kfoldfun

Cross-validate function for classification

Syntax

vals = kfoldfun(CVMdl,fun)

Description

vals = kfoldfun(CVMdl,fun) cross-validates the function fun by applying fun to the data stored in the cross-validated model CVMdl. You must pass fun as a function handle.

Examples

Estimate Classification Loss Using Custom Loss Function

Train a classification tree classifier, and then cross-validate it using a custom k-fold loss function.

Load Fisher’s iris data set.

load fisheriris

Train a classification tree classifier.

Mdl = fitctree(meas,species);

Mdl is a ClassificationTree model.

Cross-validate Mdl using the default 10-fold cross-validation. Compute the classification error (proportion of misclassified observations) for the validation-fold observations.

rng(1); % For reproducibility
CVMdl = crossval(Mdl);
L = kfoldLoss(CVMdl,'LossFun','classiferror')

L = 
0.0467

Examine the result when the cost of misclassifying a flower as versicolor is 10, and the cost of any other misclassification is 1. Create the custom function noversicolor (shown at the end of this example). This function attributes a cost of 10 for misclassifying a flower as versicolor, and a cost of 1 for any other misclassification.

Compute the mean misclassification error with the noversicolor cost.

mean(kfoldfun(CVMdl,@noversicolor))

ans = 
0.2267

This code creates the function noversicolor.

function averageCost = noversicolor(CMP,~,~,~,Xtest,Ytest,~)
% noversicolor Example custom cross-validation function
%    Attributes a cost of 10 for misclassifying versicolor irises, and 1 for
%    the other irises.  This example function requires the fisheriris data
%    set.
Ypredict = predict(CMP,Xtest);
misclassified = not(strcmp(Ypredict,Ytest)); % Different result
classifiedAsVersicolor = strcmp(Ypredict,'versicolor'); % Index of bad decisions
cost = sum(misclassified) + ...
    9*sum(misclassified & classifiedAsVersicolor); % Total differences
averageCost = cost/numel(Ytest); % Average error
end

Input Arguments

CVMdl — Cross-validated model

ClassificationPartitionedModel object | ClassificationPartitionedEnsemble object | ClassificationPartitionedGAM object

Cross-validated model, specified as a ClassificationPartitionedModel object, ClassificationPartitionedEnsemble object, or ClassificationPartitionedGAM object.

fun — Cross-validated function

function handle

Cross-validated function, specified as a function handle. fun has the syntax:

testvals = fun(CMP,Xtrain,Ytrain,Wtrain,Xtest,Ytest,Wtest)

• CMP is a compact model stored in one element of the CVMdl.Trained property.

• Xtrain is the training matrix of predictor values.

• Ytrain is the training array of response values.

• Wtrain are the training weights for observations.

• Xtest and Ytest are the test data, with associated weights Wtest.

• The returned value testvals must have the same size across all folds.

Data Types: function_handle

Output Arguments

vals — Cross-validation results

numeric matrix

Cross-validation results, returned as a numeric matrix. vals contains the arrays of testvals output, concatenated vertically over all folds. For example, if testvals from every fold is a numeric vector of length N, kfoldfun returns a KFold-by-N numeric matrix with one row per fold.

Data Types: double

Extended Capabilities

GPU Arrays

Accelerate code by running on a graphics processing unit (GPU) using Parallel Computing Toolbox™.

Usage notes and limitations:

• fun must accept gpuArray inputs.

• This function fully supports GPU arrays for the following cross-validated model objects:

  • Ensemble classifier trained with fitcensemble

  • k-nearest neighbor classifier trained with fitcknn

  • Support vector machine classifier trained with fitcsvm

  • Binary decision tree for multiclass classification trained with fitctree

  • Neural network for classification trained with fitcnet

For more information, see Run MATLAB Functions on a GPU (Parallel Computing Toolbox).

Version History

Introduced in R2011a

R2024b: Specify GPU arrays for neural network models (requires Parallel Computing Toolbox)

kfoldfun fully supports GPU arrays for ClassificationPartitionedModel models trained using fitcnet.