cvshrink

Cross-validate regularization of linear discriminant

Syntax

err = cvshrink(obj) [err,gamma] = cvshrink(obj) [err,gamma,delta] = cvshrink(obj) [err,gamma,delta,numpred] = cvshrink(obj) [err,...] = cvshrink(obj,Name,Value)

Description

err = cvshrink(obj) returns a vector of cross-validated classification error values for differing values of the regularization parameter Gamma.

[err,gamma] = cvshrink(obj) also returns the vector of Gamma values.

[err,gamma,delta] = cvshrink(obj) also returns the vector of Delta values.

[err,gamma,delta,numpred] = cvshrink(obj) returns the vector of number of nonzero predictors for each setting of the parameters Gamma and Delta.

[err,...] = cvshrink(obj,Name,Value) cross validates with additional options specified by one or more Name,Value pair arguments.

Input Arguments

obj

Discriminant analysis classifier, produced using fitcdiscr.

Name-Value Pair Arguments

Specify optional comma-separated pairs of Name,Value arguments. Name is the argument name and Value is the corresponding value. Name must appear inside quotes. You can specify several name and value pair arguments in any order as Name1,Value1,...,NameN,ValueN.

`'delta'`	Scalar `delta` — `cvshrink` uses this value of `delta` with every value of `gamma` for regularization. Row vector `delta` — For each `i` and `j`, `cvshrink` uses `delta(j)` with `gamma(i)` for regularization. Matrix `delta` — The number of rows of `delta` must equal the number of elements in `gamma`. For each `i` and `j`, `cvshrink` uses `delta(i,j)` with `gamma(i)` for regularization. Default: `0`
`'gamma'`	Vector of Gamma values for cross-validation. Default: `0:0.1:1`
`'NumDelta'`	Number of Delta intervals for cross-validation. For every value of Gamma, `cvshrink` cross-validates the discriminant using `NumDelta + 1` values of Delta, uniformly spaced from zero to the maximal Delta at which all predictors are eliminated for this value of Gamma. If you set `delta`, `cvshrink` ignores `NumDelta`. Default: `0`
`'NumGamma'`	Number of Gamma intervals for cross-validation. `cvshrink` cross-validates the discriminant using `NumGamma + 1` values of Gamma, uniformly spaced from `MinGamma` to `1`. If you set `gamma`, `cvshrink` ignores `NumGamma`. Default: `10`
`'verbose'`	Verbosity level, an integer from `0` to `2`. Higher values give more progress messages. Default: `0`

Output Arguments

`err`	Numeric vector or matrix of errors. `err` is the misclassification error rate, meaning the average fraction of misclassified data over all folds. If `delta` is a scalar (default), `err(i)` is the misclassification error rate for `obj` regularized with `gamma(i)`. If `delta` is a vector, `err(i,j)` is the misclassification error rate for `obj` regularized with `gamma(i)` and `delta(j)`. If `delta` is a matrix, `err(i,j)` is the misclassification error rate for `obj` regularized with `gamma(i)` and `delta(i,j)`.
`gamma`	Vector of Gamma values used for regularization. See Gamma and Delta.
`delta`	Vector or matrix of Delta values used for regularization. See Gamma and Delta. If you give a scalar for the `delta` name-value pair, the output `delta` is a row vector the same size as `gamma`, with entries equal to the input scalar. If you give a row vector for the `delta` name-value pair, the output `delta` is a matrix with the same number of columns as the row vector, and with the number of rows equal to the number of elements of `gamma`. The output `delta(i,j)` is equal to the input `delta(j)`. If you give a matrix for the `delta` name-value pair, the output `delta` is the same as the input matrix. The number of rows of `delta` must equal the number of elements in `gamma`.
`numpred`	Numeric vector or matrix containing the number of predictors in the model at various regularizations. `numpred` has the same size as `err`. If `delta` is a scalar (default), `numpred(i)` is the number of predictors for `obj` regularized with `gamma(i)` and `delta`. If `delta` is a vector, `numpred(i,j)` is the number of predictors for `obj` regularized with `gamma(i)` and `delta(j)`. If `delta` is a matrix, `numpred(i,j)` is the number of predictors for `obj` regularized with `gamma(i)` and `delta(i,j)`.

Examples

expand all

Regularize Data with Many Predictors

Open Live Script

Regularize a discriminant analysis classifier, and view the tradeoff between the number of predictors in the model and the classification accuracy.

Create a linear discriminant analysis classifier for the ovariancancer data. Set the SaveMemory and FillCoeffs options to keep the resulting model reasonably small.

load ovariancancer
obj = fitcdiscr(obs,grp,...
    'SaveMemory','on','FillCoeffs','off');

Use 10 levels of Gamma and 10 levels of Delta to search for good parameters. This search is time-consuming. Set Verbose to 1 to view the progress.

rng('default') % for reproducibility
[err,gamma,delta,numpred] = cvshrink(obj,...
    'NumGamma',9,'NumDelta',9,'Verbose',1);

Done building cross-validated model.
Processing Gamma step 1 out of 10.
Processing Gamma step 2 out of 10.
Processing Gamma step 3 out of 10.
Processing Gamma step 4 out of 10.
Processing Gamma step 5 out of 10.
Processing Gamma step 6 out of 10.
Processing Gamma step 7 out of 10.
Processing Gamma step 8 out of 10.
Processing Gamma step 9 out of 10.
Processing Gamma step 10 out of 10.

Plot the classification error rate against the number of predictors.

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

More About

expand all

Gamma and Delta

Regularization is the process of finding a small set of predictors that yield an effective predictive model. For linear discriminant analysis, there are two parameters, γ and δ, that control regularization as follows. cvshrink helps you select appropriate values of the parameters.

Let Σ represent the covariance matrix of the data X, and let $\hat{X}$ be the centered data (the data X minus the mean by class). Define

$D = diag ({\hat{X}}^{T} * \hat{X}) .$

The regularized covariance matrix $\tilde{Σ}$ is

$\tilde{Σ} = (1 - γ) Σ + γ D .$

Whenever γ ≥ MinGamma, $\tilde{Σ}$ is nonsingular.

Let μ_k be the mean vector for those elements of X in class k, and let μ₀ be the global mean vector (the mean of the rows of X). Let C be the correlation matrix of the data X, and let $\tilde{C}$ be the regularized correlation matrix:

$\tilde{C} = (1 - γ) C + γ I,$

where I is the identity matrix.

The linear term in the regularized discriminant analysis classifier for a data point x is

${(x - μ_{0})}^{T} {\tilde{Σ}}^{- 1} (μ_{k} - μ_{0}) = [{(x - μ_{0})}^{T} D^{- 1 / 2}] [{\tilde{C}}^{- 1} D^{- 1 / 2} (μ_{k} - μ_{0})] .$

The parameter δ enters into this equation as a threshold on the final term in square brackets. Each component of the vector $[{\tilde{C}}^{- 1} D^{- 1 / 2} (μ_{k} - μ_{0})]$ is set to zero if it is smaller in magnitude than the threshold δ. Therefore, for class k, if component j is thresholded to zero, component j of x does not enter into the evaluation of the posterior probability.

The DeltaPredictor property is a vector related to this threshold. When δ ≥ DeltaPredictor(i), all classes k have

$| {\tilde{C}}^{- 1} D^{- 1 / 2} (μ_{k} - μ_{0}) | \leq δ .$

Therefore, when δ ≥ DeltaPredictor(i), the regularized classifier does not use predictor i.

Tips

Examine the err and numpred outputs to see the tradeoff between cross-validated error and number of predictors. When you find a satisfactory point, set the corresponding gamma and delta properties in the model using dot notation. For example, if (i,j) is the location of the satisfactory point, set
```
obj.Gamma = gamma(i);
obj.Delta = delta(i,j);
```

Documentation