
return mask
class SelectFpr(_AbstractUnivariateFilter):
"""Filter: Select the pvalues below alpha based on a FPR test.
FPR test stands for False Positive Rate test. It controls the total
amount of false detections.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
alpha : float, optional
The highest p-value for features to be kept.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
"""
def __init__(self, score_func=f_classif, alpha=5e-2):
self.alpha = alpha
super(SelectFpr, self).__init__(score_func)
def _get_support_mask(self):
alpha = self.alpha
return self.pvalues_ < alpha
class SelectFdr(_AbstractUnivariateFilter):
"""Filter: Select the p-values for an estimated false discovery rate
This uses the Benjamini-Hochberg procedure. ``alpha`` is the target false
discovery rate.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
alpha : float, optional
The highest uncorrected p-value for features to keep.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
"""
def __init__(self, score_func=f_classif, alpha=5e-2):
self.alpha = alpha
super(SelectFdr, self).__init__(score_func)
def _get_support_mask(self):
alpha = self.alpha
sv = np.sort(self.pvalues_)
threshold = sv[sv < alpha * np.arange(len(self.pvalues_))].max()
return self.pvalues_ <= threshold
class SelectFwe(_AbstractUnivariateFilter):
"""Filter: Select the p-values corresponding to Family-wise error rate
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
alpha : float, optional
The highest uncorrected p-value for features to keep.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
"""
def __init__(self, score_func=f_classif, alpha=5e-2):
self.alpha = alpha
super(SelectFwe, self).__init__(score_func)
def _get_support_mask(self):
alpha = self.alpha
return (self.pvalues_ < alpha / len(self.pvalues_))
######################################################################
# Generic filter
######################################################################
class GenericUnivariateSelect(_AbstractUnivariateFilter):
"""Univariate feature selector with configurable strategy.
Parameters
----------
score_func : callable
Function taking two arrays X and y, and returning a pair of arrays
(scores, pvalues).
mode : {'percentile', 'k_best', 'fpr', 'fdr', 'fwe'}
Feature selection mode.
param : float or int depending on the feature selection mode
Parameter of the corresponding mode.
Attributes
----------
scores_ : array-like, shape=(n_features,)
Scores of features.
pvalues_ : array-like, shape=(n_features,)
p-values of feature scores.
"""
_selection_modes = {'percentile': SelectPercentile,
'k_best': SelectKBest,
'fpr': SelectFpr,
'fdr': SelectFdr,
'fwe': SelectFwe,
}
def __init__(self, score_func=f_classif, mode='percentile', param=1e-5):
if not callable(score_func):
raise TypeError(
"The score function should be a callable, %r (type %s) "
"was passed." % (score_func, type(score_func)))
if mode not in self._selection_modes:
raise ValueError(
"The mode passed should be one of %s, %r, (type %s) "
"was passed." % (
self._selection_modes.keys(),
mode, type(mode)))
super(GenericUnivariateSelect, self).__init__(score_func)
self.mode = mode
self.param = param
def _get_support_mask(self):
selector = self._selection_modes[self.mode](lambda x: x)
selector.pvalues_ = self.pvalues_
selector.scores_ = self.scores_
# Now perform some acrobatics to set the right named parameter in
# the selector
possible_params = selector._get_param_names()
possible_params.remove('score_func')
selector.set_params(**{possible_params[0]: self.param})
return selector._get_support_mask()