loss

Classification loss for neural network classifier

Since R2021a

Syntax

L = loss(Mdl,Tbl,ResponseVarName)

L = loss(Mdl,Tbl,Y)

L = loss(Mdl,X,Y)

L = loss(___,Name,Value)

Description

L = loss(Mdl,Tbl,ResponseVarName) returns the classification loss for the trained neural network classifier Mdl using the predictor data in table Tbl and the class labels in the ResponseVarName table variable.

L is returned as a scalar value that represents the classification error by default.

L = loss(Mdl,Tbl,Y) returns the classification loss for the classifier Mdl using the predictor data in table Tbl and the class labels in vector Y.

L = loss(Mdl,X,Y) returns the classification loss for the trained neural network classifier Mdl using the predictor data X and the corresponding class labels in Y.

L = loss(___,Name,Value) specifies options using one or more name-value arguments in addition to any of the input argument combinations in previous syntaxes. For example, you can specify that columns in the predictor data correspond to observations, specify the loss function, or supply observation weights.

Note

If the predictor data in X or Tbl contains any missing values and LossFun is not set to "classifcost", "classiferror", or "mincost", the loss function can return NaN. For more details, see loss can return NaN for predictor data with missing values.

Examples

Test Set Classification Error of Neural Network

Calculate the test set classification error of a neural network classifier.

Load the patients data set. Create a table from the data set. Each row corresponds to one patient, and each column corresponds to a diagnostic variable. Use the Smoker variable as the response variable, and the rest of the variables as predictors.

load patients
tbl = table(Diastolic,Systolic,Gender,Height,Weight,Age,Smoker);

Separate the data into a training set tblTrain and a test set tblTest by using a stratified holdout partition. The software reserves approximately 30% of the observations for the test data set and uses the rest of the observations for the training data set.

rng("default") % For reproducibility of the partition
c = cvpartition(tbl.Smoker,"Holdout",0.30);
trainingIndices = training(c);
testIndices = test(c);
tblTrain = tbl(trainingIndices,:);
tblTest = tbl(testIndices,:);

Train a neural network classifier using the training set. Specify the Smoker column of tblTrain as the response variable. Specify to standardize the numeric predictors.

Mdl = fitcnet(tblTrain,"Smoker", ...
    "Standardize",true);

Calculate the test set classification error. Classification error is the default loss type for neural network classifiers.

testError = loss(Mdl,tblTest,"Smoker")

testError = 
0.0671

testAccuracy = 1 - testError

testAccuracy = 
0.9329

The neural network model correctly classifies approximately 93% of the test set observations.

Select Features to Include in Neural Network Classifier

Perform feature selection by comparing test set classification margins, edges, errors, and predictions. Compare the test set metrics for a model trained using all the predictors to the test set metrics for a model trained using only a subset of the predictors.

Load the sample file fisheriris.csv, which contains iris data including sepal length, sepal width, petal length, petal width, and species type. Read the file into a table.

fishertable = readtable('fisheriris.csv');

Separate the data into a training set trainTbl and a test set testTbl by using a stratified holdout partition. The software reserves approximately 30% of the observations for the test data set and uses the rest of the observations for the training data set.

rng("default")
c = cvpartition(fishertable.Species,"Holdout",0.3);
trainTbl = fishertable(training(c),:);
testTbl = fishertable(test(c),:);

Train one neural network classifier using all the predictors in the training set, and train another classifier using all the predictors except PetalWidth. For both models, specify Species as the response variable, and standardize the predictors.

allMdl = fitcnet(trainTbl,"Species","Standardize",true);
subsetMdl = fitcnet(trainTbl,"Species ~ SepalLength + SepalWidth + PetalLength", ...
    "Standardize",true);

Calculate the test set classification margins for the two models. Because the test set includes only 45 observations, display the margins using bar graphs.

For each observation, the classification margin is the difference between the classification score for the true class and the maximal score for the false classes. Because neural network classifiers return classification scores that are posterior probabilities, margin values close to 1 indicate confident classifications and negative margin values indicate misclassifications.

tiledlayout(2,1)

% Top axes
ax1 = nexttile;
allMargins = margin(allMdl,testTbl);
bar(ax1,allMargins)
xlabel(ax1,"Observation")
ylabel(ax1,"Margin")
title(ax1,"All Predictors")

% Bottom axes
ax2 = nexttile;
subsetMargins = margin(subsetMdl,testTbl);
bar(ax2,subsetMargins)
xlabel(ax2,"Observation")
ylabel(ax2,"Margin")
title(ax2,"Subset of Predictors")

Compare the test set classification edge, or mean of the classification margins, of the two models.

allEdge = edge(allMdl,testTbl)

allEdge = 
0.8198

subsetEdge = edge(subsetMdl,testTbl)

subsetEdge = 
0.9556

Based on the test set classification margins and edges, the model trained on a subset of the predictors seems to outperform the model trained on all the predictors.

Compare the test set classification error of the two models.

allError = loss(allMdl,testTbl);
allAccuracy = 1-allError

allAccuracy = 
0.9111

subsetError = loss(subsetMdl,testTbl);
subsetAccuracy = 1-subsetError

subsetAccuracy = 
0.9778

Again, the model trained using only a subset of the predictors seems to perform better than the model trained using all the predictors.

Visualize the test set classification results using confusion matrices.

allLabels = predict(allMdl,testTbl);
figure
confusionchart(testTbl.Species,allLabels)
title("All Predictors")

subsetLabels = predict(subsetMdl,testTbl);
figure
confusionchart(testTbl.Species,subsetLabels)
title("Subset of Predictors")

The model trained using all the predictors misclassifies four of the test set observations. The model trained using a subset of the predictors misclassifies only one of the test set observations.

Given the test set performance of the two models, consider using the model trained using all the predictors except PetalWidth.

Input Arguments

Mdl — Trained neural network classifier

ClassificationNeuralNetwork model object | CompactClassificationNeuralNetwork model object

Trained neural network classifier, specified as a ClassificationNeuralNetwork model object or CompactClassificationNeuralNetwork model object returned by fitcnet or compact, respectively.

Tbl — Sample data

table

Sample data, specified as a table. Each row of Tbl corresponds to one observation, and each column corresponds to one predictor variable. Optionally, Tbl can contain an additional column for the response variable. Tbl must contain all of the predictors used to train Mdl. Multicolumn variables and cell arrays other than cell arrays of character vectors are not allowed.

• If Tbl contains the response variable used to train Mdl, then you do not need to specify ResponseVarName or Y.

• If you trained Mdl using sample data contained in a table, then the input data for loss must also be in a table.

• If you set 'Standardize',true in fitcnet when training Mdl, then the software standardizes the numeric columns of the predictor data using the corresponding means and standard deviations.

Data Types: table

ResponseVarName — Response variable name

name of variable in Tbl

Response variable name, specified as the name of a variable in Tbl. If Tbl contains the response variable used to train Mdl, then you do not need to specify ResponseVarName.

If you specify ResponseVarName, then you must specify it as a character vector or string scalar. For example, if the response variable is stored as Tbl.Y, then specify ResponseVarName as 'Y'. Otherwise, the software treats all columns of Tbl, including Tbl.Y, as predictors.

The response variable must be a categorical, character, or string array; a logical or numeric vector; or a cell array of character vectors. If the response variable is a character array, then each element must correspond to one row of the array.

Data Types: char | string

Y — Class labels

categorical array | character array | string array | logical vector | numeric vector | cell array of character vectors

Class labels, specified as a categorical, character, or string array; logical or numeric vector; or cell array of character vectors.

• The data type of Y must be the same as the data type of Mdl.ClassNames. (The software treats string arrays as cell arrays of character vectors.)

• The distinct classes in Y must be a subset of Mdl.ClassNames.

• If Y is a character array, then each element must correspond to one row of the array.

• The length of Y must be equal to the number of observations in X or Tbl.

Data Types: categorical | char | string | logical | single | double | cell

X — Predictor data

numeric matrix

Predictor data, specified as a numeric matrix. By default, loss assumes that each row of X corresponds to one observation, and each column corresponds to one predictor variable.

Note

If you orient your predictor matrix so that observations correspond to columns and specify 'ObservationsIn','columns', then you might experience a significant reduction in computation time.

The length of Y and the number of observations in X must be equal.

If you set 'Standardize',true in fitcnet when training Mdl, then the software standardizes the numeric columns of the predictor data using the corresponding means and standard deviations.

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: loss(Mdl,Tbl,"Response","LossFun","crossentropy") specifies to compute the cross-entropy loss for the model Mdl.

LossFun — Loss function

'mincost' (default) | 'binodeviance' | 'classifcost' | 'classiferror' | 'crossentropy' | 'exponential' | 'hinge' | 'logit' | 'quadratic' | function handle

Loss function, specified as a built-in loss function name or a function handle.

• This table lists the available loss functions. Specify one using its corresponding character vector or string scalar.

  Value	Description
  'binodeviance'	Binomial deviance
  'classifcost'	Observed misclassification cost
  'classiferror'	Misclassified rate in decimal
  'crossentropy'	Cross-entropy loss (for neural networks only)
  'exponential'	Exponential loss
  'hinge'	Hinge loss
  'logit'	Logistic loss
  'mincost'	Minimal expected misclassification cost (for classification scores that are posterior probabilities)
  'quadratic'	Quadratic loss

  For more details on loss functions, see Classification Loss.

• To specify a custom loss function, use function handle notation. The function must have this form:

  lossvalue = lossfun(C,S,W,Cost)

  • The output argument lossvalue is a scalar.

  • You specify the function name (lossfun).

  • C is an n-by-K logical matrix with rows indicating the class to which the corresponding observation belongs. n is the number of observations in Tbl or X, and K is the number of distinct classes (numel(Mdl.ClassNames)). The column order corresponds to the class order in Mdl.ClassNames. Create C by setting C(p,q) = 1, if observation p is in class q, for each row. Set all other elements of row p to 0.

  • S is an n-by-K numeric matrix of classification scores. The column order corresponds to the class order in Mdl.ClassNames. S is a matrix of classification scores, similar to the output of predict.

  • W is an n-by-1 numeric vector of observation weights.

  • Cost is a K-by-K numeric matrix of misclassification costs. For example, Cost = ones(K) – eye(K) specifies a cost of 0 for correct classification and 1 for misclassification.

Example: 'LossFun','crossentropy'

Data Types: char | string | function_handle

ObservationsIn — Predictor data observation dimension

'rows' (default) | 'columns'

Predictor data observation dimension, specified as 'rows' or 'columns'.

Note

If you orient your predictor matrix so that observations correspond to columns and specify 'ObservationsIn','columns', then you might experience a significant reduction in computation time. You cannot specify 'ObservationsIn','columns' for predictor data in a table.

Data Types: char | string

Weights — Observation weights

nonnegative numeric vector | name of variable in Tbl

Observation weights, specified as a nonnegative numeric vector or the name of a variable in Tbl. The software weights each observation in X or Tbl with the corresponding value in Weights. The length of Weights must equal the number of observations in X or Tbl.

If you specify the input data as a table Tbl, then Weights can be the name of a variable in Tbl that contains a numeric vector. In this case, you must specify Weights as a character vector or string scalar. For example, if the weights vector W is stored as Tbl.W, then specify it as 'W'.

By default, Weights is ones(n,1), where n is the number of observations in X or Tbl.

If you supply weights, then loss computes the weighted classification loss and normalizes weights to sum to the value of the prior probability in the respective class.

Data Types: single | double | char | string

More About

Classification Loss

Classification loss functions measure the predictive inaccuracy of classification models. When you compare the same type of loss among many models, a lower loss indicates a better predictive model.

Consider the following scenario.

• L is the weighted average classification loss.

• n is the sample size.

• For binary classification:

  • y_j is the observed class label. The software codes it as –1 or 1, indicating the negative or positive class (or the first or second class in the ClassNames property), respectively.

  • f(X_j) is the positive-class classification score for observation (row) j of the predictor data X.

  • m_j = y_jf(X_j) is the classification score for classifying observation j into the class corresponding to y_j. Positive values of m_j indicate correct classification and do not contribute much to the average loss. Negative values of m_j indicate incorrect classification and contribute significantly to the average loss.

• For algorithms that support multiclass classification (that is, K ≥ 3):

  • y_j^* is a vector of K – 1 zeros, with 1 in the position corresponding to the true, observed class y_j. For example, if the true class of the second observation is the third class and K = 4, then y₂^* = [0 0 1 0]′. The order of the classes corresponds to the order in the ClassNames property of the input model.

  • f(X_j) is the length K vector of class scores for observation j of the predictor data X. The order of the scores corresponds to the order of the classes in the ClassNames property of the input model.

  • m_j = y_j^*′f(X_j). Therefore, m_j is the scalar classification score that the model predicts for the true, observed class.

• The weight for observation j is w_j. The software normalizes the observation weights so that they sum to the corresponding prior class probability stored in the Prior property. Therefore,

  $${\displaystyle \sum _{j=1}^n{w}_j}=1.$$

Given this scenario, the following table describes the supported loss functions that you can specify by using the LossFun name-value argument.

Loss Function	Value of LossFun	Equation
Binomial deviance	"binodeviance"	$$L={\displaystyle \sum _{j=1}^n{w}_j\log \left\{1+\exp \left[-2m_j\right]\right\}}.$$
Observed misclassification cost	"classifcost"	$$L={\displaystyle \sum _{j=1}^n{w}_j}{c}_{y_j{\widehat{y}}_j},$$ where $${\widehat{y}}_j$$ is the class label corresponding to the class with the maximal score, and $$c_{y_j{\widehat{y}}_j}$$ is the user-specified cost of classifying an observation into class $${\widehat{y}}_j$$ when its true class is yj.
Misclassified rate in decimal	"classiferror"	$$L={\displaystyle \sum _{j=1}^n{w}_j}I\left\{{\widehat{y}}_j\ne y_j\right\},$$ where I{·} is the indicator function.
Cross-entropy loss	"crossentropy"	"crossentropy" is appropriate only for neural network models. The weighted cross-entropy loss is $$L=-{\displaystyle \sum _{j=1}^n\frac{{\tilde{w}}_j\log (m_j)}{Kn}},$$ where the weights $${\tilde{w}}_j$$ are normalized to sum to n instead of 1.
Exponential loss	"exponential"	$$L={\displaystyle \sum _{j=1}^n{w}_j\exp \left(-m_j\right)}.$$
Hinge loss	"hinge"	$$L={\displaystyle \sum }_{j=1}^n{w}_j\max \left\{0,1-m_j\right\}.$$
Logit loss	"logit"	$$L={\displaystyle \sum _{j=1}^n{w}_j\log \left(1+\exp \left(-m_j\right)\right)}.$$
Minimal expected misclassification cost	"mincost"	"mincost" is appropriate only if classification scores are posterior probabilities. The software computes the weighted minimal expected classification cost using this procedure for observations j = 1,...,n. Estimate the expected misclassification cost of classifying the observation Xj into the class k: $${\gamma }_{jk}={\left(f{\left(X_j\right)}^{\prime }C\right)}_k.$$ f(Xj) is the column vector of class posterior probabilities for the observation Xj. C is the cost matrix stored in the Cost property of the model. For observation j, predict the class label corresponding to the minimal expected misclassification cost: $${\widehat{y}}_j=\underset{k=1,\mathrm{...},K}{\text{argmin}}{\gamma }_{jk}.$$ Using C, identify the cost incurred (cj) for making the prediction. The weighted average of the minimal expected misclassification cost loss is $$L={\displaystyle \sum _{j=1}^n{w}_j{c}_j}.$$
Quadratic loss	"quadratic"	$$L={\displaystyle \sum _{j=1}^n{w}_j{\left(1-m_j\right)}^2}.$$

If you use the default cost matrix (whose element value is 0 for correct classification and 1 for incorrect classification), then the loss values for "classifcost", "classiferror", and "mincost" are identical. For a model with a nondefault cost matrix, the "classifcost" loss is equivalent to the "mincost" loss most of the time. These losses can be different if prediction into the class with maximal posterior probability is different from prediction into the class with minimal expected cost. Note that "mincost" is appropriate only if classification scores are posterior probabilities.

This figure compares the loss functions (except "classifcost", "crossentropy", and "mincost") over the score m for one observation. Some functions are normalized to pass through the point (0,1).

Extended Capabilities

GPU Arrays

Accelerate code by running on a graphics processing unit (GPU) using Parallel Computing Toolbox™. (since R2024b)

This function fully supports GPU arrays. For more information, see Run MATLAB Functions on a GPU (Parallel Computing Toolbox).

Version History

Introduced in R2021a

R2024b: Specify GPU arrays (requires Parallel Computing Toolbox)

loss fully supports GPU arrays.

R2022a: Default LossFun value has changed

Starting in R2022a, the loss function uses the "mincost" option (minimal expected misclassification cost) as the default value for the LossFun name-value argument. The "mincost" option is appropriate when classification scores are posterior probabilities. In previous releases, the default value was "classiferror".

You do not need to make any changes to your code for the default cost matrix (whose element value is 0 for correct classification and 1 for incorrect classification).

R2022a: loss can return NaN for predictor data with missing values

The loss function no longer omits an observation with a NaN score when computing the weighted average classification loss. Therefore, loss can now return NaN when the predictor data X or the predictor variables in Tbl contain any missing values, and the name-value argument LossFun is not specified as "classifcost", "classiferror", or "mincost". In most cases, if the test set observations do not contain missing predictors, the loss function does not return NaN.

This change improves the automatic selection of a classification model when you use fitcauto. Before this change, the software might select a model (expected to best classify new data) with few non-NaN predictors.

If loss in your code returns NaN, you can update your code to avoid this result by doing one of the following:

• Remove or replace the missing values by using rmmissing or fillmissing, respectively.

• Specify the name-value argument LossFun as "classifcost", "classiferror", or "mincost".

The following table shows the classification models for which the loss object function might return NaN. For more details, see the Compatibility Considerations for each loss function.

Model Type	Full or Compact Model Object	loss Object Function
Discriminant analysis classification model	ClassificationDiscriminant, CompactClassificationDiscriminant	loss
Ensemble of learners for classification	ClassificationEnsemble, CompactClassificationEnsemble	loss
Gaussian kernel classification model	ClassificationKernel	loss
k-nearest neighbor classification model	ClassificationKNN	loss
Linear classification model	ClassificationLinear	loss
Neural network classification model	ClassificationNeuralNetwork, CompactClassificationNeuralNetwork	loss
Support vector machine (SVM) classification model	ClassificationSVM, CompactClassificationSVM	loss