rocmetrics

Receiver operating characteristic (ROC) curve and performance metrics for binary and multiclass classifiers

Since R2022a

Description

Create a rocmetrics object to evaluate the performance of a classification model using receiver operating characteristic (ROC) curves or other performance metrics. rocmetrics supports both binary and multiclass problems.

For each class, rocmetrics computes performance metrics for a one-versus-all ROC curve. You can compute metrics for an average ROC curve by using the average function. After computing metrics for ROC curves, you can plot them by using the plot function.

By default, rocmetrics computes the false positive rates (FPR) and the true positive rates (TPR) to obtain a ROC curve and the area under the ROC curve (AUC). You can compute additional metrics by specifying the AdditionalMetrics name-value argument when you create an object or by calling the addMetrics function after you create an object. A rocmetrics object stores the computed metrics and AUC values in the Metrics and AUC properties, respectively.

rocmetrics computes pointwise confidence intervals for the performance metrics when you set the NumBootstraps value to a positive integer or when you specify cross-validated data for the true class labels (Labels), classification scores (Scores), and observation weights (Weights). For details, see Pointwise Confidence Intervals.

Creation

Syntax

rocObj = rocmetrics(Labels,Scores,ClassNames)

rocObj = rocmetrics(Labels,Scores,ClassNames,Name=Value)

Description

rocObj = rocmetrics(Labels,Scores,ClassNames) creates a rocmetrics object using the true class labels in Labels and the classification scores in Scores. Specify Labels as a vector of length n, and specify Scores as a matrix of size n-by-K, where n is the number of observations, and K is the number of classes. ClassNames specifies the column order in Scores.

The Metrics and AUC properties contain the performance metrics and AUC value for each class for which you specify Scores and ClassNames.

If you specify cross-validated data in Labels and Scores as cell arrays, then rocmetrics computes confidence intervals for the performance metrics.

rocObj = rocmetrics(Labels,Scores,ClassNames,Name=Value) specifies additional options using one or more name-value arguments. For example, NumBootstraps=100 draws 100 bootstrap samples to compute confidence intervals for the performance metrics.

Input Arguments

Labels — True class labels

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

True class labels, specified as a numeric vector, logical vector, categorical vector, character array, string array, or cell array of character vectors. You can also specify Labels as a cell array of one of these types for cross-validated data.

• For data that is not cross-validated, the length of Labels and the number of rows in Scores must be equal.

• For cross-validated data, you must specify Labels, Scores, and Weights as cell arrays with the same number of elements. rocmetrics treats an element in the cell arrays as data from one cross-validation fold and computes pointwise confidence intervals for the performance metrics. The length of Labels{i} and the number of rows in Scores{i} must be equal.

Each row of Labels or Labels{i} represents the true label of one observation.

This argument sets the Labels property.

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

Scores — Classification scores

numeric matrix | cell array of numeric matrices

Classification scores, specified as a numeric matrix or a cell array of numeric matrices.

Each row of the matrix in Scores contains the classification scores of one observation for all classes specified in ClassNames. The column order of Scores must match the class order in ClassNames.

• For a matrix input, Score(j,k) is the classification score of observation j for class ClassNames(k). You can specify Scores by using the second output argument of the predict function of a classification model object for both binary classification and multiclass classification. For example, predict of ClassificationTree returns classification scores as an n-by-K matrix, where n is the number of observations and K is the number classes. Pass the output to rocmetrics.

  The number of rows in Scores and the length of Labels must be equal. rocmetrics adjusts scores for each class relative to the scores for the rest of the classes. For details, see Adjusted Scores for Multiclass Classification Problem.

• For a vector input, Score(j) is the classification score of observation j for the class specified in ClassNames.

  • ClassNames must contain only one class.

  • Prior must be a two-element vector with Prior(1) representing the prior probability for the specified class.

  • Cost must be a 2-by-2 matrix containing [Cost(P|P),Cost(N|P);Cost(P|N),Cost(N|N)], where P is a positive class (the class for which you specify classification scores), and N is a negative class.

  • The length of Scores and the length of Labels must be equal.

  If you want to display the model operating point when you plot the ROC curve using the plot function, the values in Score(j) must be the posterior probability. This restriction applies only to a vector input.

• For cross-validated data, you must specify Labels, Scores, and Weights as cell arrays with the same number of elements. rocmetrics treats an element in the cell arrays as data from one cross-validation fold and computes pointwise confidence intervals for the performance metrics. Score{i}(j,k) is the classification score of observation j in element i for class ClassNames(k). The number of rows in Scores{i} and the length of Labels{i} must be equal.

For more information, see Classification Score Input for rocmetrics.

This argument sets the Scores property.

Data Types: single | double | cell

ClassNames — Class names

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

Class names, specified as a numeric vector, logical vector, categorical vector, character array, string array, or cell array of character vectors. ClassNames must have the same data type as the true labels in Labels. The values in ClassNames must appear in Labels.

• If you specify classification scores for only one class in Scores, ClassNames specifies only the name of this class.

• Otherwise, ClassNames specifies the order of the classes in Scores, Cost, and Prior.

This argument sets the ClassNames property.

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

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.

Example: FixedMetric="FalsePositiveRate",FixedMetricValues=0:0.01:1 holds the FPR values fixed at 0:0.01:1.

Performance Metrics

AdditionalMetrics — Additional model performance metrics

[] (default) | character vector | string array | function handle | cell array

Additional model performance metrics to compute, specified as a character vector or string scalar of the built-in metric name, string array of names, function handle (@metricName), or cell array of names or function handles. A rocmetrics object always computes the false positive rates (FPR) and the true positive rates (TPR) to obtain a ROC curve. Therefore, you do not have to specify to compute FPR and TPR.

• Built-in metrics — Specify one of the following built-in metric names by using a character vector or string scalar. You can specify more than one by using a string array.

  Name	Description
  "TruePositives" or "tp"	Number of true positives (TP)
  "FalseNegatives" or "fn"	Number of false negatives (FN)
  "FalsePositives" or "fp"	Number of false positives (FP)
  "TrueNegatives" or "tn"	Number of true negatives (TN)
  "SumOfTrueAndFalsePositives" or "tp+fp"	Sum of TP and FP
  "RateOfPositivePredictions" or "rpp"	Rate of positive predictions (RPP), (TP+FP)/(TP+FN+FP+TN)
  "RateOfNegativePredictions" or "rnp"	Rate of negative predictions (RNP), (TN+FN)/(TP+FN+FP+TN)
  "Accuracy" or "accu"	Accuracy, (TP+TN)/(TP+FN+FP+TN)
  "FalseNegativeRate", "fnr", or "miss"	False negative rate (FNR), or miss rate, FN/(TP+FN)
  "TrueNegativeRate", "tnr", or "spec"	True negative rate (TNR), or specificity, TN/(TN+FP)
  "PositivePredictiveValue", "ppv", or "prec"	Positive predictive value (PPV), or precision, TP/(TP+FP)
  "NegativePredictiveValue" or "npv"	Negative predictive value (NPV), TN/(TN+FN)
  "ExpectedCost" or "ecost"	Expected cost, (TP*cost(P|P)+FN*cost(N|P)+FP*cost(P|N)+TN*cost(N|N))/(TP+FN+FP+TN), where cost is a 2-by-2 misclassification cost matrix containing [0,cost(N|P);cost(P|N),0]. cost(N|P) is the cost of misclassifying a positive class (P) as a negative class (N), and cost(P|N) is the cost of misclassifying a negative class as a positive class. The software converts the K-by-K matrix specified by the Cost name-value argument of rocmetrics to a 2-by-2 matrix for each one-versus-all binary problem. For details, see Misclassification Cost Matrix.

  The software computes the scale vector using the prior class probabilities (Prior) and the number of classes in Labels, and then scales the performance metrics according to this scale vector. For details, see Performance Metrics.

• Custom metric — Specify a custom metric by using a function handle. A custom function that returns a performance metric must have this form:

  metric = customMetric(C,scale,cost)

  • The output argument metric is a scalar value.

  • A custom metric is a function of the confusion matrix (C), scale vector (scale), and cost matrix (cost). The software finds these input values for each one-versus-all binary problem. For details, see Performance Metrics.

    • C is a 2-by-2 confusion matrix consisting of [TP,FN;FP,TN].

    • scale is a 2-by-1 scale vector.

    • cost is a 2-by-2 misclassification cost matrix.

  The software does not support cross-validation for a custom metric. Instead, you can specify to use bootstrap when you create a rocmetrics object.

Note that the positive predictive value (PPV) is NaN for the reject-all threshold for which TP = FP = 0, and the negative predictive value (NPV) is NaN for the accept-all threshold for which TN = FN = 0. For more details, see Thresholds, Fixed Metric, and Fixed Metric Values.

Example: AdditionalMetrics=["Accuracy","PositivePredictiveValue"]

Example: AdditionalMetrics={"Accuracy",@m1,@m2} specifies the accuracy metric and the custom metrics m1 and m2 as additional metrics. rocmetrics stores the custom metric values as variables named CustomMetric1 and CustomMetric2 in the Metrics property.

Data Types: char | string | cell | function_handle

FixedMetric — Fixed metric

"Thresholds" (default) | "FalsePositiveRate" | "TruePositiveRate" | metric specified by AdditionalMetrics

Fixed metric, specified as "Thresholds", "FalsePositiveRate" (or "fpr"), "TruePositiveRate" (or "tpr"), or a metric specified by the AdditionalMetrics name-value argument. To hold a custom metric fixed, specify FixedMetric as "CustomMetricN", where N is the number that refers to the custom metric. For example, specify "CustomMetric1" to use the first custom metric specified by AdditionalMetrics as the fixed metric.

rocmetrics finds the ROC curves and other metric values that correspond to the fixed values (FixedMetricValues) of the fixed metric (FixedMetric), and stores the values in the Metrics property as a table. For more details, see Thresholds, Fixed Metric, and Fixed Metric Values.

If rocmetrics computes confidence intervals, it uses one of two methods for the computation, depending on the FixedMetric value:

• If FixedMetric is "Thresholds" (default), rocmetrics uses threshold averaging.

• If FixedMetric is a nondefault value, rocmetrics uses vertical averaging.

For details, see Pointwise Confidence Intervals.

Example: FixedMetric="TruePositiveRate"

Data Types: char | string

FixedMetricValues — Values for fixed metric

"all" (default) | numeric vector

Values for the fixed metric (FixedMetric), specified as "all" or a numeric vector.

rocmetrics finds the ROC curves and other metric values that correspond to the fixed values (FixedMetricValues) of the fixed metric (FixedMetric), and stores the values in the Metrics property as a table.

The default FixedMetric value is "Thresholds", and the default FixedMetricValues value is "all". For each class, rocmetrics uses all distinct adjusted score values as threshold values and computes the performance metrics using the threshold values. Depending on the UseNearestNeighbor setting, rocmetrics uses the exact threshold values corresponding to the fixed values or the nearest threshold values. For more details, see Thresholds, Fixed Metric, and Fixed Metric Values.

If rocmetrics computes confidence intervals, it holds FixedMetric fixed at FixedMetricValues.

• FixedMetric value is "Thresholds", and FixedMetricValues is "all" — rocmetrics computes confidence intervals at the values corresponding to all distinct threshold values.

• FixedMetric value is a performance metric, and FixedMetricValues is "all" — rocmetrics finds the metric values corresponding to all distinct threshold values, and computes confidence intervals at the values corresponding to the metric values.

For details, see Pointwise Confidence Intervals.

Example: FixedMetricValues=0:0.01:1

Data Types: single | double

NaNFlag — NaN condition

"omitnan" (default) | "includenan"

NaN condition, specified as "omitnan" or "includenan".

• "omitnan" — rocmetrics ignores all NaN score values in the input Scores and the corresponding values in Labels and Weights.

• "includenan" — rocmetrics uses the NaN score values in the input Scores for the calculation. The function adds the observations with NaN scores to false classification counts in the respective class. That is, the function counts observations with NaN scores from the positive class as false negative (FN), and counts observations with NaN scores from the negative class as false positive (FP).

For more details, see NaN Score Values.

Example: NaNFlag="includenan"

Data Types: char | string

UseNearestNeighbor — Indicator to use nearest metric values

false or 0 | true or 1

Indicator to use the nearest metric values, specified as logical 0 (false) or 1 (true).

• logical 0 (false) — rocmetrics uses the exact threshold values corresponding to the specified fixed metric values in FixedMetricValues for FixedMetric.

• logical 1 (true) — Among the adjusted input scores, rocmetrics finds a value that is the nearest to the threshold value corresponding to each specified fixed metric value.

For more details, see Thresholds, Fixed Metric, and Fixed Metric Values.

The UseNearestNeighbor value must be false if rocmetrics computes confidence intervals. Otherwise, the default value is true.

Example: UseNearestNeighbor=false

Data Types: logical

Options for Classification Model

ApplyCostToScores — Flag to apply misclassification costs to scores

false or 0 (default) | true or 1

Since R2024a

Flag to apply misclassification costs to scores for appropriate models, specified as a numeric or logical 0 (false) or 1 (true). Set ApplyCostToScores to true only when you specify scores for a k-nearest neighbor (KNN), discriminant analysis, or naive Bayes model with nondefault misclassification costs. These models use expected classification costs rather than scores to predict labels.

If you specify ApplyCostToScores as true, the software changes the scores to S*(-C), where the scores S are specified by the Scores argument, and the misclassification cost matrix C is specified by the Cost name-value argument. The rocmetrics object stores the transformed scores in the Scores property.

If you specify ApplyCostToScores as false, the software stores the untransformed scores in the Scores property of the rocmetrics object.

Example: ApplyCostToScores=true

Data Types: single | double | logical

Cost — Misclassification cost

square matrix

Misclassification cost, specified as a K-by-K square matrix C, where K is the number of unique classes in Labels. C(i,j) is the cost of classifying a point into class j if its true class is i (that is, the rows correspond to the true class and the columns correspond to the predicted class). ClassNames specifies the order of the classes.

rocmetrics converts the K-by-K matrix to a 2-by-2 matrix for each one-versus-all binary problem. For details, see Misclassification Cost Matrix.

If you specify classification scores for only one class in Scores, the Cost value must be a 2-by-2 matrix containing [0,cost(N|P);cost(P|N),0], where P is a positive class (the class for which you specify classification scores), and N is a negative class. cost(N|P) is the cost of misclassifying a positive class as a negative class, and cost(P|N) is the cost of misclassifying a negative class as a positive class.

The default value is C(i,j)=1 if i~=j, and C(i,j)=0 if i=j. The diagonal entries of a cost matrix must be zero.

This argument sets the Cost property.

Note

If you specify a misclassification cost matrix when you use scores for a KNN, discriminant analysis, or naive Bayes model, set ApplyCostToScores to true. These models use expected classification costs rather than scores to predict labels. (since R2024a)

Example: Cost=[0 2;1 0]

Data Types: single | double

Prior — Prior class probabilities

"empirical" (default) | "uniform" | vector of scalar values

Prior class probabilities, specified as one of the following:

• "empirical" determines class probabilities from class frequencies in the true class labels Labels. If you pass observation weights (Weights), rocmetrics also uses the weights to compute the class probabilities.

• "uniform" sets all class probabilities to be equal.

• Vector of scalar values, with one scalar value for each class. ClassNames specifies the order of the classes.

  If you specify classification scores for only one class in Scores, the Prior value must be a two-element vector with Prior(1) representing the prior probability for the specified class.

This argument sets the Prior property.

Example: Prior="uniform"

Data Types: single | double | char | string

Weights — Observation weights

numeric vector of positive values | cell array containing numeric vectors of positive values

Observation weights, specified as a numeric vector of positive values or a cell array containing numeric vectors of positive values.

• For data that is not cross-validated, specify Weights as a numeric vector that has the same length as Labels.

• For cross-validated data, you must specify Labels, Scores, and Weights as cell arrays with the same number of elements. rocmetrics treats an element in the cell arrays as data from one cross-validation fold and computes pointwise confidence intervals for the performance metrics. The length of Weights{i} and the length of Labels{i} must be equal.

rocmetrics weighs the observations in Labels and Scores with the corresponding values in Weights. If you set the NumBootstraps value to a positive integer, rocmetrics draws samples with replacement, using the weights as multinomial sampling probabilities.

By default, Weights is a vector of ones or a cell array containing vectors of ones.

This argument sets the Weights property.

Data Types: single | double | cell

Options for Confidence Intervals

Alpha — Significance level

0.05 (default) | scalar in the range (0,1)

Significance level for the pointwise confidence intervals, specified as a scalar in the range (0,1).

If you specify Alpha as α, then rocmetrics computes 100×(1 – α)% pointwise confidence intervals for the performance metrics.

This argument is related to computing confidence intervals. Therefore, it is valid only when you specify cross-validated data for Labels, Scores, and Weights, or when you set the NumBootstraps value to a positive integer.

Example: Alpha=0.01 specifies 99% confidence intervals.

Data Types: single | double

BootstrapOptions — Bootstrap options for parallel computation

statset("rocmetrics") (default) | structure

Bootstrap options for parallel computation, specified as a structure.

You can specify options for computing bootstrap iterations in parallel and setting random numbers during the bootstrap sampling. Create the BootstrapOptions structure with statset. This table lists the option fields and their values.

Field Name	Field Value	Default
UseParallel	Set this value to true to compute bootstrap iterations in parallel.	false
UseSubstreams	Set this value to true to run computations in parallel in a reproducible manner. To compute reproducibly, set Streams to a type that allows substreams: "mlfg6331_64" or "mrg32k3a".	false
Streams	Specify this value as a RandStream object or cell array of such objects. Use a single object except when the UseParallel value is true and the UseSubstreams value is false. In that case, use a cell array that has the same size as the parallel pool.	If you do not specify Streams, then rocmetrics uses the default stream or streams.

This argument is valid only when you specify NumBootstraps as a positive integer to compute confidence intervals using bootstrapping.

Parallel computation requires Parallel Computing Toolbox™.

Example: BootstrapOptions=statset(UseParallel=true)

Data Types: struct

BootstrapType — Bootstrap confidence interval type

"bca" (default) | "corrected percentile" | "normal" | "percentile" | "student"

Bootstrap confidence interval type, specified as one of the values in this table.

Value	Description
"bca"	Bias corrected and accelerated percentile method [8][9]. This method Involves a z0 factor computed using the proportion of bootstrap values that are less than the original sample value. To produce reasonable results when the sample is lumpy, the software computes z0 by including half of the bootstrap values that are the same as the original sample value.
"corrected percentile" or "cper"	Bias corrected percentile method [10]
"normal" or "norm"	Normal approximated interval with bootstrapped bias and standard error [11]
"percentile" or "per"	Basic percentile method
"student" or "stud"	Studentized confidence interval [8]

This argument is valid only when you specify NumBootstraps as a positive integer to compute confidence intervals using bootstrapping.

Example: BootstrapType="student"

Data Types: char | string

NumBootstraps — Number of bootstrap samples to draw

0 (default) | nonnegative integer scalar

Number of bootstrap samples to draw for computing pointwise confidence intervals, specified as a nonnegative integer scalar.

If you specify NumBootstraps as a positive integer, then rocmetrics uses NumBootstraps bootstrap samples. To create each bootstrap sample, the function randomly selects n out of the n rows of input data with replacement. The default value 0 implies that rocmetrics does not use bootstrapping.

rocmetrics computes confidence intervals by using either cross-validated data or bootstrap samples. Therefore, if you specify cross-validated data for Labels, Scores, and Weights, then NumBootstraps must be 0.

For details, see Pointwise Confidence Intervals.

Example: NumBootstraps=500

Data Types: single | double

NumBootstrapsStudentizedSE — Number of bootstrap samples to draw for studentized standard error estimate

100 (default) | positive integer scalar

Number of bootstrap samples to draw for the studentized standard error estimate, specified as a positive integer scalar.

This argument is valid only when you specify NumBootstraps as a positive integer and BootstrapType as "student" to compute studentized bootstrap confidence intervals. rocmetrics estimates the studentized standard error estimate by using NumBootstrapsStudentizedSE bootstrap data samples.

Example: NumBootstrapsStudentizedSE=500

Data Types: single | double

Properties

Performance Metrics

AUC — Area under ROC curve

numeric vector | numeric matrix

This property is read-only.

Area under the ROC curve (AUC), specified as a numeric vector or matrix.

rocmetrics computes the AUC for each one-versus-all ROC curve (that is, for each class). The column order of the AUC property value matches the class order in ClassNames.

For a binary problem where you specify Scores as a two-column matrix, this property is a 1-by-2 vector containing identical AUC values. The AUC values are identical because the overall model performance on one class is identical to the performance on the other class for a binary problem.

If rocmetrics computes confidence intervals for AUC, the AUC property value is a matrix in which the first row corresponds to the AUC values, and the second and third rows correspond to the lower and upper bounds, respectively. rocmetrics computes confidence intervals for AUC if the function also computes confidence intervals for the performance metrics and you set FixedMetric to "Thresholds" (default), "FalsePositiveRate", or "TruePositiveRate".

Data Types: single | double

Metrics — Performance metrics

table

This property is read-only.

Performance metrics, specified as a table.

The table contains performance metric values for all classes, vertically concatenated according to the class order in ClassNames. The table has a row for each unique threshold value for each class. rocmetrics determines the threshold values to use based on the value of FixedMetric, FixedMetricValues, and UseNearestNeighbor. For details, see Thresholds, Fixed Metric, and Fixed Metric Values.

The number of rows for each class in the table is the number of unique threshold values.

Each row of the table contains these variables: ClassName, Threshold, FalsePositiveRate, and TruePositiveRate, as well as a variable for each additional metric specified in AdditionalMetrics. If you specify a custom metric, rocmetrics names the metric "CustomMetricN", where N is the number that refers to the custom metric. For example, "CustomMetric1" corresponds to the first custom metric specified by AdditionalMetrics.

Each variable in the Metrics table contains a vector or a three-column matrix.

• If rocmetrics does not compute confidence intervals, each variable contains a vector.

• If rocmetrics computes confidence intervals, both ClassName and the variable for FixedMetric (Threshold, FalsePositiveRate, TruePositiveRate, or an additional metric) contain a vector, and the other variables contain a three-column matrix. The first column of the matrix corresponds to the metric values, and the second and third columns correspond to the lower and upper bounds, respectively.

Data Types: table

Classification Model Properties

You can specify the following properties when creating a rocmetrics object.

ClassNames — Class names

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

This property is read-only.

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

For details, see the input argument ClassNames, which sets this property. (The software treats character or string arrays as cell arrays of character vectors.)

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

Cost — Misclassification cost

square matrix

This property is read-only.

Misclassification cost, specified as a square matrix.

For details, see the Cost name-value argument, which sets this property.

Data Types: single | double

Labels — True class labels

numeric vector | logical vector | categorical vector | cell array

This property is read-only.

True class labels, specified as a numeric vector, logical vector, categorical vector, cell array of character vectors, or cell array of one of these types for cross-validated data.

For details, see the input argument Labels, which sets this property. (The software treats character or string arrays as cell arrays of character vectors.)

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

Prior — Prior class probabilities

numeric vector

This property is read-only.

Prior class probabilities, specified as a numeric vector.

For details, see the Prior name-value argument, which sets this property. If you specify this argument as a character vector or string scalar ("empirical" or "uniform"), rocmetrics computes the prior probabilities and stores the Prior property as a numeric vector.

Data Types: single | double

Scores — Classification scores

numeric matrix | cell array of numeric matrices

This property is read-only.

Classification scores, specified as a numeric matrix or a cell array of numeric matrices.

For details, see the input argument Scores, which sets this property.

Note

If you specify the ApplyCostToScores name-value argument as true, the software stores the transformed scores S*(-C), where the scores S are specified by the Scores argument, and the misclassification cost matrix C is specified by the Cost name-value argument. (since R2024a)

Data Types: single | double | cell

Weights — Observation weights

numeric vector of positive values | cell array containing numeric vectors of positive values

This property is read-only.

Observation weights, specified as a numeric vector of positive values or a cell array containing numeric vectors of positive values.

For details, see the Weights name-value argument, which sets this property.

Data Types: single | double | cell

Object Functions

addMetrics	Compute additional classification performance metrics
average	Compute performance metrics for average receiver operating characteristic (ROC) curve in multiclass problem
plot	Plot receiver operating characteristic (ROC) curves and other performance curves

Examples

Plot ROC Curve for Binary Classifier

Compute the performance metrics (FPR and TPR) for a binary classification problem by creating a rocmetrics object, and plot a ROC curve by using plot function.

Load the ionosphere data set. This data set has 34 predictors (X) and 351 binary responses (Y) for radar returns, either bad ('b') or good ('g').

load ionosphere

Partition the data into training and test sets. Use approximately 80% of the observations to train a support vector machine (SVM) model, and 20% of the observations to test the performance of the trained model on new data. Partition the data using cvpartition.

rng("default") % For reproducibility of the partition
c = cvpartition(Y,Holdout=0.20);
trainingIndices = training(c); % Indices for the training set
testIndices = test(c); % Indices for the test set
XTrain = X(trainingIndices,:);
YTrain = Y(trainingIndices);
XTest = X(testIndices,:);
YTest = Y(testIndices);

Train an SVM classification model.

Mdl = fitcsvm(XTrain,YTrain);

Compute the classification scores for the test set.

[~,Scores] = predict(Mdl,XTest);
size(Scores)

ans = 1×2

    70     2

The output Scores is a matrix of size 70-by-2. The column order of Scores follows the class order in Mdl. Display the class order stored in Mdl.ClassNames.

Mdl.ClassNames

ans = 2x1 cell
    {'b'}
    {'g'}

Create a rocmetrics object by using the true labels in YTest and the classification scores in Scores. Specify the column order of Scores using Mdl.ClassNames.

rocObj = rocmetrics(YTest,Scores,Mdl.ClassNames);

rocObj is a rocmetrics object that stores the AUC values and performance metrics for each class in the AUC and Metrics properties. Display the AUC property.

rocObj.AUC

ans = 1×2

    0.8587    0.8587

For a binary classification problem, the AUC values are equal to each other.

The table in Metrics contains the performance metric values for both classes, vertically concatenated according to the class order. Find the rows for the first class in the table, and display the first eight rows.

idx = strcmp(rocObj.Metrics.ClassName,Mdl.ClassNames(1));
head(rocObj.Metrics(idx,:))

    ClassName    Threshold    FalsePositiveRate    TruePositiveRate
    _________    _________    _________________    ________________

      {'b'}       15.545              0                     0      
      {'b'}       15.545              0                  0.04      
      {'b'}       15.105              0                  0.08      
      {'b'}       11.424              0                  0.16      
      {'b'}       10.077              0                   0.2      
      {'b'}       9.9716              0                  0.24      
      {'b'}       9.9417              0                  0.28      
      {'b'}       9.0338              0                  0.32      

Plot the ROC curve for each class by using the plot function.

plot(rocObj)

For each class, the plot function plots a ROC curve and displays a filled circle marker at the model operating point. The legend displays the class name and AUC value for each curve.

Note that you do not need to examine ROC curves for both classes in a binary classification problem. The two ROC curves are symmetric, and the AUC values are identical. A TPR of one class is a true negative rate (TNR) of the other class, and TNR is 1-FPR. Therefore, a plot of TPR versus FPR for one class is the same as a plot of 1-FPR versus 1-TPR for the other class.

Plot the ROC curve for the first class only by specifying the ClassNames name-value argument.

plot(rocObj,ClassNames=Mdl.ClassNames(1))

Plot ROC Curves for Multiclass Classifier

Compute the performance metrics (FPR and TPR) for a multiclass classification problem by creating a rocmetrics object, and plot a ROC curve for each class by using the plot function. Specify the AverageROCType name-value argument of plot to create the average ROC curve for the multiclass problem.

Load the fisheriris data set. The matrix meas contains flower measurements for 150 different flowers. The vector species lists the species for each flower. species contains three distinct flower names.

load fisheriris

Train a classification tree that classifies observations into one of the three labels. Cross-validate the model using 10-fold cross-validation.

rng("default") % For reproducibility
Mdl = fitctree(meas,species,Crossval="on");

Compute the classification scores for validation-fold observations.

[~,Scores] = kfoldPredict(Mdl);
size(Scores)

ans = 1×2

   150     3

The output Scores is a matrix of size 150-by-3. The column order of Scores follows the class order in Mdl. Display the class order stored in Mdl.ClassNames.

Mdl.ClassNames

ans = 3x1 cell
    {'setosa'    }
    {'versicolor'}
    {'virginica' }

Create a rocmetrics object by using the true labels in species and the classification scores in Scores. Specify the column order of Scores using Mdl.ClassNames.

rocObj = rocmetrics(species,Scores,Mdl.ClassNames);

rocObj is a rocmetrics object that stores the AUC values and performance metrics for each class in the AUC and Metrics properties. Display the AUC property.

rocObj.AUC

ans = 1×3

    1.0000    0.9636    0.9636

The table in Metrics contains the performance metric values for all three classes, vertically concatenated according to the class order. Find and display the rows for the second class in the table.

idx = strcmp(rocObj.Metrics.ClassName,Mdl.ClassNames(2));
rocObj.Metrics(idx,:)

ans=13×4 table
      ClassName       Threshold    FalsePositiveRate    TruePositiveRate
    ______________    _________    _________________    ________________

    {'versicolor'}           1              0                    0      
    {'versicolor'}           1           0.01                  0.7      
    {'versicolor'}     0.95455           0.02                  0.8      
    {'versicolor'}     0.91304           0.03                  0.9      
    {'versicolor'}        -0.2           0.04                  0.9      
    {'versicolor'}    -0.33333           0.06                  0.9      
    {'versicolor'}        -0.6           0.08                  0.9      
    {'versicolor'}    -0.86957           0.12                 0.92      
    {'versicolor'}    -0.91111           0.16                 0.96      
    {'versicolor'}    -0.95122           0.31                 0.96      
    {'versicolor'}    -0.95238           0.38                 0.98      
    {'versicolor'}    -0.95349           0.44                 0.98      
    {'versicolor'}          -1              1                    1      

Plot the ROC curve for each class. Specify AverageROCType="micro" to compute the performance metrics for the average ROC curve using the micro-averaging method.

plot(rocObj,AverageROCType="micro")

The filled circle markers indicate the model operating points. The legend displays the class name and AUC value for each curve.

Plot ROC Curves for Discriminant Analysis Classifier with Misclassification Costs

k-nearest neighbor (KNN), discriminant analysis, and naive Bayes classifiers use expected classification costs rather than scores to predict labels. When you want to use nondefault misclassification costs to create ROC curves for these models, set the ApplyCostToScores name-value argument of the rocmetrics function to true.

Read the sample file CreditRating_Historical.dat into a table. The predictor data consists of financial ratios and industry sector information for a list of corporate customers. The response variable consists of credit ratings assigned by a rating agency.

creditrating = readtable("CreditRating_Historical.dat");

Because each value in the ID variable is a unique customer ID, that is, length(unique(creditrating.ID)) is equal to the number of observations in creditrating, the ID variable is a poor predictor. Remove the ID variable from the table.

creditrating = removevars(creditrating,"ID");

Combine all the A ratings into one rating. Do the same for the B and C ratings, so that the response variable has three distinct ratings. Among the three ratings, A is considered the best and C the worst.

Rating = categorical(creditrating.Rating);
Rating = mergecats(Rating,["AAA","AA","A"],"A");
Rating = mergecats(Rating,["BBB","BB","B"],"B");
Rating = mergecats(Rating,["CCC","CC","C"],"C");
creditrating.Rating = Rating;

Assume that specific costs are associated with misclassifying the credit ratings of customers. Create a matrix variable that contains the misclassification costs. Create another variable that specifies the class names and their order in the matrix variable.

classificationCosts = [0 100 200; 500 0 100; 1000 500 0];
classNames = categorical(["A","B","C"]);

The costs indicate that classifying a customer with bad credit as a customer with good credit is more costly than classifying a customer with good credit as a customer with bad credit. For example, the cost of misclassifying a C rating customer as an A rating customer is $1000.

Partition the data into training and test sets. Use 75% of the observations to train a discriminant analysis classifier, and 25% of the observations to test the performance of the trained model on new data.

rng("default") % For reproducibility
c = cvpartition(creditrating.Rating,"Holdout",0.25);
trainRatings = creditrating(training(c),:);
testRatings = creditrating(test(c),:);

Train a discriminant analysis classifier. Specify the misclassification costs.

mdl = fitcdiscr(trainRatings,"Rating",Cost=classificationCosts, ...
    ClassNames=classNames);

Predict the class labels, scores, and expected classification costs for the observations in the test set.

[labels,scores,expectedCosts] = predict(mdl,testRatings);

For each observation, the predicted class label corresponds to the minimum expected classification cost among all classes rather than the greatest score (or posterior probability).

For example, display the predictions for the first observation in the test set.

firstLabel = labels(1)

firstLabel = categorical
     B 

firstScores = array2table(scores(1,:),VariableNames=["A","B","C"])

firstScores=1×3 table
       A          B           C     
    _______    _______    __________

    0.70807    0.29193    4.7141e-13

firstExpectedCosts = array2table(expectedCosts(1,:), ...
    VariableNames=["A","B","C"])

firstExpectedCosts=1×3 table
      A         B         C   
    ______    ______    ______

    145.96    70.807    170.81

The predicted label corresponds to class B, which has the lowest expected classification cost, even though class A has the greatest posterior probability.

Create a rocmetrics object by using the true labels in testRatings and the classification scores in scores. Specify the column order of scores. To use nondefault misclassification costs and scores returned by a discriminant analysis model, specify the Cost and ApplyCostToScores name-value arguments.

roc = rocmetrics(testRatings.Rating,scores,classNames, ...
    Cost=classificationCosts,ApplyCostToScores=true);

Notice that the scores stored in rocmetrics are the negative expected classification costs.

isequal(roc.Scores,-expectedCosts)

ans = logical
   1

Plot the ROC curve for each class by using the plot function.

plot(roc,ClassNames=classNames)

For each class, the plot function plots a curve. The filled circle markers indicate the model operating points.

Train a cross-validated discriminant analysis classifier by using the entire creditrating data set.

cvmdl = fitcdiscr(creditrating,"Rating",Cost=classificationCosts, ...
    ClassNames=classNames,CrossVal="on")

cvmdl = 
  ClassificationPartitionedModel
    CrossValidatedModel: 'Discriminant'
         PredictorNames: {'WC_TA'  'RE_TA'  'EBIT_TA'  'MVE_BVTD'  'S_TA'  'Industry'}
           ResponseName: 'Rating'
        NumObservations: 3932
                  KFold: 10
              Partition: [1x1 cvpartition]
             ClassNames: [A    B    C]
         ScoreTransform: 'none'

The fitcdiscr function creates a ClassificationPartitionedModel object of type Discriminant (CrossValidatedModel property value). To create the cross-validated model, the function completes these steps:

1. Randomly partition the data into 10 sets.

2. For each set, reserve the set as validation data, and train the model using the other 9 sets.

3. Store the 10 compact trained models in a 10-by-1 cell vector in the Trained property of the cross-validated model object.

cvmdl.Trained

ans=10×1 cell array
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}
    {1x1 classreg.learning.classif.CompactClassificationDiscriminant}

Predict the class label, scores, and expected classification costs for each observation.

[cvlabels,cvscores,cvexpectedCosts] = kfoldPredict(cvmdl);

Plot the ROC curve for each class.

cvroc = rocmetrics(creditrating.Rating,cvscores,classNames, ...
    Cost=classificationCosts,ApplyCostToScores=true);
plot(cvroc,ClassNames=classNames)

The cross-validation results are similar to the previous test set results.

Specify Scores as Vector

For generated samples containing outliers, train an isolation forest model and compute anomaly scores by using the iforest function. iforest returns scores as a vector. Use the scores to create a rocmetrics object. Plot the precision-recall curve using the anomaly scores, and find the model operating point for the isolation forest model.

Use a Gaussian copula to generate random data points from a bivariate distribution.

rng("default")
rho = [1,0.05;0.05,1];
n = 1000;
u = copularnd("Gaussian",rho,n);

Add noise to 5% of randomly selected observations to make the observations outliers.

noise = randperm(n,0.05*n);
true_tf = false(n,1);
true_tf(noise) = true;
u(true_tf,1) = u(true_tf,1)*5;

Train an isolation forest model by using the iforest function. Specify the fraction of anomalies in the training observations as 0.05.

[f,tf,scores] = iforest(u,ContaminationFraction=0.05);

f is an IsolationForest object. iforest also returns the anomaly indicators (tf) and anomaly scores (scores) for the training data. iforest determines the threshold value (f.ScoreThreshold) so that the function detects the specified fraction of training observations as anomalies.

Check the performance of the IsolationForest object by plotting the precision-recall curve and computing the area under the curve (AUC) value. Create a rocmetrics object by using the true anomaly indicators (true_tf) and anomaly scores (scores). A score value close to 1 indicates an anomaly, as does the value true in true_tf. Therefore, specify the class name for scores as true. Specify the AdditionalMetrics name-value argument to compute the precision values (or positive predictive values).

rocObj = rocmetrics(true_tf,scores,true,AdditionalMetrics="PositivePredictiveValue");

Plot the curve by using the plot function of rocmetrics. Specify the y-axis metric as precision (or positive predictive value) and the x-axis metric as recall (or true positive rate). Display a filled circle at the model operating point corresponding to f.ScoreThreshold. Compute the area under the precision-recall curve using the trapezoidal method of the trapz function, and display the value in the legend.

r = plot(rocObj,YAxisMetric="PositivePredictiveValue",XAxisMetric="TruePositiveRate");
hold on
idx = find(rocObj.Metrics.Threshold>=f.ScoreThreshold,1,'last');
scatter(rocObj.Metrics.TruePositiveRate(idx), ...
    rocObj.Metrics.PositivePredictiveValue(idx), ...
    [],r.Color,"filled")
xyData = rmmissing([r.XData r.YData]);
auc = trapz(xyData(:,1),xyData(:,2));
legend(join([r.DisplayName " (AUC = " string(auc) ")"],""),"true Model Operating Point")
xlabel("Recall")
ylabel("Precision")
title("Precision-Recall Curve")
hold off

Compute Confidence Intervals Using Bootstrapping

Compute the confidence intervals for FPR and TPR for fixed threshold values by using bootstrap samples, and plot the confidence intervals for TPR on the ROC curve by using the plot function.

Load the ionosphere data set. This data set has 34 predictors (X) and 351 binary responses (Y) for radar returns, either bad ('b') or good ('g').

load ionosphere

Partition the data into training and test sets. Use approximately 80% of the observations to train a support vector machine (SVM) model, and 20% of the observations to test the performance of the trained model on new data. Partition the data using cvpartition.

rng("default") % For reproducibility of the partition
c = cvpartition(Y,Holdout=0.20);
trainingIndices = training(c); % Indices for the training set
testIndices = test(c); % Indices for the test set
XTrain = X(trainingIndices,:);
YTrain = Y(trainingIndices);
XTest = X(testIndices,:);
YTest = Y(testIndices);

Train an SVM classification model.

Mdl = fitcsvm(XTrain,YTrain);

Compute the classification scores for the test set.

[~,Scores] = predict(Mdl,XTest);

Create a rocmetrics object by using the true labels in YTest and the classification scores in Scores. Specify the column order of Scores using Mdl.ClassNames. Specify NumBootstraps as 100 to use 100 bootstrap samples to compute the confidence intervals.

rocObj = rocmetrics(YTest,Scores,Mdl.ClassNames, ...
    NumBootstraps=100);

Find the rows for the second class in the table of the Metrics property, and display the first eight rows.

idx = strcmp(rocObj.Metrics.ClassName,Mdl.ClassNames(2));
head(rocObj.Metrics(idx,:))

    ClassName    Threshold        FalsePositiveRate                 TruePositiveRate        
    _________    _________    __________________________    ________________________________

      {'g'}        7.196         0          0          0           0           0           0
      {'g'}        7.196         0          0          0    0.022222           0    0.093023
      {'g'}       6.2583         0          0          0    0.044444           0     0.11969
      {'g'}       5.5719         0          0          0    0.066667    0.020988     0.16024
      {'g'}       5.5643         0          0          0    0.088889    0.022635     0.18805
      {'g'}       5.4618      0.04          0    0.22222    0.088889    0.022635     0.18805
      {'g'}       5.3667      0.08          0       0.28    0.088889    0.022635     0.18805
      {'g'}       5.1525      0.08          0       0.28     0.11111    0.045035     0.19532

Each row of the table contains the metric value and its confidence intervals for FPR and TPR for a fixed threshold value. The Threshold variable is a column vector, and the FalsePositiveRate and TruePositiveRate variables are three-column matrices. The first column of the matrices corresponds to the metric values, and the second and third columns correspond to the lower and upper bounds, respectively.

Plot the ROC curve and the confidence intervals for TPR. Specify ShowConfidenceIntervals=true to show the confidence intervals, and specify one class to plot by using the ClassNames name-value argument.

plot(rocObj,ShowConfidenceIntervals=true,ClassNames=Mdl.ClassNames(2))

The shaded area around the ROC curve indicates the confidence intervals. The confidence intervals represent the uncertainty of the curve due to the variance in the test set for the trained model.

Compute Confidence Intervals with Cross-Validated Input Data

Compute the confidence intervals for FPR and TPR for fixed threshold values by using cross-validated data, and plot the confidence intervals for TPR on the ROC curve by using the plot function.

Load the fisheriris data set. The matrix meas contains flower measurements for 150 different flowers. The vector species lists the species for each flower. species contains three distinct flower names.

load fisheriris

Train a naive Bayes model that classifies observations into one of the three labels. Cross-validate the model using 10-fold cross-validation.

rng("default") % For reproducibility
Mdl = fitcnb(meas,species,Crossval="on");

Compute the classification scores for validation-fold observations.

[~,Scores] = kfoldPredict(Mdl);

Store the cross-validated scores and the corresponding true labels in cell arrays, so that each element in the cell arrays corresponds to one validation fold.

cv = Mdl.Partition;
numTestSets = cv.NumTestSets;
cvLabels = cell(numTestSets,1);
cvScores = cell(numTestSets,1);
for i = 1:numTestSets
    testIdx = test(cv,i);
    cvLabels{i} = species(testIdx);
    cvScores{i} = Scores(testIdx,:);
end

Create a rocmetrics object using the cell arrays. If you specify true labels and scores by using cell arrays, rocmetrics computes the confidence intervals.

rocObj = rocmetrics(cvLabels,cvScores,Mdl.ClassNames);

Plot the ROC curve and the confidence intervals for TPR. Specify ShowConfidenceIntervals=true to show the confidence intervals.

plot(rocObj,ShowConfidenceIntervals=true)

The shaded area around each curve indicates the confidence intervals. The widths of the confidence intervals for setosa are 0 for nonzero false positive rates, so the plot does not have a shaded area for setosa. The confidence intervals reflect the uncertainty in the model due to the variance in the training and test sets.

Compare Classification Models

Train three different classification models: decision tree model, generalized additive model, and naive Bayes model. Compare the performance of the three models on a test data set using the ROC curves and the AUC values.

Load the 1994 census data stored in census1994.mat. The data set consists of demographic data from the US Census Bureau to predict whether an individual makes over $50,000 per year.

load census1994

census1994 contains the training data set adultdata and the test data set adulttest. Display the unique values in the response variable salary.

classNames = unique(adultdata.salary)

classNames = 2x1 categorical
     <=50K 
     >50K 

Train the three models by passing the training data adultdata and specifying the response variable name "salary". Specify the order of the classes by using the ClassNames name-value argument.

MdlTree = fitctree(adultdata,"salary",ClassNames=classNames);
MdlGAM = fitcgam(adultdata,"salary",ClassNames=classNames); 
MdlNB = fitcnb(adultdata,"salary",ClassNames=classNames);

Compute the classification scores for the test data set adulttest using the trained models.

[~,ScoresTree] = predict(MdlTree,adulttest);
[~,ScoresGAM] = predict(MdlGAM,adulttest);
[~,ScoresNB] = predict(MdlNB,adulttest);

Create a rocmetrics object for each model.

rocTree = rocmetrics(adulttest.salary,ScoresTree,classNames);
rocGAM = rocmetrics(adulttest.salary,ScoresGAM,classNames);
rocNB = rocmetrics(adulttest.salary,ScoresNB,classNames);

Plot the ROC curve for each model. By default, the plot function displays the class names and the AUC values in the legend. To include the model names in the legend instead of the class names, modify the DisplayName property of the ROCCurve object returned by the plot function.

figure
c = cell(3,1);
g = cell(3,1);
[c{1},g{1}] = plot(rocTree,ClassNames=classNames(1));
hold on
[c{2},g{2}] = plot(rocGAM,ClassNames=classNames(1));
[c{3},g{3}] = plot(rocNB,ClassNames=classNames(1));
modelNames = ["Decision Tree Model", ...
    "Generalized Additive Model","Naive Bayes Model"];
for i = 1 : 3
    c{i}.DisplayName = replace(c{i}.DisplayName, ...
        string(classNames(1)),modelNames(i));
    g{i}(1).DisplayName = join([modelNames(i),"Operating Point"]);
end
hold off

The generalized additive model (MdlGAM) has the highest AUC value, and the decision tree model (MdlTree) has the lowest. This result suggests that MdlGAM has better average performance for the test data set than MdlTree and MdlNB.

Find Model Operating Point and Optimal Operating Point

Find the model operating point and the optimal operating point for a binary classification model. Classify observations in a test data set by using a new threshold corresponding to the optimal operating point.

Load the ionosphere data set. This data set has 34 predictors (X) and 351 binary responses (Y) for radar returns, either bad (b) or good (g).

load ionosphere

Partition the data into training and test sets. Use approximately 75% of the observations to train a support vector machine (SVM) model, and 25% of the observations to test the performance of the trained model on new data. Partition the data using cvpartition.

rng("default") % For reproducibility of the partition
c = cvpartition(Y,Holdout=0.25);
trainingIndices = training(c); % Indices for the training set
testIndices = test(c); % Indices for the test set
XTrain = X(trainingIndices,:);
YTrain = Y(trainingIndices);
XTest = X(testIndices,:);
YTest = Y(testIndices);

Train an SVM classification model.

Mdl = fitcsvm(XTrain,YTrain);

Display the class order stored in Mdl.ClassNames.

Mdl.ClassNames

ans = 2x1 cell
    {'b'}
    {'g'}

Compute the classification scores for the test set.

[Y1,Scores] = predict(Mdl,XTest);

Create a rocmetrics object by using the true labels in YTest and the classification scores in Scores. Specify the column order of Scores using Mdl.ClassNames.

rocObj = rocmetrics(YTest,Scores,Mdl.ClassNames);

Find the model operating point in the Metrics property of rocObj for class b. The predict function classifies an observation into the class yielding a larger score, which corresponds to the class with a nonnegative adjusted score. That is, the typical threshold value used by the predict function is 0. Among the rows in the Metrics property of rocObj for class b, find the point that has the smallest nonnegative threshold value. The point on the curve indicates identical performance to the performance of the threshold value 0.

idx_b = strcmp(rocObj.Metrics.ClassName,"b");
X = rocObj.Metrics(idx_b,:).FalsePositiveRate;
Y = rocObj.Metrics(idx_b,:).TruePositiveRate;
T = rocObj.Metrics(idx_b,:).Threshold;
idx_model = find(T>=0,1,"last");
modelpt = [T(idx_model) X(idx_model) Y(idx_model)]

modelpt = 1×3

    1.2654    0.0179    0.5806

For binary classification, an optimal operating point that minimizes the average misclassification cost is a point at which the ROC curve intersects a straight line with slope $$m$$, where $$m$$ is defined as

$\mathit{m}=\frac{\mathrm{cost}\left(\mathit{P}|\mathit{N}\right)-\mathrm{cost}\left(\mathit{N}|\mathit{N}\right)}{\mathrm{cost}\left(\mathit{N}|\mathit{P}\right)-\mathrm{cost}\left(\mathit{P}|\mathit{P}\right)}\cdot \frac{\mathit{n}}{\mathit{p}}$.

$$p$$ is the total number of observations in the positive class, and $$n$$ is the total number of observations in the negative class. The cost values are the components of the cost matrix $$C$$:

$$\mathit{C}=\left[\begin{array}{cc}\mathrm{cost}\left(\mathit{P}|\mathit{P}\right)& \mathrm{cost}\left(\mathit{N}|\mathit{P}\right)\\ \mathrm{cost}\left(\mathit{P}|\mathit{N}\right)& \mathrm{cost}\left(\mathit{N}|\mathit{N}\right)\end{array}\right]$$

cost(N|P) is the cost of misclassifying a positive class as a negative class, and cost(P|N) is the cost of misclassifying a negative class as a positive class. According to the class order in Mdl.ClassNames, the positive class P corresponds to class b.

Among the points on the ROC curve that intersect a line with slope $$m$$, choose one that is closest to the perfect classifier point (FPR = 0, TPR = 1), which the perfect ROC curve passes.

Find the optimal operating point for the positive class b.

p = sum(strcmp(YTest,"b"));  
n = sum(~strcmp(YTest,"b")); 
cost = Mdl.Cost;
m = (cost(2,1)-cost(2,2))/(cost(1,2)-cost(1,1))*n/p;
[~,idx_opt] = min(X - Y/m);
optpt = [T(idx_opt) X(idx_opt) Y(idx_opt)]

optpt = 1×3

   -1.1978    0.1071    0.7742

Plot the ROC curve for class b by using the plot function, and display the optimal operating point by using the scatter function.

figure
r = plot(rocObj,ClassNames="b");
hold on 
scatter(optpt(2),optpt(3),"filled", ...
    DisplayName="b Optimal Operating Point");

Display the model operating point and the optimal operating point.

array2table([modelpt;optpt], ...
    RowNames=["Model Operating Point" "Optimal Operating Point"], ...
    VariableNames=["Threshold" "FalsePositiveRate" "TruePositiveRate"])

ans=2×3 table
                               Threshold    FalsePositiveRate    TruePositiveRate
                               _________    _________________    ________________

    Model Operating Point        1.2654         0.017857             0.58065     
    Optimal Operating Point     -1.1978          0.10714             0.77419     

Classify XTest using the optimal operating point. Assign an observation whose adjusted score is greater than or equal to the optimal threshold to the positive class b.

s = Scores(:,1) - Scores(:,2);
idx_b_opt = (s >= optpt(1));
Y2 = cell(size(YTest));
Y2(idx_b_opt) = {'b'};
Y2(~idx_b_opt) = {'g'};

Display the adjusted scores for the observations that have different labels in Y1 (labels from the predict function) and Y2 (labels from the optimal threshold optpt(1)).

s(~strcmp(Y1,Y2))

ans = 11×1

   -1.1703
   -0.8445
   -0.8235
   -0.4546
   -1.0719
   -0.4612
   -0.2191
   -1.1978
   -1.0114
   -1.1552
      ⋮

Eleven observations have adjusted scores less than 0 but greater than or equal to the optimal threshold.

Compute Performance Metrics for Specified Classes and Thresholds

After training a model for a multiclass classification problem, create a rocmetrics object for classes of interest only. Specify FixedMetricValues so that rocmetrics computes the performance metrics for the specified threshold values.

Read the sample file CreditRating_Historical.dat into a table. The predictor data consists of financial ratios and industry sector information for a list of corporate customers. The response variable consists of credit ratings assigned by a rating agency. Preview the first few rows of the data set.

creditrating = readtable("CreditRating_Historical.dat");
head(creditrating)

     ID      WC_TA     RE_TA     EBIT_TA    MVE_BVTD    S_TA     Industry    Rating 
    _____    ______    ______    _______    ________    _____    ________    _______

    62394     0.013     0.104     0.036      0.447      0.142        3       {'BB' }
    48608     0.232     0.335     0.062      1.969      0.281        8       {'A'  }
    42444     0.311     0.367     0.074      1.935      0.366        1       {'A'  }
    48631     0.194     0.263     0.062      1.017      0.228        4       {'BBB'}
    43768     0.121     0.413     0.057      3.647      0.466       12       {'AAA'}
    39255    -0.117    -0.799      0.01      0.179      0.082        4       {'CCC'}
    62236     0.087     0.158     0.049      0.816      0.324        2       {'BBB'}
    39354     0.005     0.181     0.034      2.597      0.388        7       {'AA' }

Because each value in the ID variable is a unique customer ID, that is, length(unique(creditrating.ID)) is equal to the number of observations in creditrating, the ID variable is a poor predictor. Remove the ID variable from the table, and convert the Industry variable to a categorical variable.

creditrating = removevars(creditrating,"ID");
creditrating.Industry = categorical(creditrating.Industry);

Partition the data into training and test sets. Use approximately 80% of the observations to train a neural network model, and 20% of the observations to test the performance of the trained model on new data. Partition the data using cvpartition.

rng("default") % For reproducibility of the partition
c = cvpartition(creditrating.Rating,"Holdout",0.20);
trainingIndices = training(c); % Indices for the training set
testIndices = test(c); % Indices for the test set
creditTrain = creditrating(trainingIndices,:);
creditTest = creditrating(testIndices,:);

Train a neural network classifier by passing the training data creditTrain to the fitcnet function.

Mdl = fitcnet(creditTrain,"Rating");

Compute classification scores and predict credit ratings for the test set observations.

[labels,Scores] = predict(Mdl,creditTest);

The classification scores for a neural network classifier correspond to posterior probabilities.

Assume that you want to evaluate the model only for the ratings B, BB, and BBB, and ignore the rest of the ratings.

Display the order of the ratings in the model stored in the ClassNames property, and identify the classes to evaluate.

Mdl.ClassNames

ans = 7x1 cell
    {'A'  }
    {'AA' }
    {'AAA'}
    {'B'  }
    {'BB' }
    {'BBB'}
    {'CCC'}

idx_Class = [4 5 6];
classesToEvaluate = Mdl.ClassNames(idx_Class);

Find the indices of the observations for the three classes (B, BB, BBB).

idx = ismember(creditTest.Rating,classesToEvaluate);

Create a rocmetrics object using the true labels and scores for the three classes. Specify FixedMetricValues=1:-0.25:-1 so that rocmetrics computes the performance metrics for the specified threshold values.

thresholds = 1:-0.25:-1;
rocObj = rocmetrics(creditTest.Rating(idx),Scores(idx,idx_Class), ...
    classesToEvaluate,FixedMetricValues=thresholds);

Display the computed metrics stored in the Metrics property.

rocObj.Metrics

ans=27×4 table
    ClassName    Threshold    FalsePositiveRate    TruePositiveRate
    _________    _________    _________________    ________________

     {'B' }         0.9244               0                   0     
     {'B' }        0.78248               0             0.10938     
     {'B' }        0.50353        0.010309             0.26562     
     {'B' }        0.25055        0.023196             0.45312     
     {'B' }      0.0070141        0.046392              0.5625     
     {'B' }       -0.23768         0.10825              0.6875     
     {'B' }       -0.49955          0.1701             0.76562     
     {'B' }       -0.74956         0.50258             0.85938     
     {'B' }       -0.96771               1                   1     
     {'BB'}        0.95584               0                   0     
     {'BB'}        0.75096        0.044944             0.18378     
     {'BB'}        0.50175         0.10487             0.45405     
     {'BB'}        0.25636         0.14607             0.61081     
     {'BB'}      0.0026063         0.23596             0.75676     
     {'BB'}       -0.24187         0.33333             0.84865     
     {'BB'}       -0.49625          0.4794             0.93514     
      ⋮

The Metrics property contains the performance metrics for the three ratings B, BB, and BBB and the specified threshold values only. The default UseNearestNeighbor value is true if rocmetrics does not compute confidence intervals. Therefore, for each specified threshold value, rocmetrics selects an adjusted score value nearest to the specified value and uses the nearest value as a threshold. Display the specified threshold values and the actual threshold values used for each class.

idx_B = strcmp(rocObj.Metrics.ClassName,"B");
idx_BB = strcmp(rocObj.Metrics.ClassName,"BB");
idx_BBB = strcmp(rocObj.Metrics.ClassName,"BBB");
table(thresholds',rocObj.Metrics.Threshold(idx_B), ...
    rocObj.Metrics.Threshold(idx_BB), ...
    rocObj.Metrics.Threshold(idx_BBB), ...
    VariableNames=["Fixed Threshold";string(classesToEvaluate)])

ans=9×4 table
    Fixed Threshold        B           BB          BBB   
    _______________    _________    _________    ________

             1            0.9244      0.95584     0.94197
          0.75           0.78248      0.75096     0.75019
           0.5           0.50353      0.50175     0.50325
          0.25           0.25055      0.25636     0.25134
             0         0.0070141    0.0026063     0.01411
         -0.25          -0.23768     -0.24187    -0.23063
          -0.5          -0.49955     -0.49625    -0.49643
         -0.75          -0.74956     -0.74413    -0.74864
            -1          -0.96771     -0.94197    -0.96189

More About

Receiver Operating Characteristic (ROC) Curve

A ROC curve shows the true positive rate versus the false positive rate for different thresholds of classification scores.

The true positive rate and the false positive rate are defined as follows:

• True positive rate (TPR), also known as recall or sensitivity — TP/(TP+FN), where TP is the number of true positives and FN is the number of false negatives

• False positive rate (FPR), also known as fallout or 1-specificity — FP/(TN+FP), where FP is the number of false positives and TN is the number of true negatives

Each point on a ROC curve corresponds to a pair of TPR and FPR values for a specific threshold value. You can find different pairs of TPR and FPR values by varying the threshold value, and then create a ROC curve using the pairs. For each class, rocmetrics uses all distinct adjusted score values as threshold values to create a ROC curve.

For a multiclass classification problem, rocmetrics formulates a set of one-versus-all binary classification problems to have one binary problem for each class, and finds a ROC curve for each class using the corresponding binary problem. Each binary problem assumes one class as positive and the rest as negative.

For a binary classification problem, if you specify the classification scores as a matrix, rocmetrics formulates two one-versus-all binary classification problems. Each of these problems treats one class as a positive class and the other class as a negative class, and rocmetrics finds two ROC curves. Use one of the curves to evaluate the binary classification problem.

For more details, see ROC Curve and Performance Metrics.

Area Under ROC Curve (AUC)

The area under a ROC curve (AUC) corresponds to the integral of a ROC curve (TPR values) with respect to FPR from FPR = 0 to FPR = 1.

The AUC provides an aggregate performance measure across all possible thresholds. The AUC values are in the range 0 to 1, and larger AUC values indicate better classifier performance.

One-Versus-All (OVA) Coding Design

The one-versus-all (OVA) coding design reduces a multiclass classification problem to a set of binary classification problems. In this coding design, each binary classification treats one class as positive and the rest of the classes as negative. rocmetrics uses the OVA coding design for multiclass classification and evaluates the performance on each class by using the binary classification that the class is positive.

For example, the OVA coding design for three classes formulates three binary classifications:

$$\begin{array}{cccc}& \text{Binary 1}& \text{Binary }2& \text{Binary 3}\\ \text{Class 1}& 1& -1& -1\\ \text{Class 2}& -1& 1& -1\\ \text{Class 3}& -1& -1& 1\end{array}$$

Each row corresponds to a class, and each column corresponds to a binary classification problem. The first binary classification assumes that class 1 is a positive class and the rest of the classes are negative. rocmetrics evaluates the performance on the first class by using the first binary classification problem.

Model Operating Point

The model operating point represents the FPR and TPR corresponding to the typical threshold value.

The typical threshold value depends on the input format of the Scores argument (classification scores) specified when you create a rocmetrics object:

• If you specify Scores as a matrix, rocmetrics assumes that the values in Scores are the scores for a multiclass classification problem and uses adjusted score values. A multiclass classification model classifies an observation into a class that yields the largest score, which corresponds to a nonnegative score in the adjusted scores. Therefore, the threshold value is 0.

• If you specify Scores as a column vector, rocmetrics assumes that the values in Scores are posterior probabilities of the class specified in ClassNames. A binary classification model classifies an observation into a class that yields a higher posterior probability, that is, a posterior probability greater than 0.5. Therefore, the threshold value is 0.5.

For a binary classification problem, you can specify Scores as a two-column matrix or a column vector. However, if the classification scores are not posterior probabilities, you must specify Scores as a matrix. A binary classifier classifies an observation into a class that yields a larger score, which is equivalent to a class that yields a nonnegative adjusted score. Therefore, if you specify Scores as a matrix for a binary classifier, rocmetrics can find a correct model operating point using the same scheme that it applies to a multiclass classifier. If you specify classification scores that are not posterior probabilities as a vector, rocmetrics cannot identify a correct model operating point because it always uses 0.5 as a threshold for the model operating point.

The plot function displays a filled circle marker at the model operating point for each ROC curve (see ShowModelOperatingPoint). The function chooses a point corresponding to the typical threshold value. If the curve does not have a data point for the typical threshold value, the function finds a point that has the smallest threshold value greater than the typical threshold. The point on the curve indicates identical performance to the performance of the typical threshold value.

Algorithms

Adjusted Scores for Multiclass Classification Problem

For each class, rocmetrics adjusts the classification scores (input argument Scores of rocmetrics) relative to the scores for the rest of the classes if you specify Scores as a matrix. Specifically, the adjusted score for a class given an observation is the difference between the score for the class and the maximum value of the scores for the rest of the classes.

For example, if you have [s₁,s₂,s₃] in a row of Scores for a classification problem with three classes, the adjusted score values are [s₁-max(s₂,s₃),s₂-max(s₁,s₃),s₃-max(s₁,s₂)].

rocmetrics computes the performance metrics using the adjusted score values for each class.

For a binary classification problem, you can specify Scores as a two-column matrix or a column vector. Using a two-column matrix is a simpler option because the predict function of a classification object returns classification scores as a matrix, which you can pass to rocmetrics. If you pass scores in a two-column matrix, rocmetrics adjusts scores in the same way that it adjusts scores for multiclass classification, and it computes performance metrics for both classes. You can use the metric values for one of the two classes to evaluate the binary classification problem. The metric values for a class returned by rocmetrics when you pass a two-column matrix are equivalent to the metric values returned by rocmetrics when you specify classification scores for the class as a column vector.

References

[1] Fawcett, T. “ROC Graphs: Notes and Practical Considerations for Researchers”, Machine Learning 31, no. 1 (2004): 1–38.

[2] Zweig, M., and G. Campbell. “Receiver-Operating Characteristic (ROC) Plots: A Fundamental Evaluation Tool in Clinical Medicine.” Clinical Chemistry 39, no. 4 (1993): 561–577.

[3] Davis, J., and M. Goadrich. “The Relationship Between Precision-Recall and ROC Curves.” Proceedings of ICML ’06, 2006, pp. 233–240.

[4] Moskowitz, C. S., and M. S. Pepe. “Quantifying and Comparing the Predictive Accuracy of Continuous Prognostic Factors for Binary Outcomes.” Biostatistics 5, no. 1 (2004): 113–27.

[5] Huang, Y., M. S. Pepe, and Z. Feng. “Evaluating the Predictiveness of a Continuous Marker.” U. Washington Biostatistics Paper Series, 2006, 250–61.

[6] Briggs, W. M., and R. Zaretzki. “The Skill Plot: A Graphical Technique for Evaluating Continuous Diagnostic Tests.” Biometrics 64, no. 1 (2008): 250–256.

[7] Bettinger, R. “Cost-Sensitive Classifier Selection Using the ROC Convex Hull Method.” SAS Institute, 2003.

[8] DiCiccio, Thomas J., and Bradley Efron. “Bootstrap Confidence Intervals.” Statistical Science 11, no. 3 (1996): 189–228.

[9] Efron, Bradley, and Robert J. Tibshirani. An Introduction to the Bootstrap. New York: Chapman & Hall, 1993.

[10] Efron, Bradley. The Jackknife, the Bootstrap and Other Resampling Plans. Philadelphia: The Society for Industrial and Applied Mathematics, 1982.

[11] Davison, A. C., and D. V. Hinkley. Bootstrap Methods and Their Applications. Cambridge University Press, 1997.

Extended Capabilities

Automatic Parallel Support

Accelerate code by automatically running computation in parallel using Parallel Computing Toolbox™.

To run bootstrap iterations in parallel, specify the BootstrapOptions name-value argument when you create a rocmetrics object and set the UseParallel field of the options structure to true using statset.

For example: NumBootstraps=1000,BootstrapOptions=statset(UseParallel=true)

For more information about parallel computing, see Run MATLAB Functions with Automatic Parallel Support (Parallel Computing Toolbox).

Version History

Introduced in R2022a

R2024a: Apply misclassification costs to scores for k-nearest neighbor (KNN), discriminant analysis, and naive Bayes models

If you use nondefault misclassification costs by specifying the Cost name-value argument, and you use scores for a KNN, discriminant analysis, or naive Bayes model, set the ApplyCostToScores name-value argument to true. Because these models use expected classification costs rather than scores to predict labels, this setting enables the function to transform the scores appropriately.

For an example, see Plot ROC Curves for Discriminant Analysis Classifier with Misclassification Costs.