How to get ROC curve using SVM?

I have extracted the features of finger knuckles using LBP and now want to classify using SVM. I am using following code and I get values of F1 score , recall, percision, and accuracy. Now I also want to plot ROC curve. Plz help me which changes are required to get ROC for one-vs-all classicification using SVM. Feat file is attached.

clc

clear

close all

%%

Data = load ("Feat13.mat");

Features = Data.Feat (:,1:end-1);

Labels = Data.Feat (:,end);

% Features = Data.Feat (1:200,1:end-1); % for 1st 200 only

% Labels = Data.Feat (1:200,end);

% [m,n] = size(Features);

[m,n] = size(Data.Feat);

P = 0.50;

idx = randperm(m) ;

Training = Data.Feat(idx(1:round(P*m)),:);

Testing = Data.Feat(idx(round(P*m)+1:end),:);

Train_Features = Training(:,1:end-1);

Train_Labels = Training(:,end);

Test_Features = Testing(:,1:end-1);

Test_Labels = Testing(:,end);

rng(1); % For reproducibility

% SVMModel = fitcecoc(Train_Features,Train_Labels);

t = templateSVM('Standardize',true,'KernelFunction','linear');

SVMModel = fitcecoc(Train_Features,Train_Labels,'Learners',t);

error = resubLoss(SVMModel)

[Pred_TrainLabels,Pred_TrainScore] = predict(SVMModel,Train_Features);

[Pred_TestLabels,Pred_TestScore] = predict(SVMModel,Test_Features);

% ROC_data = roc_curve(Test_Labels,Pred_TestLabels)

% [Pred_WholeLabels,Pred_WholeScore] = predict(SVMModel,Features);

% ROC_data = roc_curve(Labels,Pred_wholeLabels)

% ROC_data = roc_curve(Labels,Pred_WholeLabels)

%

% [X,Y,T,AUC] = perfcurve(Labels,Pred_WholeScore,'5')

% plot(X,Y)

%% % For whole labels and scores

% [tpr,fpr,thresholds] = roc(Labels,Pred_WholeLabels);

% plotroc(Labels,Pred_WholeScore())

% %% For Test labels and scores only

% [tpr,fpr,thresholds] = roc(Labels,Pred_TestLabels);

% plotroc(Labels,Pred_TestScore())

% [c_matrixp,Result]= confusion.getMatrix(Test_Labels,Pred_TestLabels);

% [c_matrixp,Result]= confusion.getMatrix(Labels,Pred_WholeLabels);

fig = figure;

cm = confusionchart(Test_Labels,Pred_TestLabels,'RowSummary','row-normalized','ColumnSummary','column-normalized');

% cm = confusionchart(Labels,Pred_WholeLabels,'RowSummary','row-normalized','ColumnSummary','column-normalized');

cm.Title = 'Finger Creases Classification Using SVM';

cm.RowSummary = 'row-normalized';

cm.ColumnSummary = 'column-normalized';

[m, order] = confusionmat(Test_Labels,Pred_TestLabels);

Diagonal=diag(m);

sum_rows=sum(m,2);

Precision=Diagonal./sum_rows;

Overall_Precision=mean(Precision)

sum_col=sum(m,1);

recall=Diagonal./sum_col';

recall(isnan(recall))=0;

overall_recall=mean(recall)

F1_Score=2*((Overall_Precision*overall_recall)/(Overall_Precision+overall_recall))

accuracy = sum(Test_Labels == Pred_TestLabels,'all')/numel(Pred_TestLabels)

Answer 1

[Rohit]

You can compute a ROC curve and other performance curves by creating a rocmetrics object which support both binary and multiclass classification.

Refer to the below documentation links for further reference-

https://www.mathworks.com/help/stats/performance-curves.html

https://www.mathworks.com/help/stats/rocmetrics.html