non_negative_factorization#
- sklearn.decomposition.non_negative_factorization(X, W=None, H=None, n_components='auto', *, init=None, update_H=True, solver='cd', beta_loss='frobenius', tol=0.0001, max_iter=200, alpha_W=0.0, alpha_H='same', l1_ratio=0.0, random_state=None, verbose=0, shuffle=False)[source]#
Compute Non-negative Matrix Factorization (NMF).
Find two non-negative matrices (W, H) whose product approximates the non- negative matrix X. This factorization can be used for example for dimensionality reduction, source separation or topic extraction.
The objective function is:
\[ \begin{align}\begin{aligned}L(W, H) &= 0.5 * ||X - WH||_{loss}^2\\ &+ alpha\_W * l1\_ratio * n\_features * ||vec(W)||_1\\ &+ alpha\_H * l1\_ratio * n\_samples * ||vec(H)||_1\\ &+ 0.5 * alpha\_W * (1 - l1\_ratio) * n\_features * ||W||_{Fro}^2\\ &+ 0.5 * alpha\_H * (1 - l1\_ratio) * n\_samples * ||H||_{Fro}^2,\end{aligned}\end{align} \]where \(||A||_{Fro}^2 = \sum_{i,j} A_{ij}^2\) (Frobenius norm) and \(||vec(A)||_1 = \sum_{i,j} abs(A_{ij})\) (Elementwise L1 norm)
The generic norm \(||X - WH||_{loss}^2\) may represent the Frobenius norm or another supported beta-divergence loss. The choice between options is controlled by the
beta_loss
parameter.The regularization terms are scaled by
n_features
forW
and byn_samples
forH
to keep their impact balanced with respect to one another and to the data fit term as independent as possible of the sizen_samples
of the training set.The objective function is minimized with an alternating minimization of W and H. If H is given and update_H=False, it solves for W only.
Note that the transformed data is named W and the components matrix is named H. In the NMF literature, the naming convention is usually the opposite since the data matrix X is transposed.
- Parameters:
- X{array-like, sparse matrix} of shape (n_samples, n_features)
Constant matrix.
- Warray-like of shape (n_samples, n_components), default=None
If
init='custom'
, it is used as initial guess for the solution. Ifupdate_H=False
, it is initialised as an array of zeros, unlesssolver='mu'
, then it is filled with values calculated bynp.sqrt(X.mean() / self._n_components)
. IfNone
, uses the initialisation method specified ininit
.- Harray-like of shape (n_components, n_features), default=None
If
init='custom'
, it is used as initial guess for the solution. Ifupdate_H=False
, it is used as a constant, to solve for W only. IfNone
, uses the initialisation method specified ininit
.- n_componentsint or {‘auto’} or None, default=’auto’
Number of components. If
None
, all features are kept. Ifn_components='auto'
, the number of components is automatically inferred fromW
orH
shapes.Changed in version 1.4: Added
'auto'
value.Changed in version 1.6: Default value changed from
None
to'auto'
.- init{‘random’, ‘nndsvd’, ‘nndsvda’, ‘nndsvdar’, ‘custom’}, default=None
Method used to initialize the procedure.
Valid options:
None: ‘nndsvda’ if n_components < n_features, otherwise ‘random’.
‘random’: non-negative random matrices, scaled with:
sqrt(X.mean() / n_components)
‘nndsvd’: Nonnegative Double Singular Value Decomposition (NNDSVD) initialization (better for sparseness)
‘nndsvda’: NNDSVD with zeros filled with the average of X (better when sparsity is not desired)
‘nndsvdar’: NNDSVD with zeros filled with small random values (generally faster, less accurate alternative to NNDSVDa for when sparsity is not desired)
‘custom’: If
update_H=True
, use custom matrices W and H which must both be provided. Ifupdate_H=False
, then only custom matrix H is used.
Changed in version 0.23: The default value of
init
changed from ‘random’ to None in 0.23.Changed in version 1.1: When
init=None
and n_components is less than n_samples and n_features defaults tonndsvda
instead ofnndsvd
.- update_Hbool, default=True
Set to True, both W and H will be estimated from initial guesses. Set to False, only W will be estimated.
- solver{‘cd’, ‘mu’}, default=’cd’
Numerical solver to use:
‘cd’ is a Coordinate Descent solver that uses Fast Hierarchical Alternating Least Squares (Fast HALS).
‘mu’ is a Multiplicative Update solver.
Added in version 0.17: Coordinate Descent solver.
Added in version 0.19: Multiplicative Update solver.
- beta_lossfloat or {‘frobenius’, ‘kullback-leibler’, ‘itakura-saito’}, default=’frobenius’
Beta divergence to be minimized, measuring the distance between X and the dot product WH. Note that values different from ‘frobenius’ (or 2) and ‘kullback-leibler’ (or 1) lead to significantly slower fits. Note that for beta_loss <= 0 (or ‘itakura-saito’), the input matrix X cannot contain zeros. Used only in ‘mu’ solver.
Added in version 0.19.
- tolfloat, default=1e-4
Tolerance of the stopping condition.
- max_iterint, default=200
Maximum number of iterations before timing out.
- alpha_Wfloat, default=0.0
Constant that multiplies the regularization terms of
W
. Set it to zero (default) to have no regularization onW
.Added in version 1.0.
- alpha_Hfloat or “same”, default=”same”
Constant that multiplies the regularization terms of
H
. Set it to zero to have no regularization onH
. If “same” (default), it takes the same value asalpha_W
.Added in version 1.0.
- l1_ratiofloat, default=0.0
The regularization mixing parameter, with 0 <= l1_ratio <= 1. For l1_ratio = 0 the penalty is an elementwise L2 penalty (aka Frobenius Norm). For l1_ratio = 1 it is an elementwise L1 penalty. For 0 < l1_ratio < 1, the penalty is a combination of L1 and L2.
- random_stateint, RandomState instance or None, default=None
Used for NMF initialisation (when
init
== ‘nndsvdar’ or ‘random’), and in Coordinate Descent. Pass an int for reproducible results across multiple function calls. See Glossary.- verboseint, default=0
The verbosity level.
- shufflebool, default=False
If true, randomize the order of coordinates in the CD solver.
- Returns:
- Wndarray of shape (n_samples, n_components)
Solution to the non-negative least squares problem.
- Hndarray of shape (n_components, n_features)
Solution to the non-negative least squares problem.
- n_iterint
Actual number of iterations.
References
[1]“Fast local algorithms for large scale nonnegative matrix and tensor factorizations” Cichocki, Andrzej, and P. H. A. N. Anh-Huy. IEICE transactions on fundamentals of electronics, communications and computer sciences 92.3: 708-721, 2009.
[2]“Algorithms for nonnegative matrix factorization with the beta-divergence” Fevotte, C., & Idier, J. (2011). Neural Computation, 23(9).
Examples
>>> import numpy as np >>> X = np.array([[1,1], [2, 1], [3, 1.2], [4, 1], [5, 0.8], [6, 1]]) >>> from sklearn.decomposition import non_negative_factorization >>> W, H, n_iter = non_negative_factorization( ... X, n_components=2, init='random', random_state=0)