# lasso

Lasso or elastic net regularization for linear models

## Description

example

B = lasso(X,y) returns fitted least-squares regression coefficients for linear models of the predictor data X and the response y. Each column of B corresponds to a particular regularization coefficient in Lambda. By default, lasso performs lasso regularization using a geometric sequence of Lambda values.

example

B = lasso(X,y,Name,Value) fits regularized regressions with additional options specified by one or more name-value pair arguments. For example, 'Alpha',0.5 sets elastic net as the regularization method, with the parameter Alpha equal to 0.5.

example

[B,FitInfo] = lasso(___) also returns the structure FitInfo, which contains information about the fit of the models, using any of the input arguments in the previous syntaxes.

## Examples

collapse all

Construct a data set with redundant predictors and identify those predictors by using lasso.

Create a matrix X of 100 five-dimensional normal variables. Create a response vector y from just two components of X, and add a small amount of noise.

rng default % For reproducibility
X = randn(100,5);
weights = [0;2;0;-3;0]; % Only two nonzero coefficients
y = X*weights + randn(100,1)*0.1; % Small added noise

Construct the default lasso fit.

B = lasso(X,y);

Find the coefficient vector for the 25th Lambda value in B.

B(:,25)
ans = 5×1

0
1.6093
0
-2.5865
0

lasso identifies and removes the redundant predictors.

Create sample data with predictor variable X and response variable $y=0+2X+\epsilon$.

rng('default') % For reproducibility
X = rand(100,1);
y = 2*X + randn(100,1)/10;

Specify a regularization value, and find the coefficient of the regression model without an intercept term.

lambda = 1e-03;
B = lasso(X,y,'Lambda',lambda,'Intercept',false)
Warning: When the 'Intercept' value is false, the 'Standardize' value is set to false.
B = 1.9825

Plot the real values (points) against the predicted values (line).

scatter(X,y)
hold on
x = 0:0.1:1;
plot(x,x*B)
hold off

Construct a data set with redundant predictors and identify those predictors by using cross-validated lasso.

Create a matrix X of 100 five-dimensional normal variables. Create a response vector y from two components of X, and add a small amount of noise.

rng default % For reproducibility
X = randn(100,5);
weights = [0;2;0;-3;0]; % Only two nonzero coefficients
y = X*weights + randn(100,1)*0.1; % Small added noise

Construct the lasso fit by using 10-fold cross-validation with labeled predictor variables.

[B,FitInfo] = lasso(X,y,'CV',10,'PredictorNames',{'x1','x2','x3','x4','x5'});

Display the variables in the model that corresponds to the minimum cross-validated mean squared error (MSE).

idxLambdaMinMSE = FitInfo.IndexMinMSE;
minMSEModelPredictors = FitInfo.PredictorNames(B(:,idxLambdaMinMSE)~=0)
minMSEModelPredictors = 1x2 cell
{'x2'}    {'x4'}

Display the variables in the sparsest model within one standard error of the minimum MSE.

idxLambda1SE = FitInfo.Index1SE;
sparseModelPredictors = FitInfo.PredictorNames(B(:,idxLambda1SE)~=0)
sparseModelPredictors = 1x2 cell
{'x2'}    {'x4'}

In this example, lasso identifies the same predictors for the two models and removes the redundant predictors.

Visually examine the cross-validated error of various levels of regularization.

Load the sample data.

Create a design matrix with interactions and no constant term.

X = [x1 x2 x3];
D = x2fx(X,'interaction');
D(:,1) = []; % No constant term

Construct the lasso fit using 10-fold cross-validation. Include the FitInfo output so you can plot the result.

rng default % For reproducibility
[B,FitInfo] = lasso(D,y,'CV',10);

Plot the cross-validated fits.

lassoPlot(B,FitInfo,'PlotType','CV');
legend('show') % Show legend

The green circle and dotted line locate the Lambda with minimum cross-validation error. The blue circle and dotted line locate the point with minimum cross-validation error plus one standard error.

Predict students' exam scores using lasso and the elastic net method.

Split the data into training and test sets.

n = length(y);
c = cvpartition(n,'HoldOut',0.3);
idxTrain = training(c,1);
idxTest = ~idxTrain;
XTrain = X(idxTrain,:);
yTrain = y(idxTrain);
XTest = X(idxTest,:);
yTest = y(idxTest);

Find the coefficients of a regularized linear regression model using 10-fold cross-validation and the elastic net method with Alpha = 0.75. Use the largest Lambda value such that the mean squared error (MSE) is within one standard error of the minimum MSE.

[B,FitInfo] = lasso(XTrain,yTrain,'Alpha',0.75,'CV',10);
idxLambda1SE = FitInfo.Index1SE;
coef = B(:,idxLambda1SE);
coef0 = FitInfo.Intercept(idxLambda1SE);

Predict exam scores for the test data. Compare the predicted values to the actual exam grades using a reference line.

yhat = XTest*coef + coef0;
hold on
scatter(yTest,yhat)
plot(yTest,yTest)
hold off

Create a matrix X of N p-dimensional normal variables, where N is large and p = 1000. Create a response vector y from the model y = beta0 + X*p, where beta0 is a constant, along with additive noise.

rng default % For reproducibility
N = 1e4; % Number of samples
p = 1e3; % Number of features
X = randn(N,p);
beta = randn(p,1); % Multiplicative coefficients
beta0 = randn; % Additive term
y = beta0 + X*beta + randn(N,1); % Last term is noise

Construct the default lasso fit. Time the creation.

B = lasso(X,y,"UseCovariance",false); % Warm up lasso for reliable timing data
tic
B = lasso(X,y,"UseCovariance",false);
timefalse = toc
timefalse = 3.8125

Construct the lasso fit using the covariance matrix. Time the creation.

B2 = lasso(X,y,"UseCovariance",true); % Warm up lasso for reliable timing data
tic
B2 = lasso(X,y,"UseCovariance",true);
timetrue = toc
timetrue = 0.3062

The fitting time with the covariance matrix is much less than the time without it. View the speedup factor that results from using the covariance matrix.

speedup = timefalse/timetrue
speedup = 12.4503

Check that the returned coefficients B and B2 are similar.

norm(B-B2)/norm(B)
ans = 5.2508e-15

The results are virtually identical.

## Input Arguments

collapse all

Predictor data, specified as a numeric matrix. Each row represents one observation, and each column represents one predictor variable.

Data Types: single | double

Response data, specified as a numeric vector. y has length n, where n is the number of rows of X. The response y(i) corresponds to the ith row of X.

Data Types: single | double

### Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes.

Example: lasso(X,y,'Alpha',0.75,'CV',10) performs elastic net regularization with 10-fold cross-validation. The 'Alpha',0.75 name-value pair argument sets the parameter used in the elastic net optimization.

Absolute error tolerance used to determine the convergence of the ADMM Algorithm, specified as the comma-separated pair consisting of 'AbsTol' and a positive scalar. The algorithm converges when successive estimates of the coefficient vector differ by an amount less than AbsTol.

Note

This option applies only when you use lasso on tall arrays. See Extended Capabilities for more information.

Example: 'AbsTol',1e–3

Data Types: single | double

Weight of lasso (L1) versus ridge (L2) optimization, specified as the comma-separated pair consisting of 'Alpha' and a positive scalar value in the interval (0,1]. The value Alpha = 1 represents lasso regression, Alpha close to 0 approaches ridge regression, and other values represent elastic net optimization. See Elastic Net.

Example: 'Alpha',0.5

Data Types: single | double

Initial values for x-coefficients in ADMM Algorithm, specified as the comma-separated pair consisting of 'B0' and a numeric vector.

Note

This option applies only when you use lasso on tall arrays. See Extended Capabilities for more information.

Data Types: single | double

Size of the covariance matrix in megabytes, specified as a positive scalar or 'maximal'. The lasso function can use a covariance matrix for fitting when the UseCovariance argument is true or 'auto'.

If UseCovariance is true or 'auto' and CacheSize is 'maximal', lasso can attempt to allocate a covariance matrix that exceeds the available memory. In this case, MATLAB® issues an error.

Example: 'CacheSize','maximal'

Data Types: double | char | string

Cross-validation specification for estimating the mean squared error (MSE), specified as the comma-separated pair consisting of 'CV' and one of the following:

• 'resubstitution'lasso uses X and y to fit the model and to estimate the MSE without cross-validation.

• Positive scalar integer Klasso uses K-fold cross-validation.

• cvpartition object cvplasso uses the cross-validation method expressed in cvp. You cannot use a 'leaveout' partition with lasso.

Example: 'CV',3

Maximum number of nonzero coefficients in the model, specified as the comma-separated pair consisting of 'DFmax' and a positive integer scalar. lasso returns results only for Lambda values that satisfy this criterion.

Example: 'DFmax',5

Data Types: single | double

Flag for fitting the model with the intercept term, specified as the comma-separated pair consisting of 'Intercept' and either true or false. The default value is true, which indicates to include the intercept term in the model. If Intercept is false, then the returned intercept value is 0.

Example: 'Intercept',false

Data Types: logical

Regularization coefficients, specified as the comma-separated pair consisting of 'Lambda' and a vector of nonnegative values. See Lasso.

• If you do not supply Lambda, then lasso calculates the largest value of Lambda that gives a nonnull model. In this case, LambdaRatio gives the ratio of the smallest to the largest value of the sequence, and NumLambda gives the length of the vector.

• If you supply Lambda, then lasso ignores LambdaRatio and NumLambda.

• If Standardize is true, then Lambda is the set of values used to fit the models with the X data standardized to have zero mean and a variance of one.

The default is a geometric sequence of NumLambda values, with only the largest value able to produce B = 0.

Example: 'Lambda',linspace(0,1)

Data Types: single | double

Ratio of the smallest to the largest Lambda values when you do not supply Lambda, specified as the comma-separated pair consisting of 'LambdaRatio' and a positive scalar.

If you set LambdaRatio = 0, then lasso generates a default sequence of Lambda values and replaces the smallest one with 0.

Example: 'LambdaRatio',1e–2

Data Types: single | double

Maximum number of iterations allowed, specified as the comma-separated pair consisting of 'MaxIter' and a positive integer scalar.

If the algorithm executes MaxIter iterations before reaching the convergence tolerance RelTol, then the function stops iterating and returns a warning message.

The function can return more than one warning when NumLambda is greater than 1.

Default values are 1e5 for standard data and 1e4 for tall arrays.

Example: 'MaxIter',1e3

Data Types: single | double

Number of Monte Carlo repetitions for cross-validation, specified as the comma-separated pair consisting of 'MCReps' and a positive integer scalar.

• If CV is 'resubstitution' or a cvpartition of type 'resubstitution', then MCReps must be 1.

• If CV is a cvpartition of type 'holdout', then MCReps must be greater than 1.

Example: 'MCReps',5

Data Types: single | double

Number of Lambda values lasso uses when you do not supply Lambda, specified as the comma-separated pair consisting of 'NumLambda' and a positive integer scalar. lasso can return fewer than NumLambda fits if the residual error of the fits drops below a threshold fraction of the variance of y.

Example: 'NumLambda',50

Data Types: single | double

Option to cross-validate in parallel and specify the random streams, specified as the comma-separated pair consisting of 'Options' and a structure. Parallel computation requires Parallel Computing Toolbox™.

Create the Options structure with statset. The option fields are:

• UseParallel — Set to true to compute in parallel. The default is false.

• UseSubstreams — Set to true to compute in parallel in a reproducible fashion. For reproducibility, set Streams to a type allowing substreams: 'mlfg6331_64' or 'mrg32k3a'. The default is false.

• Streams — A RandStream object or cell array consisting of one such object. If you do not specify Streams, then lasso uses the default stream.

Example: 'Options',statset('UseParallel',true)

Data Types: struct

Names of the predictor variables, in the order in which they appear in X, specified as the comma-separated pair consisting of 'PredictorNames' and a string array or cell array of character vectors.

Example: 'PredictorNames',{'x1','x2','x3','x4'}

Data Types: string | cell

Convergence threshold for the coordinate descent algorithm [3], specified as the comma-separated pair consisting of 'RelTol' and a positive scalar. The algorithm terminates when successive estimates of the coefficient vector differ in the L2 norm by a relative amount less than RelTol.

Example: 'RelTol',5e–3

Data Types: single | double

Augmented Lagrangian parameter ρ for the ADMM Algorithm, specified as the comma-separated pair consisting of 'Rho' and a positive scalar. The default is automatic selection.

Note

This option applies only when you use lasso on tall arrays. See Extended Capabilities for more information.

Example: 'Rho',2

Data Types: single | double

Flag for standardizing the predictor data X before fitting the models, specified as the comma-separated pair consisting of 'Standardize' and either true or false. If Standardize is true, then the X data is scaled to have zero mean and a variance of one. Standardize affects whether the regularization is applied to the coefficients on the standardized scale or the original scale. The results are always presented on the original data scale.

If Intercept is false, then the software sets Standardize to false, regardless of the Standardize value you specify.

X and y are always centered when Intercept is true.

Example: 'Standardize',false

Data Types: logical

Indication to use a covariance matrix for fitting, specified as 'auto' or a logical scalar.

• 'auto' causes lasso to attempt to use a covariance matrix for fitting when the number of observations is greater than the number of problem variables. This attempt can fail when memory is insufficient. To find out whether lasso used a covariance matrix for fitting, examine the UseCovariance field of the FitInfo output.

• true causes lasso to use a covariance matrix for fitting as long as the required size does not exceed CacheSize. If the required covariance matrix size exceeds CacheSize, lasso issues a warning and does not use a covariance matrix for fitting.

• false causes lasso not to use a covariance matrix for fitting.

Using a covariance matrix for fitting can be faster than not using one, but can require more memory. See Use Correlation Matrix for Fitting Lasso. The speed increase can negatively affect numerical stability. For details, see Coordinate Descent Algorithm.

Example: 'UseCovariance',true

Data Types: logical | char | string

Initial value of the scaled dual variable u in the ADMM Algorithm, specified as the comma-separated pair consisting of 'U0' and a numeric vector.

Note

This option applies only when you use lasso on tall arrays. See Extended Capabilities for more information.

Data Types: single | double

Observation weights, specified as the comma-separated pair consisting of 'Weights' and a nonnegative vector. Weights has length n, where n is the number of rows of X. The lasso function scales Weights to sum to 1.

Data Types: single | double

## Output Arguments

collapse all

Fitted coefficients, returned as a numeric matrix. B is a p-by-L matrix, where p is the number of predictors (columns) in X, and L is the number of Lambda values. You can specify the number of Lambda values using the NumLambda name-value pair argument.

The coefficient corresponding to the intercept term is a field in FitInfo.

Data Types: single | double

Fit information of the linear models, returned as a structure with the fields described in this table.

Field in FitInfoDescription
InterceptIntercept term β0 for each linear model, a 1-by-L vector
LambdaLambda parameters in ascending order, a 1-by-L vector
AlphaValue of the Alpha parameter, a scalar
DFNumber of nonzero coefficients in B for each value of Lambda, a 1-by-L vector
MSEMean squared error (MSE), a 1-by-L vector
PredictorNamesValue of the PredictorNames parameter, stored as a cell array of character vectors
UseCovarianceLogical value indicating whether the covariance matrix was used in fitting. If the covariance was computed and used, this field is true. Otherwise, this field is false.

If you set the CV name-value pair argument to cross-validate, the FitInfo structure contains these additional fields.

Field in FitInfoDescription
SEStandard error of MSE for each Lambda, as calculated during cross-validation, a 1-by-L vector
LambdaMinMSELambda value with the minimum MSE, a scalar
Lambda1SELargest Lambda value such that MSE is within one standard error of the minimum MSE, a scalar
IndexMinMSEIndex of Lambda with the value LambdaMinMSE, a scalar
Index1SEIndex of Lambda with the value Lambda1SE, a scalar

collapse all

### Lasso

For a given value of λ, a nonnegative parameter, lasso solves the problem

$\underset{{\beta }_{0},\beta }{\mathrm{min}}\left(\frac{1}{2N}\sum _{i=1}^{N}{\left({y}_{i}-{\beta }_{0}-{x}_{i}^{T}\beta \right)}^{2}+\lambda \sum _{j=1}^{p}|{\beta }_{j}|\right).$

• N is the number of observations.

• yi is the response at observation i.

• xi is data, a vector of length p at observation i.

• λ is a nonnegative regularization parameter corresponding to one value of Lambda.

• The parameters β0 and β are a scalar and a vector of length p, respectively.

As λ increases, the number of nonzero components of β decreases.

The lasso problem involves the L1 norm of β, as contrasted with the elastic net algorithm.

### Elastic Net

For α strictly between 0 and 1, and nonnegative λ, elastic net solves the problem

$\underset{{\beta }_{0},\beta }{\mathrm{min}}\left(\frac{1}{2N}\sum _{i=1}^{N}{\left({y}_{i}-{\beta }_{0}-{x}_{i}^{T}\beta \right)}^{2}+\lambda {P}_{\alpha }\left(\beta \right)\right),$

where

${P}_{\alpha }\left(\beta \right)=\frac{\left(1-\alpha \right)}{2}{‖\beta ‖}_{2}^{2}+\alpha {‖\beta ‖}_{1}=\sum _{j=1}^{p}\left(\frac{\left(1-\alpha \right)}{2}{\beta }_{j}^{2}+\alpha |{\beta }_{j}|\right).$

Elastic net is the same as lasso when α = 1. For other values of α, the penalty term Pα(β) interpolates between the L1 norm of β and the squared L2 norm of β. As α shrinks toward 0, elastic net approaches ridge regression.

## Algorithms

collapse all

### Coordinate Descent Algorithm

lasso fits many values of λ simultaneously by an efficient procedure named coordinate descent, based on Friedman, Tibshirani, and Hastie [3]. The procedure has two main code paths depending on whether the fitting uses a covariance matrix. You can affect this choice with the UseCovariance name-value argument.

When lasso uses a covariance matrix to fit N data points and D predictors, the fitting has a rough computational complexity of D*D. Without a covariance matrix, the computational complexity is roughly N*D. So, typically, using a covariance matrix can be faster when N > D, and the default 'auto' setting of the UseCovariance argument makes this choice. Using a covariance matrix causes lasso to subtract larger numbers than otherwise, which can be less numerically stable. For details of the algorithmic differences, see [3]. For one comparison of timing and accuracy differences, see Use Correlation Matrix for Fitting Lasso.

When operating on tall arrays, lasso uses an algorithm based on the Alternating Direction Method of Multipliers (ADMM) [5]. The notation used here is the same as in the reference paper. This method solves problems of the form

Minimize $l\left(x\right)+g\left(z\right)$

Subject to $Ax+Bz=c$

Using this notation, the lasso regression problem is

Minimize $l\left(x\right)+g\left(z\right)=\frac{1}{2}{‖Ax-b‖}_{2}^{2}+\lambda {‖z‖}_{1}$

Subject to $x-z=0$

Because the loss function $l\left(x\right)=\frac{1}{2}{‖Ax-b‖}_{2}^{2}$ is quadratic, the iterative updates performed by the algorithm amount to solving a linear system of equations with a single coefficient matrix but several right-hand sides. The updates performed by the algorithm during each iteration are

$\begin{array}{l}{x}^{k+1}={\left({A}^{T}A+\rho I\right)}^{-1}\left({A}^{T}b+\rho \left({z}^{k}-{u}^{k}\right)\right)\\ {z}^{k+1}={S}_{\lambda /\rho }\left({x}^{k+1}+{u}^{k}\right)\\ {u}^{k+1}={u}^{k}+{x}^{k+1}-{z}^{k+1}\end{array}$

A is the dataset (a tall array), x contains the coefficients, ρ is the penalty parameter (augmented Lagrangian parameter), b is the response (a tall array), and S is the soft thresholding operator.

${S}_{\kappa }\left(a\right)=\left\{\begin{array}{c}\begin{array}{cc}a-\kappa ,\text{\hspace{0.17em}}& a>\kappa \end{array}\\ \begin{array}{cc}0,\text{\hspace{0.17em}}& |a|\text{\hspace{0.17em}}\le \kappa \text{\hspace{0.17em}}\end{array}\\ \begin{array}{cc}a+\kappa ,\text{\hspace{0.17em}}& a<\kappa \text{\hspace{0.17em}}\end{array}\end{array}.$

lasso solves the linear system using Cholesky factorization because the coefficient matrix ${A}^{T}A+\rho I$ is symmetric and positive definite. Because $\rho$ does not change between iterations, the Cholesky factorization is cached between iterations.

Even though A and b are tall arrays, they appear only in the terms ${A}^{T}A$ and ${A}^{T}b$. The results of these two matrix multiplications are small enough to fit in memory, so they are precomputed and the iterative updates between iterations are performed entirely within memory.

## References

[1] Tibshirani, R. “Regression Shrinkage and Selection via the Lasso.” Journal of the Royal Statistical Society. Series B, Vol. 58, No. 1, 1996, pp. 267–288.

[2] Zou, H., and T. Hastie. “Regularization and Variable Selection via the Elastic Net.” Journal of the Royal Statistical Society. Series B, Vol. 67, No. 2, 2005, pp. 301–320.

[3] Friedman, J., R. Tibshirani, and T. Hastie. “Regularization Paths for Generalized Linear Models via Coordinate Descent.” Journal of Statistical Software. Vol. 33, No. 1, 2010. https://www.jstatsoft.org/v33/i01

[4] Hastie, T., R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. 2nd edition. New York: Springer, 2008.

[5] Boyd, S. “Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers.” Foundations and Trends in Machine Learning. Vol. 3, No. 1, 2010, pp. 1–122.

## Version History

Introduced in R2011b