Machine Learning with MATLAB

Lasso Regularization

This example demonstrates the use of lasso for feature selection by looking at a dataset and identifying predictors of diabetes in a population. The dataset contains 10 predictors. The goal is to identify important predictors and discard those that are unnecessary.

View the complete set of data and functions for this demonstration.

Download Data

filename = 'diabetes.txt'; urlwrite('',filename); 

Import data

Once the file is saved, you can import data into MATLAB as a table using the Import Tool with default options. Alternatively, you can use the following code which can be auto generated from the Import Tool:

formatSpec = '%f%f%f%f%f%f%f%f%f%f%f%[^\n\r]';
fileID = fopen(filename,'r');
dataArray = textscan(fileID, formatSpec, 'Delimiter', '\t', 'HeaderLines' ,1, 'ReturnOnError', false);
diabetes = table(dataArray{1:end-1}, 'VariableNames', {'AGE','SEX','BMI','BP','S1','S2','S3','S4','S5','S6','Y'});
clearvars filename delimiter startRow formatSpec fileID dataArray ans;

% Delete the file
delete diabetes.txt

Read the Predictors and Response Variables from the Table

predNames = diabetes.Properties.VariableNames(1:end-1);
X = diabetes{:,1:end-1};
y = diabetes{:,end};

Perform Lasso Regularization

[beta, FitInfo] = lasso(X,y,'Standardize',true,'CV',10,'PredictorNames',predNames);

hlplot = get(gca,'Children');

% Generating colors for each line in the plot
colors = hsv(numel(hlplot));
for ii = 1:numel(hlplot)

set(gcf,'Units','Normalized','Position',[0.2 0.4 0.5 0.35])

Larger values of lambda appear on the left side of the graph, which means that there is increased regularization. As the lambda value increases, the number of nonzero predictors also increases.

Important Predictors

As a rule of thumb, one standard-error value is often used for choosing a smaller model with a good fit.

lam = FitInfo.Index1SE;
isImportant = beta(:,lam) ~= 0;
    'BMI'    'BP'    'S3'    'S5'

Fit a Linear Model with the Terms for Comparison

mdlFull = fitlm(X,y,'Intercept',false);
Linear regression model:
    y ~ x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8 + x9 + x10

Estimated Coefficients:
           Estimate    SE         tStat       pValue    
    x1     0.022296    0.22256     0.10018       0.92025
    x2      -26.073     5.9561     -4.3775    1.5074e-05
    x3       5.3537    0.73462      7.2877    1.5112e-12
    x4       1.0178     0.2304      4.4175    1.2635e-05
    x5       1.2636    0.33044      3.8239    0.00015068
    x6      -1.2849     0.3468     -3.7051    0.00023877
    x7      -3.0683    0.37189     -8.2505    1.9259e-15
    x8       -5.508     5.5883    -0.98565       0.32486
    x9       5.5034     9.4293     0.58365       0.55976
    x10     0.12339     0.2788     0.44256        0.6583

Number of observations: 442, Error degrees of freedom: 432
Root Mean Squared Error: 55.6

Compare the MSE for regularized and unregularized models.

disp(['Lasso MSE: ', num2str(FitInfo.MSE(lam))])
disp(['Full  MSE: ', num2str(mdlFull.MSE)])
Lasso MSE: 3176.5163
Full  MSE: 3092.896

The mean squared error (MSE) of the fit using only the important predictors as determined by lasso, is quite close to the error from the linear model that uses all the predictors. Lasso is often used to prevent overfitting or remove redundant predictors to improve model accuracy.