scikit-learn/scikits/learn/pca.py

432 lines
14 KiB
Python

""" Principal Component Analysis
"""
# Author: Alexandre Gramfort <alexandre.gramfort@inria.fr>
# Olivier Grisel <olivier.grisel@ensta.org>
# License: BSD Style.
import numpy as np
from scipy import linalg
from .base import BaseEstimator
from .utils.extmath import fast_logdet
from .utils.extmath import fast_svd
from .utils.extmath import safe_sparse_dot
def _assess_dimension_(spectrum, rank, n_samples, dim):
"""Compute the likelihood of a rank rank dataset
The dataset is assumed to be embedded in gaussian noise of shape(n,
dimf) having spectrum spectrum.
Parameters
----------
spectrum: array of shape (n)
data spectrum
rank: int,
tested rank value
n_samples: int,
number of samples
dim: int,
embedding/empirical dimension
Returns
-------
ll: float,
The log-likelihood
Notes
-----
This implements the method of Thomas P. Minka:
Automatic Choice of Dimensionality for PCA. NIPS 2000: 598-604
"""
if rank > dim:
raise ValueError("the dimension cannot exceed dim")
from scipy.special import gammaln
pu = -rank * np.log(2)
for i in range(rank):
pu += gammaln((dim - i) / 2) - np.log(np.pi) * (dim - i) / 2
pl = np.sum(np.log(spectrum[:rank]))
pl = -pl * n_samples / 2
if rank == dim:
pv = 0
v = 1
else:
v = np.sum(spectrum[rank:dim]) / (dim - rank)
pv = -np.log(v) * n_samples * (dim - rank) / 2
m = dim * rank - rank * (rank + 1) / 2
pp = np.log(2 * np.pi) * (m + rank + 1) / 2
pa = 0
spectrum_ = spectrum.copy()
spectrum_[rank:dim] = v
for i in range(rank):
for j in range(i + 1, dim):
pa += (np.log((spectrum[i] - spectrum[j])
* (1. / spectrum_[j] - 1. / spectrum_[i]))
+ np.log(n_samples))
ll = pu + pl + pv + pp - pa / 2 - rank * np.log(n_samples) / 2
return ll
def _infer_dimension_(spectrum, n, p):
"""This method infers the dimension of a dataset of shape (n, p)
The dataset is described by its spectrum `spectrum`.
"""
ll = []
for rank in range(min(n, p, len(spectrum))):
ll.append(_assess_dimension_(spectrum, rank, n, p))
ll = np.array(ll)
return ll.argmax()
class PCA(BaseEstimator):
"""Principal component analysis (PCA)
Linear dimensionality reduction using Singular Value Decomposition of the
data and keeping only the most significant singular vectors to project the
data to a lower dimensional space.
This implementation uses the scipy.linalg implementation of the singular
value decomposition. It only works for dense arrays and is not scalable to
large dimensional data.
The time complexity of this implementation is O(n ** 3) assuming
n ~ n_samples ~ n_features.
Parameters
----------
n_components: int, none or string
Number of components to keep.
if n_components is not set all components are kept:
n_components == min(n_samples, n_features)
if n_components == 'mle', Minka's MLE is used to guess the dimension
if 0 < n_components < 1, select the number of components such that
the explained variance ratio is greater
than n_components
copy: bool
If False, data passed to fit are overwritten
whiten: bool, optional
When True (False by default) the components_ vectors are divided
by n_samples times singular values to ensure uncorrelated outputs
with unit component-wise variances.
Whitening will remove some information from the transformed signal
(the relative variance scales of the components) but can sometime
improve the predictive accuracy of the downstream estimators by
making there data respect some hard-wired assumptions.
Attributes
----------
components_: array, [n_features, n_components]
Components with maximum variance.
explained_variance_ratio_: array, [n_components]
Percentage of variance explained by each of the selected components.
k is not set then all components are stored and the sum of
explained variances is equal to 1.0
Notes
-----
For n_components='mle', this class uses the method of Thomas P. Minka:
Automatic Choice of Dimensionality for PCA. NIPS 2000: 598-604
Examples
--------
>>> import numpy as np
>>> from scikits.learn.pca import PCA
>>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]])
>>> pca = PCA(n_components=2)
>>> pca.fit(X)
PCA(copy=True, n_components=2, whiten=False)
>>> print pca.explained_variance_ratio_
[ 0.99244289 0.00755711]
See also
--------
ProbabilisticPCA
RandomizedPCA
"""
def __init__(self, n_components=None, copy=True, whiten=False):
self.n_components = n_components
self.copy = copy
self.whiten = whiten
def fit(self, X, y=None, **params):
"""Fit the model from data in X.
Parameters
----------
X: array-like, shape (n_samples, n_features)
Training vector, where n_samples in the number of samples
and n_features is the number of features.
Returns
-------
self : object
Returns the instance itself.
"""
self._set_params(**params)
X = np.atleast_2d(X)
n_samples, n_features = X.shape
if self.copy:
X = X.copy()
# Center data
self.mean_ = np.mean(X, axis=0)
X -= self.mean_
U, S, V = linalg.svd(X, full_matrices=False)
self.explained_variance_ = (S ** 2) / n_samples
self.explained_variance_ratio_ = self.explained_variance_ / \
self.explained_variance_.sum()
if self.whiten:
n = X.shape[0]
self.components_ = np.dot(V.T, np.diag(1.0 / S)) * np.sqrt(n)
else:
self.components_ = V.T
if self.n_components == 'mle':
self.n_components = _infer_dimension_(self.explained_variance_,
n_samples, X.shape[1])
elif 0 < self.n_components and self.n_components < 1.0:
# number of components for which the cumulated explained variance
# percentage is superior to the desired threshold
n_remove = np.sum(self.explained_variance_ratio_.cumsum() >=
self.n_components) - 1
self.n_components = n_features - n_remove
if self.n_components is not None:
self.components_ = self.components_[:, :self.n_components]
self.explained_variance_ = \
self.explained_variance_[:self.n_components]
self.explained_variance_ratio_ = \
self.explained_variance_ratio_[:self.n_components]
return self
def transform(self, X):
"""Apply the dimension reduction learned on the train data."""
X_transformed = X - self.mean_
X_transformed = np.dot(X_transformed, self.components_)
return X_transformed
def inverse_transform(self, X):
"""Return an input X_original whose transform would be X
Note: if whitening is enabled, inverse_transform does not compute the
exact inverse operation as transform.
"""
return np.dot(X, self.components_.T) + self.mean_
class ProbabilisticPCA(PCA):
"""Additional layer on top of PCA that add a probabilistic evaluation
""" + PCA.__doc__
def fit(self, X, y=None, homoscedastic=True):
"""Additionally to PCA.fit, learns a covariance model
Parameters
----------
X: array of shape(n_samples, n_dim)
The data to fit
homoscedastic: bool, optional,
If True, average variance across remaining dimensions
"""
PCA.fit(self, X)
self.dim = X.shape[1]
Xr = X - self.mean_
Xr -= np.dot(np.dot(Xr, self.components_), self.components_.T)
n_samples = X.shape[0]
if self.dim <= self.n_components:
delta = np.zeros(self.dim)
elif homoscedastic:
delta = (Xr ** 2).sum() * np.ones(self.dim) \
/ (n_samples * self.dim)
else:
delta = (Xr ** 2).mean(0) / (self.dim - self.n_components)
self.covariance_ = np.diag(delta)
for k in range(self.n_components):
add_cov = np.dot(
self.components_[:, k:k + 1], self.components_[:, k:k + 1].T)
self.covariance_ += self.explained_variance_[k] * add_cov
return self
def score(self, X):
"""Return a score associated to new data
Parameters
----------
X: array of shape(n_samples, n_dim)
The data to test
Returns
-------
ll: array of shape (n_samples),
log-likelihood of each row of X under the current model
"""
Xr = X - self.mean_
log_like = np.zeros(X.shape[0])
self.precision_ = np.linalg.inv(self.covariance_)
for i in range(X.shape[0]):
log_like[i] = -.5 * np.dot(np.dot(self.precision_, Xr[i]), Xr[i])
log_like += fast_logdet(self.precision_) - \
self.dim / 2 * np.log(2 * np.pi)
return log_like
class RandomizedPCA(BaseEstimator):
"""Principal component analysis (PCA) using randomized SVD
Linear dimensionality reduction using approximated Singular Value
Decomposition of the data and keeping only the most significant
singular vectors to project the data to a lower dimensional space.
This implementation uses a randomized SVD implementation and can
handle both scipy.sparse and numpy dense arrays as input.
Parameters
----------
n_components: int
Maximum number of components to keep: default is 50.
copy: bool
If False, data passed to fit are overwritten
iterated_power: int, optional
Number of iteration for the power method. 3 by default.
whiten: bool, optional
When True (False by default) the components_ vectors are divided
by the singular values to ensure uncorrelated outputs with unit
component-wise variances.
Whitening will remove some information from the transformed signal
(the relative variance scales of the components) but can sometime
improve the predictive accuracy of the downstream estimators by
making there data respect some hard-wired assumptions.
Attributes
----------
components_: array, [n_features, n_components]
Components with maximum variance.
explained_variance_ratio_: array, [n_components]
Percentage of variance explained by each of the selected components.
k is not set then all components are stored and the sum of
explained variances is equal to 1.0
Examples
--------
>>> import numpy as np
>>> from scikits.learn.pca import RandomizedPCA
>>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]])
>>> pca = RandomizedPCA(n_components=2)
>>> pca.fit(X)
RandomizedPCA(copy=True, n_components=2, iterated_power=3, whiten=False)
>>> print pca.explained_variance_ratio_
[ 0.99244289 0.00755711]
See also
--------
PCA
ProbabilisticPCA
Notes
-------
References:
* Finding structure with randomness: Stochastic algorithms for
constructing approximate matrix decompositions Halko, et al., 2009
(arXiv:909)
* A randomized algorithm for the decomposition of matrices
Per-Gunnar Martinsson, Vladimir Rokhlin and Mark Tygert
"""
def __init__(self, n_components, copy=True, iterated_power=3,
whiten=False):
self.n_components = n_components
self.copy = copy
self.iterated_power = iterated_power
self.whiten = whiten
self.mean_ = None
def fit(self, X, y=None, **params):
"""Fit the model to the data X.
Parameters
----------
X: array-like or scipy.sparse matrix, shape (n_samples, n_features)
Training vector, where n_samples in the number of samples and
n_features is the number of features.
Returns
-------
self : object
Returns the instance itself.
"""
self._set_params(**params)
if not hasattr(X, 'todense'):
X = np.atleast_2d(X)
n_samples = X.shape[0]
if self.copy:
X = X.copy()
if not hasattr(X, 'todense'):
# not a sparse matrix, ensure this is a 2D array
X = np.atleast_2d(X)
# Center data
self.mean_ = np.mean(X, axis=0)
X -= self.mean_
U, S, V = fast_svd(X, self.n_components, q=self.iterated_power)
self.explained_variance_ = (S ** 2) / n_samples
self.explained_variance_ratio_ = self.explained_variance_ / \
self.explained_variance_.sum()
if self.whiten:
n = X.shape[0]
self.components_ = np.dot(V.T, np.diag(1.0 / S)) * np.sqrt(n)
else:
self.components_ = V.T
return self
def transform(self, X):
"""Apply the dimension reduction learned on the training data."""
if self.mean_ is not None:
X = X - self.mean_
X = safe_sparse_dot(X, self.components_)
return X
def inverse_transform(self, X):
"""Return an reconstructed input whose transform would be X"""
X_original = safe_sparse_dot(X, self.components_.T)
if self.mean_ is not None:
X_original = X_original + self.mean_
return X_original