libsvm_plotroc.m

function [auROC auPRC] = plotroc(y,x,params)
% Reused and adapted from the Matlab code of LIBSVM.
% plotroc draws the recevier operating characteristic(ROC) curve.
%
%  auc = plotroc(training_label, training_instance [, libsvm_options -v cv_fold]) 
%  Use cross-validation on training data to get decision values and plot ROC curve.
%
%  auc = plotroc(testing_label, testing_instance, model) 
%  Use the given model to predict testing data and obtain decision values
%  for ROC
%
% Example:
%   
% 	load('heart_scale.mat'); 
% 	plotroc(heart_scale_label, heart_scale_inst,'-v 5');
%
%	[y,x] = libsvmread('heart_scale');
%   	model = svmtrain(y,x);
% 	plotroc(y,x,model);

	rand('state',0); % reset random seed
	if nargin < 2
		help plotroc
		return
	elseif isempty(y) | isempty(x)
		error('Input data is empty');
	elseif sum(y == 1) + sum(y == -1) ~= length(y)
		error('ROC is only applicable to binary classes with labels 1, -1'); % check the trainig_file is binary
	elseif exist('params') && ~ischar(params)
		model = params;
		[predict_label,mse,deci] = libsvmpredict(y,x,model); % the procedure for predicting
		[auROC auPRC] = roc_curve(deci*model.Label(1),y);
	elseif exist('params') && strcmp(params, 'personal')
		[auROC auPRC] = roc_curve(x, y);
		[prc_stack_x, prc_stack_y, prc_thre, auPRC] = perfcurve(y, x, 1, 'xCrit', 'reca', 'yCrit', 'prec');
		%% This condition takes care of scenario when
		%% I predict the values and wish to get the auROC.
		%% 'y' in that case holds the ground_truth labels and
		%% 'x' holds the SVM predicted values.
	else
		if ~exist('params')
			params = [];
		end
		[param,fold] = proc_argv(params); % specify each parameter
		if fold <= 1
			error('The number of folds must be greater than 1');
		else	
			[deci,label_y] = get_cv_deci(y,x,param,fold); % get the value of decision and label after cross-calidation
			[auROC auPRC] = roc_curve(deci,label_y); % plot ROC curve
		end
	end
end

function [resu,fold] = proc_argv(params)
	resu=params;
	fold=5;
	if ~isempty(params) && ~isempty(regexp(params,'-v'))
        [fold_val,fold_start,fold_end] = regexp(params,'-v\s+\d+','match','start','end');
        if ~isempty(fold_val)
            [temp1,fold] = strread([fold_val{:}],'%s %u');
            resu([fold_start:fold_end]) = [];
        else
            error('Number of CV folds must be specified by "-v cv_fold"');
        end
    end
end

function [deci,label_y] = get_cv_deci(prob_y,prob_x,param,nr_fold)
	l=length(prob_y);
	deci = ones(l,1);
	label_y = ones(l,1);	 
	rand_ind = randperm(l); 
	for i=1:nr_fold % Cross training : folding
		test_ind=rand_ind([floor((i-1)*l/nr_fold)+1:floor(i*l/nr_fold)]');
		train_ind = [1:l]';
		train_ind(test_ind) = [];
		model = libsvmtrain(prob_y(train_ind),prob_x(train_ind,:),param);	   
		[predict_label,mse,subdeci] = libsvmpredict(prob_y(test_ind),prob_x(test_ind,:),model);
		deci(test_ind) = subdeci.*model.Label(1);
		label_y(test_ind) = prob_y(test_ind);
	end
end

function [auROC, auPRC] = roc_curve(deci,label_y)
	[val,ind] = sort(deci,'descend');
	roc_y = label_y(ind);
	stack_x = cumsum(roc_y == -1)/sum(roc_y == -1);
	stack_y = cumsum(roc_y == 1)/sum(roc_y == 1);
	auROC = sum((stack_x(2:length(roc_y),1)-stack_x(1:length(roc_y)-1,1)).*stack_y(2:length(roc_y),1));

        %Comment the above lines if using perfcurve of statistics toolbox
        [stack_x,stack_y,thre,auPRC]=perfcurve(label_y , deci, 1, 'xCrit', 'reca', 'yCrit', 'prec');
	%plot(stack_x,stack_y);
	%xlabel('False Positive Rate');
	%ylabel('True Positive Rate');
	%title(['ROC curve of (AUC = ' num2str(auc) ' )']);
end