Main Content

Regularize Discriminant Analysis Classifier

This example shows how to make a more robust and simpler model by trying to remove predictors without hurting the predictive power of the model. This is especially important when you have many predictors in your data. Linear discriminant analysis uses the two regularization parameters, Gamma and Delta, to identify and remove redundant predictors. The cvshrink method helps identify appropriate settings for these parameters.

Load data and create a classifier.

Create a linear discriminant analysis classifier for the ovariancancer data. Set the SaveMemory and FillCoeffs name-value pair arguments to keep the resulting model reasonably small. For computational ease, this example uses a random subset of about one third of the predictors to train the classifier.

load ovariancancer
rng(1); % For reproducibility
numPred = size(obs,2);
obs = obs(:,randsample(numPred,ceil(numPred/3)));
Mdl = fitcdiscr(obs,grp,'SaveMemory','on','FillCoeffs','off');

Cross validate the classifier.

Use 25 levels of Gamma and 25 levels of Delta to search for good parameters. This search is time consuming. Set Verbose to 1 to view the progress.

[err,gamma,delta,numpred] = cvshrink(Mdl,...
    'NumGamma',24,'NumDelta',24,'Verbose',1);
Done building cross-validated model.
Processing Gamma step 1 out of 25.
Processing Gamma step 2 out of 25.
Processing Gamma step 3 out of 25.
Processing Gamma step 4 out of 25.
Processing Gamma step 5 out of 25.
Processing Gamma step 6 out of 25.
Processing Gamma step 7 out of 25.
Processing Gamma step 8 out of 25.
Processing Gamma step 9 out of 25.
Processing Gamma step 10 out of 25.
Processing Gamma step 11 out of 25.
Processing Gamma step 12 out of 25.
Processing Gamma step 13 out of 25.
Processing Gamma step 14 out of 25.
Processing Gamma step 15 out of 25.
Processing Gamma step 16 out of 25.
Processing Gamma step 17 out of 25.
Processing Gamma step 18 out of 25.
Processing Gamma step 19 out of 25.
Processing Gamma step 20 out of 25.
Processing Gamma step 21 out of 25.
Processing Gamma step 22 out of 25.
Processing Gamma step 23 out of 25.
Processing Gamma step 24 out of 25.
Processing Gamma step 25 out of 25.

Examine the quality of the regularized classifiers.

Plot the number of predictors against the error.

plot(err,numpred,'k.')
xlabel('Error rate')
ylabel('Number of predictors')

Figure contains an axes object. The axes object with xlabel Error rate, ylabel Number of predictors contains 25 objects of type line. One or more of the lines displays its values using only markers

Examine the lower-left part of the plot more closely.

axis([0 .1 0 1000])

Figure contains an axes object. The axes object with xlabel Error rate, ylabel Number of predictors contains 25 objects of type line. One or more of the lines displays its values using only markers

There is a clear tradeoff between lower number of predictors and lower error.

Choose an optimal tradeoff between model size and accuracy.

Multiple pairs of Gamma and Delta values produce about the same minimal error. Display the indices of these pairs and their values.

First, find the minimal error value.

minerr = min(min(err))
minerr = 0.0139

Find the subscripts of err producing minimal error.

[p,q] = find(err < minerr + 1e-4);

Convert from subscripts to linear indices.

idx = sub2ind(size(delta),p,q);

Display the Gamma and Delta values.

[gamma(p) delta(idx)]
ans = 4×2

    0.7202    0.1145
    0.7602    0.1131
    0.8001    0.1128
    0.8001    0.1410

These points have as few as 29% of the total predictors with nonzero coefficients in the model.

numpred(idx)/ceil(numPred/3)*100
ans = 4×1

   39.8051
   38.9805
   36.8066
   28.7856

To further lower the number of predictors, you must accept larger error rates. For example, to choose the Gamma and Delta that give the lowest error rate with 200 or fewer predictors.

low200 = min(min(err(numpred <= 200)));
lownum = min(min(numpred(err == low200)));
[low200 lownum]
ans = 1×2

    0.0185  173.0000

You need 173 predictors to achieve an error rate of 0.0185, and this is the lowest error rate among those that have 200 predictors or fewer.

Display the Gamma and Delta that achieve this error/number of predictors.

[r,s] = find((err == low200) & (numpred == lownum));
[gamma(r); delta(r,s)]
ans = 2×1

    0.6403
    0.2399

Set the regularization parameters.

To set the classifier with these values of Gamma and Delta, use dot notation.

Mdl.Gamma = gamma(r);
Mdl.Delta = delta(r,s);

Heatmap plot

To compare the cvshrink calculation to that in Guo, Hastie, and Tibshirani [1], plot heatmaps of error and number of predictors against Gamma and the index of the Delta parameter. (The Delta parameter range depends on the value of the Gamma parameter. So to get a rectangular plot, use the Delta index, not the parameter itself.)

% Create the Delta index matrix
indx = repmat(1:size(delta,2),size(delta,1),1);

figure
subplot(1,2,1)
imagesc(err)
colorbar
colormap('jet')
title('Classification error')
xlabel('Delta index')
ylabel('Gamma index')

subplot(1,2,2)
imagesc(numpred)
colorbar
title('Number of predictors in the model')
xlabel('Delta index')
ylabel('Gamma index')

Figure contains 2 axes objects. Axes object 1 with title Classification error, xlabel Delta index, ylabel Gamma index contains an object of type image. Axes object 2 with title Number of predictors in the model, xlabel Delta index, ylabel Gamma index contains an object of type image.

You see the best classification error when Delta is small, but fewest predictors when Delta is large.

References

[1] Guo, Y., T. Hastie, and R. Tibshirani. "Regularized Discriminant Analysis and Its Application in Microarray." Biostatistics, Vol. 8, No. 1, pp. 86–100, 2007.

See Also

Functions

Objects

Related Topics