1.1. Linear Models¶
The following are a set of methods intended for regression in which the target value is expected to be a linear combination of the features. In mathematical notation, if \(\hat{y}\) is the predicted value.
Across the module, we designate the vector \(w = (w_1,
..., w_p)\) as coef_
and \(w_0\) as intercept_
.
To perform classification with generalized linear models, see Logistic regression.
1.1.1. Ordinary Least Squares¶
LinearRegression
fits a linear model with coefficients
\(w = (w_1, ..., w_p)\) to minimize the residual sum
of squares between the observed targets in the dataset, and the
targets predicted by the linear approximation. Mathematically it
solves a problem of the form:
LinearRegression
will take in its fit
method arrays X
, y
and will store the coefficients \(w\) of the linear model in its
coef_
member:
>>> from sklearn import linear_model
>>> reg = linear_model.LinearRegression()
>>> reg.fit([[0, 0], [1, 1], [2, 2]], [0, 1, 2])
LinearRegression()
>>> reg.coef_
array([0.5, 0.5])
The coefficient estimates for Ordinary Least Squares rely on the independence of the features. When features are correlated and the columns of the design matrix \(X\) have an approximately linear dependence, the design matrix becomes close to singular and as a result, the leastsquares estimate becomes highly sensitive to random errors in the observed target, producing a large variance. This situation of multicollinearity can arise, for example, when data are collected without an experimental design.
1.1.1.1. NonNegative Least Squares¶
It is possible to constrain all the coefficients to be nonnegative, which may
be useful when they represent some physical or naturally nonnegative
quantities (e.g., frequency counts or prices of goods).
LinearRegression
accepts a boolean positive
parameter: when set to True
NonNegative Least Squares are then applied.
1.1.1.2. Ordinary Least Squares Complexity¶
The least squares solution is computed using the singular value
decomposition of X. If X is a matrix of shape (n_samples, n_features)
this method has a cost of
\(O(n_{\text{samples}} n_{\text{features}}^2)\), assuming that
\(n_{\text{samples}} \geq n_{\text{features}}\).
1.1.2. Ridge regression and classification¶
1.1.2.1. Regression¶
Ridge
regression addresses some of the problems of
Ordinary Least Squares by imposing a penalty on the size of the
coefficients. The ridge coefficients minimize a penalized residual sum
of squares:
The complexity parameter \(\alpha \geq 0\) controls the amount of shrinkage: the larger the value of \(\alpha\), the greater the amount of shrinkage and thus the coefficients become more robust to collinearity.
As with other linear models, Ridge
will take in its fit
method
arrays X
, y
and will store the coefficients \(w\) of the linear model in
its coef_
member:
>>> from sklearn import linear_model
>>> reg = linear_model.Ridge(alpha=.5)
>>> reg.fit([[0, 0], [0, 0], [1, 1]], [0, .1, 1])
Ridge(alpha=0.5)
>>> reg.coef_
array([0.34545455, 0.34545455])
>>> reg.intercept_
0.13636...
Note that the class Ridge
allows for the user to specify that the
solver be automatically chosen by setting solver="auto"
. When this option
is specified, Ridge
will choose between the "lbfgs"
, "cholesky"
,
and "sparse_cg"
solvers. Ridge
will begin checking the conditions
shown in the following table from top to bottom. If the condition is true,
the corresponding solver is chosen.
Solver 
Condition 
‘lbfgs’ 
The 
‘cholesky’ 
The input array X is not sparse. 
‘sparse_cg’ 
None of the above conditions are fulfilled. 
1.1.2.2. Classification¶
The Ridge
regressor has a classifier variant:
RidgeClassifier
. This classifier first converts binary targets to
{1, 1}
and then treats the problem as a regression task, optimizing the
same objective as above. The predicted class corresponds to the sign of the
regressor’s prediction. For multiclass classification, the problem is
treated as multioutput regression, and the predicted class corresponds to
the output with the highest value.
It might seem questionable to use a (penalized) Least Squares loss to fit a
classification model instead of the more traditional logistic or hinge
losses. However, in practice, all those models can lead to similar
crossvalidation scores in terms of accuracy or precision/recall, while the
penalized least squares loss used by the RidgeClassifier
allows for
a very different choice of the numerical solvers with distinct computational
performance profiles.
The RidgeClassifier
can be significantly faster than e.g.
LogisticRegression
with a high number of classes because it can
compute the projection matrix \((X^T X)^{1} X^T\) only once.
This classifier is sometimes referred to as a Least Squares Support Vector Machines with a linear kernel.
1.1.2.3. Ridge Complexity¶
This method has the same order of complexity as Ordinary Least Squares.
1.1.2.4. Setting the regularization parameter: leaveoneout CrossValidation¶
RidgeCV
implements ridge regression with builtin
crossvalidation of the alpha parameter. The object works in the same way
as GridSearchCV except that it defaults to LeaveOneOut CrossValidation:
>>> import numpy as np
>>> from sklearn import linear_model
>>> reg = linear_model.RidgeCV(alphas=np.logspace(6, 6, 13))
>>> reg.fit([[0, 0], [0, 0], [1, 1]], [0, .1, 1])
RidgeCV(alphas=array([1.e06, 1.e05, 1.e04, 1.e03, 1.e02, 1.e01, 1.e+00, 1.e+01,
1.e+02, 1.e+03, 1.e+04, 1.e+05, 1.e+06]))
>>> reg.alpha_
0.01
Specifying the value of the cv attribute will trigger the use of
crossvalidation with GridSearchCV
, for
example cv=10
for 10fold crossvalidation, rather than LeaveOneOut
CrossValidation.
References Click for more details ¶
“Notes on Regularized Least Squares”, Rifkin & Lippert (technical report, course slides).
1.1.3. Lasso¶
The Lasso
is a linear model that estimates sparse coefficients.
It is useful in some contexts due to its tendency to prefer solutions
with fewer nonzero coefficients, effectively reducing the number of
features upon which the given solution is dependent. For this reason,
Lasso and its variants are fundamental to the field of compressed sensing.
Under certain conditions, it can recover the exact set of nonzero
coefficients (see
Compressive sensing: tomography reconstruction with L1 prior (Lasso)).
Mathematically, it consists of a linear model with an added regularization term. The objective function to minimize is:
The lasso estimate thus solves the minimization of the leastsquares penalty with \(\alpha w_1\) added, where \(\alpha\) is a constant and \(w_1\) is the \(\ell_1\)norm of the coefficient vector.
The implementation in the class Lasso
uses coordinate descent as
the algorithm to fit the coefficients. See Least Angle Regression
for another implementation:
>>> from sklearn import linear_model
>>> reg = linear_model.Lasso(alpha=0.1)
>>> reg.fit([[0, 0], [1, 1]], [0, 1])
Lasso(alpha=0.1)
>>> reg.predict([[1, 1]])
array([0.8])
The function lasso_path
is useful for lowerlevel tasks, as it
computes the coefficients along the full path of possible values.
Note
Feature selection with Lasso
As the Lasso regression yields sparse models, it can thus be used to perform feature selection, as detailed in L1based feature selection.
References Click for more details ¶
The following two references explain the iterations used in the coordinate descent solver of scikitlearn, as well as the duality gap computation used for convergence control.
“Regularization Path For Generalized linear Models by Coordinate Descent”, Friedman, Hastie & Tibshirani, J Stat Softw, 2010 (Paper).
“An InteriorPoint Method for LargeScale L1Regularized Least Squares,” S. J. Kim, K. Koh, M. Lustig, S. Boyd and D. Gorinevsky, in IEEE Journal of Selected Topics in Signal Processing, 2007 (Paper)
1.1.3.1. Setting regularization parameter¶
The alpha
parameter controls the degree of sparsity of the estimated
coefficients.
1.1.3.1.1. Using crossvalidation¶
scikitlearn exposes objects that set the Lasso alpha
parameter by
crossvalidation: LassoCV
and LassoLarsCV
.
LassoLarsCV
is based on the Least Angle Regression algorithm
explained below.
For highdimensional datasets with many collinear features,
LassoCV
is most often preferable. However, LassoLarsCV
has
the advantage of exploring more relevant values of alpha
parameter, and
if the number of samples is very small compared to the number of
features, it is often faster than LassoCV
.
1.1.3.1.2. Informationcriteria based model selection¶
Alternatively, the estimator LassoLarsIC
proposes to use the
Akaike information criterion (AIC) and the Bayes Information criterion (BIC).
It is a computationally cheaper alternative to find the optimal value of alpha
as the regularization path is computed only once instead of k+1 times
when using kfold crossvalidation.
Indeed, these criteria are computed on the insample training set. In short, they penalize the overoptimistic scores of the different Lasso models by their flexibility (cf. to “Mathematical details” section below).
However, such criteria need a proper estimation of the degrees of freedom of the solution, are derived for large samples (asymptotic results) and assume the correct model is candidates under investigation. They also tend to break when the problem is badly conditioned (e.g. more features than samples).
1.1.3.1.3. AIC and BIC criteria¶
The definition of AIC (and thus BIC) might differ in the literature. In this section, we give more information regarding the criterion computed in scikitlearn.
Mathematical details Click for more details ¶
The AIC criterion is defined as:
where \(\hat{L}\) is the maximum likelihood of the model and \(d\) is the number of parameters (as well referred to as degrees of freedom in the previous section).
The definition of BIC replace the constant \(2\) by \(\log(N)\):
where \(N\) is the number of samples.
For a linear Gaussian model, the maximum loglikelihood is defined as:
where \(\sigma^2\) is an estimate of the noise variance, \(y_i\) and \(\hat{y}_i\) are respectively the true and predicted targets, and \(n\) is the number of samples.
Plugging the maximum loglikelihood in the AIC formula yields:
The first term of the above expression is sometimes discarded since it is a constant when \(\sigma^2\) is provided. In addition, it is sometimes stated that the AIC is equivalent to the \(C_p\) statistic [12]. In a strict sense, however, it is equivalent only up to some constant and a multiplicative factor.
At last, we mentioned above that \(\sigma^2\) is an estimate of the
noise variance. In LassoLarsIC
when the parameter noise_variance
is
not provided (default), the noise variance is estimated via the unbiased
estimator [13] defined as:
where \(p\) is the number of features and \(\hat{y}_i\) is the
predicted target using an ordinary least squares regression. Note, that this
formula is valid only when n_samples > n_features
.
1.1.3.1.4. Comparison with the regularization parameter of SVM¶
The equivalence between alpha
and the regularization parameter of SVM,
C
is given by alpha = 1 / C
or alpha = 1 / (n_samples * C)
,
depending on the estimator and the exact objective function optimized by the
model.
1.1.4. Multitask Lasso¶
The MultiTaskLasso
is a linear model that estimates sparse
coefficients for multiple regression problems jointly: y
is a 2D array,
of shape (n_samples, n_tasks)
. The constraint is that the selected
features are the same for all the regression problems, also called tasks.
The following figure compares the location of the nonzero entries in the coefficient matrix W obtained with a simple Lasso or a MultiTaskLasso. The Lasso estimates yield scattered nonzeros while the nonzeros of the MultiTaskLasso are full columns.
Fitting a timeseries model, imposing that any active feature be active at all times.
Examples:
Mathematical details Click for more details ¶
Mathematically, it consists of a linear model trained with a mixed \(\ell_1\) \(\ell_2\)norm for regularization. The objective function to minimize is:
where \(\text{Fro}\) indicates the Frobenius norm
and \(\ell_1\) \(\ell_2\) reads
The implementation in the class MultiTaskLasso
uses
coordinate descent as the algorithm to fit the coefficients.
1.1.5. ElasticNet¶
ElasticNet
is a linear regression model trained with both
\(\ell_1\) and \(\ell_2\)norm regularization of the coefficients.
This combination allows for learning a sparse model where few of
the weights are nonzero like Lasso
, while still maintaining
the regularization properties of Ridge
. We control the convex
combination of \(\ell_1\) and \(\ell_2\) using the l1_ratio
parameter.
Elasticnet is useful when there are multiple features that are correlated with one another. Lasso is likely to pick one of these at random, while elasticnet is likely to pick both.
A practical advantage of tradingoff between Lasso and Ridge is that it allows ElasticNet to inherit some of Ridge’s stability under rotation.
The objective function to minimize is in this case
The class ElasticNetCV
can be used to set the parameters
alpha
(\(\alpha\)) and l1_ratio
(\(\rho\)) by crossvalidation.
Examples:
References Click for more details ¶
The following two references explain the iterations used in the coordinate descent solver of scikitlearn, as well as the duality gap computation used for convergence control.
“Regularization Path For Generalized linear Models by Coordinate Descent”, Friedman, Hastie & Tibshirani, J Stat Softw, 2010 (Paper).
“An InteriorPoint Method for LargeScale L1Regularized Least Squares,” S. J. Kim, K. Koh, M. Lustig, S. Boyd and D. Gorinevsky, in IEEE Journal of Selected Topics in Signal Processing, 2007 (Paper)
1.1.6. Multitask ElasticNet¶
The MultiTaskElasticNet
is an elasticnet model that estimates sparse
coefficients for multiple regression problems jointly: Y
is a 2D array
of shape (n_samples, n_tasks)
. The constraint is that the selected
features are the same for all the regression problems, also called tasks.
Mathematically, it consists of a linear model trained with a mixed \(\ell_1\) \(\ell_2\)norm and \(\ell_2\)norm for regularization. The objective function to minimize is:
The implementation in the class MultiTaskElasticNet
uses coordinate descent as
the algorithm to fit the coefficients.
The class MultiTaskElasticNetCV
can be used to set the parameters
alpha
(\(\alpha\)) and l1_ratio
(\(\rho\)) by crossvalidation.
1.1.7. Least Angle Regression¶
Leastangle regression (LARS) is a regression algorithm for highdimensional data, developed by Bradley Efron, Trevor Hastie, Iain Johnstone and Robert Tibshirani. LARS is similar to forward stepwise regression. At each step, it finds the feature most correlated with the target. When there are multiple features having equal correlation, instead of continuing along the same feature, it proceeds in a direction equiangular between the features.
The advantages of LARS are:
It is numerically efficient in contexts where the number of features is significantly greater than the number of samples.
It is computationally just as fast as forward selection and has the same order of complexity as ordinary least squares.
It produces a full piecewise linear solution path, which is useful in crossvalidation or similar attempts to tune the model.
If two features are almost equally correlated with the target, then their coefficients should increase at approximately the same rate. The algorithm thus behaves as intuition would expect, and also is more stable.
It is easily modified to produce solutions for other estimators, like the Lasso.
The disadvantages of the LARS method include:
Because LARS is based upon an iterative refitting of the residuals, it would appear to be especially sensitive to the effects of noise. This problem is discussed in detail by Weisberg in the discussion section of the Efron et al. (2004) Annals of Statistics article.
The LARS model can be used via the estimator Lars
, or its
lowlevel implementation lars_path
or lars_path_gram
.
1.1.8. LARS Lasso¶
LassoLars
is a lasso model implemented using the LARS
algorithm, and unlike the implementation based on coordinate descent,
this yields the exact solution, which is piecewise linear as a
function of the norm of its coefficients.
>>> from sklearn import linear_model
>>> reg = linear_model.LassoLars(alpha=.1)
>>> reg.fit([[0, 0], [1, 1]], [0, 1])
LassoLars(alpha=0.1)
>>> reg.coef_
array([0.6..., 0. ])
Examples:
The Lars algorithm provides the full path of the coefficients along
the regularization parameter almost for free, thus a common operation
is to retrieve the path with one of the functions lars_path
or lars_path_gram
.
Mathematical formulation Click for more details ¶
The algorithm is similar to forward stepwise regression, but instead of including features at each step, the estimated coefficients are increased in a direction equiangular to each one’s correlations with the residual.
Instead of giving a vector result, the LARS solution consists of a
curve denoting the solution for each value of the \(\ell_1\) norm of the
parameter vector. The full coefficients path is stored in the array
coef_path_
of shape (n_features, max_features + 1)
. The first
column is always zero.
References:
Original Algorithm is detailed in the paper Least Angle Regression by Hastie et al.
1.1.9. Orthogonal Matching Pursuit (OMP)¶
OrthogonalMatchingPursuit
and orthogonal_mp
implement the OMP
algorithm for approximating the fit of a linear model with constraints imposed
on the number of nonzero coefficients (ie. the \(\ell_0\) pseudonorm).
Being a forward feature selection method like Least Angle Regression, orthogonal matching pursuit can approximate the optimum solution vector with a fixed number of nonzero elements:
Alternatively, orthogonal matching pursuit can target a specific error instead of a specific number of nonzero coefficients. This can be expressed as:
OMP is based on a greedy algorithm that includes at each step the atom most highly correlated with the current residual. It is similar to the simpler matching pursuit (MP) method, but better in that at each iteration, the residual is recomputed using an orthogonal projection on the space of the previously chosen dictionary elements.
Examples:
References Click for more details ¶
1.1.10. Bayesian Regression¶
Bayesian regression techniques can be used to include regularization parameters in the estimation procedure: the regularization parameter is not set in a hard sense but tuned to the data at hand.
This can be done by introducing uninformative priors
over the hyper parameters of the model.
The \(\ell_{2}\) regularization used in Ridge regression and classification is
equivalent to finding a maximum a posteriori estimation under a Gaussian prior
over the coefficients \(w\) with precision \(\lambda^{1}\).
Instead of setting lambda
manually, it is possible to treat it as a random
variable to be estimated from the data.
To obtain a fully probabilistic model, the output \(y\) is assumed to be Gaussian distributed around \(X w\):
where \(\alpha\) is again treated as a random variable that is to be estimated from the data.
The advantages of Bayesian Regression are:
It adapts to the data at hand.
It can be used to include regularization parameters in the estimation procedure.
The disadvantages of Bayesian regression include:
Inference of the model can be time consuming.
References Click for more details ¶
A good introduction to Bayesian methods is given in C. Bishop: Pattern Recognition and Machine learning
Original Algorithm is detailed in the book
Bayesian learning for neural networks
by Radford M. Neal
1.1.10.1. Bayesian Ridge Regression¶
BayesianRidge
estimates a probabilistic model of the
regression problem as described above.
The prior for the coefficient \(w\) is given by a spherical Gaussian:
The priors over \(\alpha\) and \(\lambda\) are chosen to be gamma
distributions, the
conjugate prior for the precision of the Gaussian. The resulting model is
called Bayesian Ridge Regression, and is similar to the classical
Ridge
.
The parameters \(w\), \(\alpha\) and \(\lambda\) are estimated
jointly during the fit of the model, the regularization parameters
\(\alpha\) and \(\lambda\) being estimated by maximizing the
log marginal likelihood. The scikitlearn implementation
is based on the algorithm described in Appendix A of (Tipping, 2001)
where the update of the parameters \(\alpha\) and \(\lambda\) is done
as suggested in (MacKay, 1992). The initial value of the maximization procedure
can be set with the hyperparameters alpha_init
and lambda_init
.
There are four more hyperparameters, \(\alpha_1\), \(\alpha_2\), \(\lambda_1\) and \(\lambda_2\) of the gamma prior distributions over \(\alpha\) and \(\lambda\). These are usually chosen to be noninformative. By default \(\alpha_1 = \alpha_2 = \lambda_1 = \lambda_2 = 10^{6}\).
Bayesian Ridge Regression is used for regression:
>>> from sklearn import linear_model
>>> X = [[0., 0.], [1., 1.], [2., 2.], [3., 3.]]
>>> Y = [0., 1., 2., 3.]
>>> reg = linear_model.BayesianRidge()
>>> reg.fit(X, Y)
BayesianRidge()
After being fitted, the model can then be used to predict new values:
>>> reg.predict([[1, 0.]])
array([0.50000013])
The coefficients \(w\) of the model can be accessed:
>>> reg.coef_
array([0.49999993, 0.49999993])
Due to the Bayesian framework, the weights found are slightly different to the ones found by Ordinary Least Squares. However, Bayesian Ridge Regression is more robust to illposed problems.
Examples:
References Click for more details ¶
Section 3.3 in Christopher M. Bishop: Pattern Recognition and Machine Learning, 2006
David J. C. MacKay, Bayesian Interpolation, 1992.
Michael E. Tipping, Sparse Bayesian Learning and the Relevance Vector Machine, 2001.
1.1.10.2. Automatic Relevance Determination  ARD¶
The Automatic Relevance Determination (as being implemented in
ARDRegression
) is a kind of linear model which is very similar to the
Bayesian Ridge Regression, but that leads to sparser coefficients \(w\)
[1] [2].
ARDRegression
poses a different prior over \(w\): it drops
the spherical Gaussian distribution for a centered elliptic Gaussian
distribution. This means each coefficient \(w_{i}\) can itself be drawn from
a Gaussian distribution, centered on zero and with a precision
\(\lambda_{i}\):
with \(A\) being a positive definite diagonal matrix and \(\text{diag}(A) = \lambda = \{\lambda_{1},...,\lambda_{p}\}\).
In contrast to the Bayesian Ridge Regression, each coordinate of \(w_{i}\) has its own standard deviation \(\frac{1}{\lambda_i}\). The prior over all \(\lambda_i\) is chosen to be the same gamma distribution given by the hyperparameters \(\lambda_1\) and \(\lambda_2\).
ARD is also known in the literature as Sparse Bayesian Learning and Relevance Vector Machine [3] [4]. For a workedout comparison between ARD and Bayesian Ridge Regression, see the example below.
Examples:
References:
Christopher M. Bishop: Pattern Recognition and Machine Learning, Chapter 7.2.1
David Wipf and Srikantan Nagarajan: A New View of Automatic Relevance Determination
Michael E. Tipping: Sparse Bayesian Learning and the Relevance Vector Machine
Tristan Fletcher: Relevance Vector Machines Explained
1.1.11. Logistic regression¶
The logistic regression is implemented in LogisticRegression
. Despite
its name, it is implemented as a linear model for classification rather than
regression in terms of the scikitlearn/ML nomenclature. The logistic
regression is also known in the literature as logit regression,
maximumentropy classification (MaxEnt) or the loglinear classifier. In this
model, the probabilities describing the possible outcomes of a single trial
are modeled using a logistic function.
This implementation can fit binary, OnevsRest, or multinomial logistic regression with optional \(\ell_1\), \(\ell_2\) or ElasticNet regularization.
Note
Regularization
Regularization is applied by default, which is common in machine learning but not in statistics. Another advantage of regularization is that it improves numerical stability. No regularization amounts to setting C to a very high value.
Note
Logistic Regression as a special case of the Generalized Linear Models (GLM)
Logistic regression is a special case of Generalized Linear Models with a Binomial / Bernoulli conditional distribution and a Logit link. The numerical output of the logistic regression, which is the predicted probability, can be used as a classifier by applying a threshold (by default 0.5) to it. This is how it is implemented in scikitlearn, so it expects a categorical target, making the Logistic Regression a classifier.
Examples
1.1.11.1. Binary Case¶
For notational ease, we assume that the target \(y_i\) takes values in the
set \(\{0, 1\}\) for data point \(i\).
Once fitted, the predict_proba
method of LogisticRegression
predicts
the probability of the positive class \(P(y_i=1X_i)\) as
As an optimization problem, binary class logistic regression with regularization term \(r(w)\) minimizes the following cost function:
where \({s_i}\) corresponds to the weights assigned by the user to a specific training sample (the vector \(s\) is formed by elementwise multiplication of the class weights and sample weights).
We currently provide four choices for the regularization term \(r(w)\) via
the penalty
argument:
penalty 
\(r(w)\) 


\(0\) 
\(\ell_1\) 
\(\w\_1\) 
\(\ell_2\) 
\(\frac{1}{2}\w\_2^2 = \frac{1}{2}w^T w\) 

\(\frac{1  \rho}{2}w^T w + \rho \w\_1\) 
For ElasticNet, \(\rho\) (which corresponds to the l1_ratio
parameter)
controls the strength of \(\ell_1\) regularization vs. \(\ell_2\)
regularization. ElasticNet is equivalent to \(\ell_1\) when
\(\rho = 1\) and equivalent to \(\ell_2\) when \(\rho=0\).
Note that the scale of the class weights and the sample weights will influence
the optimization problem. For instance, multiplying the sample weights by a
constant \(b>0\) is equivalent to multiplying the (inverse) regularization
strength C
by \(b\).
1.1.11.2. Multinomial Case¶
The binary case can be extended to \(K\) classes leading to the multinomial logistic regression, see also loglinear model.
Note
It is possible to parameterize a \(K\)class classification model using only \(K1\) weight vectors, leaving one class probability fully determined by the other class probabilities by leveraging the fact that all class probabilities must sum to one. We deliberately choose to overparameterize the model using \(K\) weight vectors for ease of implementation and to preserve the symmetrical inductive bias regarding ordering of classes, see [16]. This effect becomes especially important when using regularization. The choice of overparameterization can be detrimental for unpenalized models since then the solution may not be unique, as shown in [16].
Mathematical details Click for more details ¶
Let \(y_i \in {1, \ldots, K}\) be the label (ordinal) encoded target variable for observation \(i\).
Instead of a single coefficient vector, we now have
a matrix of coefficients \(W\) where each row vector \(W_k\) corresponds to class
\(k\). We aim at predicting the class probabilities \(P(y_i=kX_i)\) via
predict_proba
as:
The objective for the optimization becomes
Where \([P]\) represents the Iverson bracket which evaluates to \(0\)
if \(P\) is false, otherwise it evaluates to \(1\). We currently provide four choices
for the regularization term \(r(W)\) via the penalty
argument, where \(m\)
is the number of features:
penalty 
\(r(W)\) 


\(0\) 
\(\ell_1\) 
\(\W\_{1,1} = \sum_{i=1}^m\sum_{j=1}^{K}W_{i,j}\) 
\(\ell_2\) 
\(\frac{1}{2}\W\_F^2 = \frac{1}{2}\sum_{i=1}^m\sum_{j=1}^{K} W_{i,j}^2\) 

\(\frac{1  \rho}{2}\W\_F^2 + \rho \W\_{1,1}\) 
1.1.11.3. Solvers¶
The solvers implemented in the class LogisticRegression
are “lbfgs”, “liblinear”, “newtoncg”, “newtoncholesky”, “sag” and “saga”:
The following table summarizes the penalties supported by each solver:
Solvers 

Penalties 
‘lbfgs’ 
‘liblinear’ 
‘newtoncg’ 
‘newtoncholesky’ 
‘sag’ 
‘saga’ 

Multinomial + L2 penalty 
yes 
no 
yes 
no 
yes 
yes 

OVR + L2 penalty 
yes 
yes 
yes 
yes 
yes 
yes 

Multinomial + L1 penalty 
no 
no 
no 
no 
no 
yes 

OVR + L1 penalty 
no 
yes 
no 
no 
no 
yes 

ElasticNet 
no 
no 
no 
no 
no 
yes 

No penalty (‘none’) 
yes 
no 
yes 
yes 
yes 
yes 

Behaviors 

Penalize the intercept (bad) 
no 
yes 
no 
no 
no 
no 

Faster for large datasets 
no 
no 
no 
no 
yes 
yes 

Robust to unscaled datasets 
yes 
yes 
yes 
yes 
no 
no 
The “lbfgs” solver is used by default for its robustness. For large datasets
the “saga” solver is usually faster.
For large dataset, you may also consider using SGDClassifier
with loss="log_loss"
, which might be even faster but requires more tuning.
1.1.11.3.1. Differences between solvers¶
There might be a difference in the scores obtained between
LogisticRegression
with solver=liblinear
or
LinearSVC
and the external liblinear library directly,
when fit_intercept=False
and the fit coef_
(or) the data to be predicted
are zeroes. This is because for the sample(s) with decision_function
zero,
LogisticRegression
and LinearSVC
predict the
negative class, while liblinear predicts the positive class. Note that a model
with fit_intercept=False
and having many samples with decision_function
zero, is likely to be a underfit, bad model and you are advised to set
fit_intercept=True
and increase the intercept_scaling
.
Solvers’ details Click for more details ¶
The solver “liblinear” uses a coordinate descent (CD) algorithm, and relies on the excellent C++ LIBLINEAR library, which is shipped with scikitlearn. However, the CD algorithm implemented in liblinear cannot learn a true multinomial (multiclass) model; instead, the optimization problem is decomposed in a “onevsrest” fashion so separate binary classifiers are trained for all classes. This happens under the hood, so
LogisticRegression
instances using this solver behave as multiclass classifiers. For \(\ell_1\) regularizationsklearn.svm.l1_min_c
allows to calculate the lower bound for C in order to get a non “null” (all feature weights to zero) model.The “lbfgs”, “newtoncg” and “sag” solvers only support \(\ell_2\) regularization or no regularization, and are found to converge faster for some highdimensional data. Setting
multi_class
to “multinomial” with these solvers learns a true multinomial logistic regression model [5], which means that its probability estimates should be better calibrated than the default “onevsrest” setting.The “sag” solver uses Stochastic Average Gradient descent [6]. It is faster than other solvers for large datasets, when both the number of samples and the number of features are large.
The “saga” solver [7] is a variant of “sag” that also supports the nonsmooth
penalty="l1"
. This is therefore the solver of choice for sparse multinomial logistic regression. It is also the only solver that supportspenalty="elasticnet"
.The “lbfgs” is an optimization algorithm that approximates the Broyden–Fletcher–Goldfarb–Shanno algorithm [8], which belongs to quasiNewton methods. As such, it can deal with a wide range of different training data and is therefore the default solver. Its performance, however, suffers on poorly scaled datasets and on datasets with onehot encoded categorical features with rare categories.
The “newtoncholesky” solver is an exact Newton solver that calculates the hessian matrix and solves the resulting linear system. It is a very good choice for
n_samples
>>n_features
, but has a few shortcomings: Only \(\ell_2\) regularization is supported. Furthermore, because the hessian matrix is explicitly computed, the memory usage has a quadratic dependency onn_features
as well as onn_classes
. As a consequence, only the onevsrest scheme is implemented for the multiclass case.
For a comparison of some of these solvers, see [9].
References:
Christopher M. Bishop: Pattern Recognition and Machine Learning, Chapter 4.3.4
Mark Schmidt, Nicolas Le Roux, and Francis Bach: Minimizing Finite Sums with the Stochastic Average Gradient.
Aaron Defazio, Francis Bach, Simon LacosteJulien: SAGA: A Fast Incremental Gradient Method With Support for NonStrongly Convex Composite Objectives.
https://en.wikipedia.org/wiki/Broyden%E2%80%93Fletcher%E2%80%93Goldfarb%E2%80%93Shanno_algorithm
Thomas P. Minka “A comparison of numerical optimizers for logistic regression”
Note
Feature selection with sparse logistic regression
A logistic regression with \(\ell_1\) penalty yields sparse models, and can thus be used to perform feature selection, as detailed in L1based feature selection.
Note
Pvalue estimation
It is possible to obtain the pvalues and confidence intervals for coefficients in cases of regression without penalization. The statsmodels package natively supports this. Within sklearn, one could use bootstrapping instead as well.
LogisticRegressionCV
implements Logistic Regression with builtin
crossvalidation support, to find the optimal C
and l1_ratio
parameters
according to the scoring
attribute. The “newtoncg”, “sag”, “saga” and
“lbfgs” solvers are found to be faster for highdimensional dense data, due
to warmstarting (see Glossary).
1.1.12. Generalized Linear Models¶
Generalized Linear Models (GLM) extend linear models in two ways [10]. First, the predicted values \(\hat{y}\) are linked to a linear combination of the input variables \(X\) via an inverse link function \(h\) as
Secondly, the squared loss function is replaced by the unit deviance \(d\) of a distribution in the exponential family (or more precisely, a reproductive exponential dispersion model (EDM) [11]).
The minimization problem becomes:
where \(\alpha\) is the L2 regularization penalty. When sample weights are provided, the average becomes a weighted average.
The following table lists some specific EDMs and their unit deviance :
Distribution 
Target Domain 
Unit Deviance \(d(y, \hat{y})\) 

Normal 
\(y \in (\infty, \infty)\) 
\((y\hat{y})^2\) 
Bernoulli 
\(y \in \{0, 1\}\) 
\(2({y}\log\frac{y}{\hat{y}}+({1}{y})\log\frac{{1}{y}}{{1}\hat{y}})\) 
Categorical 
\(y \in \{0, 1, ..., k\}\) 
\(2\sum_{i \in \{0, 1, ..., k\}} I(y = i) y_\text{i}\log\frac{I(y = i)}{\hat{I(y = i)}}\) 
Poisson 
\(y \in [0, \infty)\) 
\(2(y\log\frac{y}{\hat{y}}y+\hat{y})\) 
Gamma 
\(y \in (0, \infty)\) 
\(2(\log\frac{\hat{y}}{y}+\frac{y}{\hat{y}}1)\) 
Inverse Gaussian 
\(y \in (0, \infty)\) 
\(\frac{(y\hat{y})^2}{y\hat{y}^2}\) 
The Probability Density Functions (PDF) of these distributions are illustrated in the following figure,
The Bernoulli distribution is a discrete probability distribution modelling a Bernoulli trial  an event that has only two mutually exclusive outcomes. The Categorical distribution is a generalization of the Bernoulli distribution for a categorical random variable. While a random variable in a Bernoulli distribution has two possible outcomes, a Categorical random variable can take on one of K possible categories, with the probability of each category specified separately.
The choice of the distribution depends on the problem at hand:
If the target values \(y\) are counts (nonnegative integer valued) or relative frequencies (nonnegative), you might use a Poisson distribution with a loglink.
If the target values are positive valued and skewed, you might try a Gamma distribution with a loglink.
If the target values seem to be heavier tailed than a Gamma distribution, you might try an Inverse Gaussian distribution (or even higher variance powers of the Tweedie family).
If the target values \(y\) are probabilities, you can use the Bernoulli distribution. The Bernoulli distribution with a logit link can be used for binary classification. The Categorical distribution with a softmax link can be used for multiclass classification.
Examples of use cases Click for more details ¶
Agriculture / weather modeling: number of rain events per year (Poisson), amount of rainfall per event (Gamma), total rainfall per year (Tweedie / Compound Poisson Gamma).
Risk modeling / insurance policy pricing: number of claim events / policyholder per year (Poisson), cost per event (Gamma), total cost per policyholder per year (Tweedie / Compound Poisson Gamma).
Credit Default: probability that a loan can’t be paid back (Bernoulli).
Fraud Detection: probability that a financial transaction like a cash transfer is a fraudulent transaction (Bernoulli).
Predictive maintenance: number of production interruption events per year (Poisson), duration of interruption (Gamma), total interruption time per year (Tweedie / Compound Poisson Gamma).
Medical Drug Testing: probability of curing a patient in a set of trials or probability that a patient will experience side effects (Bernoulli).
News Classification: classification of news articles into three categories namely Business News, Politics and Entertainment news (Categorical).
References:
McCullagh, Peter; Nelder, John (1989). Generalized Linear Models, Second Edition. Boca Raton: Chapman and Hall/CRC. ISBN 0412317605.
Jørgensen, B. (1992). The theory of exponential dispersion models and analysis of deviance. Monografias de matemática, no. 51. See also Exponential dispersion model.
1.1.12.1. Usage¶
TweedieRegressor
implements a generalized linear model for the
Tweedie distribution, that allows to model any of the above mentioned
distributions using the appropriate power
parameter. In particular:
power = 0
: Normal distribution. Specific estimators such asRidge
,ElasticNet
are generally more appropriate in this case.power = 1
: Poisson distribution.PoissonRegressor
is exposed for convenience. However, it is strictly equivalent toTweedieRegressor(power=1, link='log')
.power = 2
: Gamma distribution.GammaRegressor
is exposed for convenience. However, it is strictly equivalent toTweedieRegressor(power=2, link='log')
.power = 3
: Inverse Gaussian distribution.
The link function is determined by the link
parameter.
Usage example:
>>> from sklearn.linear_model import TweedieRegressor
>>> reg = TweedieRegressor(power=1, alpha=0.5, link='log')
>>> reg.fit([[0, 0], [0, 1], [2, 2]], [0, 1, 2])
TweedieRegressor(alpha=0.5, link='log', power=1)
>>> reg.coef_
array([0.2463..., 0.4337...])
>>> reg.intercept_
0.7638...
Examples
Practical considerations Click for more details ¶
The feature matrix X
should be standardized before fitting. This ensures
that the penalty treats features equally.
Since the linear predictor \(Xw\) can be negative and Poisson,
Gamma and Inverse Gaussian distributions don’t support negative values, it
is necessary to apply an inverse link function that guarantees the
nonnegativeness. For example with link='log'
, the inverse link function
becomes \(h(Xw)=\exp(Xw)\).
If you want to model a relative frequency, i.e. counts per exposure (time, volume, …) you can do so by using a Poisson distribution and passing \(y=\frac{\mathrm{counts}}{\mathrm{exposure}}\) as target values together with \(\mathrm{exposure}\) as sample weights. For a concrete example see e.g. Tweedie regression on insurance claims.
When performing crossvalidation for the power
parameter of
TweedieRegressor
, it is advisable to specify an explicit scoring
function,
because the default scorer TweedieRegressor.score
is a function of
power
itself.
1.1.13. Stochastic Gradient Descent  SGD¶
Stochastic gradient descent is a simple yet very efficient approach
to fit linear models. It is particularly useful when the number of samples
(and the number of features) is very large.
The partial_fit
method allows online/outofcore learning.
The classes SGDClassifier
and SGDRegressor
provide
functionality to fit linear models for classification and regression
using different (convex) loss functions and different penalties.
E.g., with loss="log"
, SGDClassifier
fits a logistic regression model,
while with loss="hinge"
it fits a linear support vector machine (SVM).
You can refer to the dedicated Stochastic Gradient Descent documentation section for more details.
1.1.14. Perceptron¶
The Perceptron
is another simple classification algorithm suitable for
large scale learning. By default:
It does not require a learning rate.
It is not regularized (penalized).
It updates its model only on mistakes.
The last characteristic implies that the Perceptron is slightly faster to train than SGD with the hinge loss and that the resulting models are sparser.
In fact, the Perceptron
is a wrapper around the SGDClassifier
class using a perceptron loss and a constant learning rate. Refer to
mathematical section of the SGD procedure
for more details.
1.1.15. Passive Aggressive Algorithms¶
The passiveaggressive algorithms are a family of algorithms for largescale
learning. They are similar to the Perceptron in that they do not require a
learning rate. However, contrary to the Perceptron, they include a
regularization parameter C
.
For classification, PassiveAggressiveClassifier
can be used with
loss='hinge'
(PAI) or loss='squared_hinge'
(PAII). For regression,
PassiveAggressiveRegressor
can be used with
loss='epsilon_insensitive'
(PAI) or
loss='squared_epsilon_insensitive'
(PAII).
References Click for more details ¶
“Online PassiveAggressive Algorithms” K. Crammer, O. Dekel, J. Keshat, S. ShalevShwartz, Y. Singer  JMLR 7 (2006)
1.1.16. Robustness regression: outliers and modeling errors¶
Robust regression aims to fit a regression model in the presence of corrupt data: either outliers, or error in the model.
1.1.16.1. Different scenario and useful concepts¶
There are different things to keep in mind when dealing with data corrupted by outliers:
Outliers in X or in y?
Outliers in the y direction
Outliers in the X direction
Fraction of outliers versus amplitude of error
The number of outlying points matters, but also how much they are outliers.
Small outliers
Large outliers
An important notion of robust fitting is that of breakdown point: the fraction of data that can be outlying for the fit to start missing the inlying data.
Note that in general, robust fitting in highdimensional setting (large
n_features
) is very hard. The robust models here will probably not work
in these settings.
Tradeoffs: which estimator ?
Scikitlearn provides 3 robust regression estimators: RANSAC, Theil Sen and HuberRegressor.
HuberRegressor should be faster than RANSAC and Theil Sen unless the number of samples are very large, i.e.
n_samples
>>n_features
. This is because RANSAC and Theil Sen fit on smaller subsets of the data. However, both Theil Sen and RANSAC are unlikely to be as robust as HuberRegressor for the default parameters.RANSAC is faster than Theil Sen and scales much better with the number of samples.
RANSAC will deal better with large outliers in the y direction (most common situation).
Theil Sen will cope better with mediumsize outliers in the X direction, but this property will disappear in highdimensional settings.
When in doubt, use RANSAC.
1.1.16.2. RANSAC: RANdom SAmple Consensus¶
RANSAC (RANdom SAmple Consensus) fits a model from random subsets of inliers from the complete data set.
RANSAC is a nondeterministic algorithm producing only a reasonable result with
a certain probability, which is dependent on the number of iterations (see
max_trials
parameter). It is typically used for linear and nonlinear
regression problems and is especially popular in the field of photogrammetric
computer vision.
The algorithm splits the complete input sample data into a set of inliers, which may be subject to noise, and outliers, which are e.g. caused by erroneous measurements or invalid hypotheses about the data. The resulting model is then estimated only from the determined inliers.
Examples
Details of the algorithm Click for more details ¶
Each iteration performs the following steps:
Select
min_samples
random samples from the original data and check whether the set of data is valid (seeis_data_valid
).Fit a model to the random subset (
base_estimator.fit
) and check whether the estimated model is valid (seeis_model_valid
).Classify all data as inliers or outliers by calculating the residuals to the estimated model (
base_estimator.predict(X)  y
)  all data samples with absolute residuals smaller than or equal to theresidual_threshold
are considered as inliers.Save fitted model as best model if number of inlier samples is maximal. In case the current estimated model has the same number of inliers, it is only considered as the best model if it has better score.
These steps are performed either a maximum number of times (max_trials
) or
until one of the special stop criteria are met (see stop_n_inliers
and
stop_score
). The final model is estimated using all inlier samples (consensus
set) of the previously determined best model.
The is_data_valid
and is_model_valid
functions allow to identify and reject
degenerate combinations of random subsamples. If the estimated model is not
needed for identifying degenerate cases, is_data_valid
should be used as it
is called prior to fitting the model and thus leading to better computational
performance.
References Click for more details ¶
“Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography” Martin A. Fischler and Robert C. Bolles  SRI International (1981)
“Performance Evaluation of RANSAC Family” Sunglok Choi, Taemin Kim and Wonpil Yu  BMVC (2009)
1.1.16.3. TheilSen estimator: generalizedmedianbased estimator¶
The TheilSenRegressor
estimator uses a generalization of the median in
multiple dimensions. It is thus robust to multivariate outliers. Note however
that the robustness of the estimator decreases quickly with the dimensionality
of the problem. It loses its robustness properties and becomes no
better than an ordinary least squares in high dimension.
Examples:
Theoretical considerations Click for more details ¶
TheilSenRegressor
is comparable to the Ordinary Least Squares
(OLS) in terms of asymptotic efficiency and as an
unbiased estimator. In contrast to OLS, TheilSen is a nonparametric
method which means it makes no assumption about the underlying
distribution of the data. Since TheilSen is a medianbased estimator, it
is more robust against corrupted data aka outliers. In univariate
setting, TheilSen has a breakdown point of about 29.3% in case of a
simple linear regression which means that it can tolerate arbitrary
corrupted data of up to 29.3%.
The implementation of TheilSenRegressor
in scikitlearn follows a
generalization to a multivariate linear regression model [14] using the
spatial median which is a generalization of the median to multiple
dimensions [15].
In terms of time and space complexity, TheilSen scales according to
which makes it infeasible to be applied exhaustively to problems with a large number of samples and features. Therefore, the magnitude of a subpopulation can be chosen to limit the time and space complexity by considering only a random subset of all possible combinations.
References:
Xin Dang, Hanxiang Peng, Xueqin Wang and Heping Zhang: TheilSen Estimators in a Multiple Linear Regression Model.
Kärkkäinen and S. Äyrämö: On Computation of Spatial Median for Robust Data Mining.
Also see the Wikipedia page
1.1.16.4. Huber Regression¶
The HuberRegressor
is different to Ridge
because it applies a
linear loss to samples that are classified as outliers.
A sample is classified as an inlier if the absolute error of that sample is
lesser than a certain threshold. It differs from TheilSenRegressor
and RANSACRegressor
because it does not ignore the effect of the outliers
but gives a lesser weight to them.
Examples:
Mathematical details Click for more details ¶
The loss function that HuberRegressor
minimizes is given by
where
It is advised to set the parameter epsilon
to 1.35 to achieve 95%
statistical efficiency.
References:
Peter J. Huber, Elvezio M. Ronchetti: Robust Statistics, Concomitant scale estimates, pg 172
The HuberRegressor
differs from using SGDRegressor
with loss set to huber
in the following ways.
HuberRegressor
is scaling invariant. Onceepsilon
is set, scalingX
andy
down or up by different values would produce the same robustness to outliers as before. as compared toSGDRegressor
whereepsilon
has to be set again whenX
andy
are scaled.HuberRegressor
should be more efficient to use on data with small number of samples whileSGDRegressor
needs a number of passes on the training data to produce the same robustness.
Note that this estimator is different from the R implementation of Robust Regression (https://stats.oarc.ucla.edu/r/dae/robustregression/) because the R implementation does a weighted least squares implementation with weights given to each sample on the basis of how much the residual is greater than a certain threshold.
1.1.17. Quantile Regression¶
Quantile regression estimates the median or other quantiles of \(y\) conditional on \(X\), while ordinary least squares (OLS) estimates the conditional mean.
Quantile regression may be useful if one is interested in predicting an interval instead of point prediction. Sometimes, prediction intervals are calculated based on the assumption that prediction error is distributed normally with zero mean and constant variance. Quantile regression provides sensible prediction intervals even for errors with nonconstant (but predictable) variance or nonnormal distribution.
Based on minimizing the pinball loss, conditional quantiles can also be
estimated by models other than linear models. For example,
GradientBoostingRegressor
can predict conditional
quantiles if its parameter loss
is set to "quantile"
and parameter
alpha
is set to the quantile that should be predicted. See the example in
Prediction Intervals for Gradient Boosting Regression.
Most implementations of quantile regression are based on linear programming
problem. The current implementation is based on
scipy.optimize.linprog
.
Examples:
Mathematical details Click for more details ¶
As a linear model, the QuantileRegressor
gives linear predictions
\(\hat{y}(w, X) = Xw\) for the \(q\)th quantile, \(q \in (0, 1)\).
The weights or coefficients \(w\) are then found by the following
minimization problem:
This consists of the pinball loss (also known as linear loss),
see also mean_pinball_loss
,
and the L1 penalty controlled by parameter alpha
, similar to
Lasso
.
As the pinball loss is only linear in the residuals, quantile regression is
much more robust to outliers than squared error based estimation of the mean.
Somewhat in between is the HuberRegressor
.
References Click for more details ¶
Koenker, R., & Bassett Jr, G. (1978). Regression quantiles. Econometrica: journal of the Econometric Society, 3350.
Portnoy, S., & Koenker, R. (1997). The Gaussian hare and the Laplacian tortoise: computability of squarederror versus absoluteerror estimators. Statistical Science, 12, 279300.
Koenker, R. (2005). Quantile Regression. Cambridge University Press.
1.1.18. Polynomial regression: extending linear models with basis functions¶
One common pattern within machine learning is to use linear models trained on nonlinear functions of the data. This approach maintains the generally fast performance of linear methods, while allowing them to fit a much wider range of data.
Mathematical details Click for more details ¶
For example, a simple linear regression can be extended by constructing polynomial features from the coefficients. In the standard linear regression case, you might have a model that looks like this for twodimensional data:
If we want to fit a paraboloid to the data instead of a plane, we can combine the features in secondorder polynomials, so that the model looks like this:
The (sometimes surprising) observation is that this is still a linear model: to see this, imagine creating a new set of features
With this relabeling of the data, our problem can be written
We see that the resulting polynomial regression is in the same class of linear models we considered above (i.e. the model is linear in \(w\)) and can be solved by the same techniques. By considering linear fits within a higherdimensional space built with these basis functions, the model has the flexibility to fit a much broader range of data.
Here is an example of applying this idea to onedimensional data, using polynomial features of varying degrees:
This figure is created using the PolynomialFeatures
transformer, which
transforms an input data matrix into a new data matrix of a given degree.
It can be used as follows:
>>> from sklearn.preprocessing import PolynomialFeatures
>>> import numpy as np
>>> X = np.arange(6).reshape(3, 2)
>>> X
array([[0, 1],
[2, 3],
[4, 5]])
>>> poly = PolynomialFeatures(degree=2)
>>> poly.fit_transform(X)
array([[ 1., 0., 1., 0., 0., 1.],
[ 1., 2., 3., 4., 6., 9.],
[ 1., 4., 5., 16., 20., 25.]])
The features of X
have been transformed from \([x_1, x_2]\) to
\([1, x_1, x_2, x_1^2, x_1 x_2, x_2^2]\), and can now be used within
any linear model.
This sort of preprocessing can be streamlined with the Pipeline tools. A single object representing a simple polynomial regression can be created and used as follows:
>>> from sklearn.preprocessing import PolynomialFeatures
>>> from sklearn.linear_model import LinearRegression
>>> from sklearn.pipeline import Pipeline
>>> import numpy as np
>>> model = Pipeline([('poly', PolynomialFeatures(degree=3)),
... ('linear', LinearRegression(fit_intercept=False))])
>>> # fit to an order3 polynomial data
>>> x = np.arange(5)
>>> y = 3  2 * x + x ** 2  x ** 3
>>> model = model.fit(x[:, np.newaxis], y)
>>> model.named_steps['linear'].coef_
array([ 3., 2., 1., 1.])
The linear model trained on polynomial features is able to exactly recover the input polynomial coefficients.
In some cases it’s not necessary to include higher powers of any single feature,
but only the socalled interaction features
that multiply together at most \(d\) distinct features.
These can be gotten from PolynomialFeatures
with the setting
interaction_only=True
.
For example, when dealing with boolean features, \(x_i^n = x_i\) for all \(n\) and is therefore useless; but \(x_i x_j\) represents the conjunction of two booleans. This way, we can solve the XOR problem with a linear classifier:
>>> from sklearn.linear_model import Perceptron
>>> from sklearn.preprocessing import PolynomialFeatures
>>> import numpy as np
>>> X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
>>> y = X[:, 0] ^ X[:, 1]
>>> y
array([0, 1, 1, 0])
>>> X = PolynomialFeatures(interaction_only=True).fit_transform(X).astype(int)
>>> X
array([[1, 0, 0, 0],
[1, 0, 1, 0],
[1, 1, 0, 0],
[1, 1, 1, 1]])
>>> clf = Perceptron(fit_intercept=False, max_iter=10, tol=None,
... shuffle=False).fit(X, y)
And the classifier “predictions” are perfect:
>>> clf.predict(X)
array([0, 1, 1, 0])
>>> clf.score(X, y)
1.0