Daniel@0: function [FPrate, TPrate, AUC, thresholds] = computeROC(confidence, testClass)
Daniel@0: % function [FPrate, TPrate, AUC, thresholds] = computeROC(confidence, testClass)
Daniel@0: %
Daniel@0: % computeROC computes the data for an ROC curve based on a classifier's confidence output.
Daniel@0: % It returns the false positive rate and the true positive rate along with
Daniel@0: % the area under the ROC curve, and the list of thresholds.
Daniel@0: %
Daniel@0: % Inputs: 
Daniel@0: %           - confidence(i) is proportional to the probability that
Daniel@0: %             testClass(i) is positive
Daniel@0: %
Daniel@0: % testClass = 0 => target absent
Daniel@0: % testClass = 1 => target present
Daniel@0: %
Daniel@0: % Based on algorithms 2 and 4 from Tom Fawcett's paper "ROC Graphs: Notes and
Daniel@0: % Practical Considerations for Data Mining Researchers" (2003)
Daniel@0: % http://www.hpl.hp.com/techreports/2003/HPL-2003-4.pdf"
Daniel@0: %
Daniel@0: % Vlad Magdin, 21 Feb 2005
Daniel@0: 
Daniel@0: % break ties in scores
Daniel@0: S = rand('state');
Daniel@0: rand('state',0);
Daniel@0: confidence = confidence + rand(size(confidence))*10^(-10);
Daniel@0: rand('state',S)
Daniel@0: [thresholds order] = sort(confidence, 'descend');
Daniel@0: testClass = testClass(order);
Daniel@0: 
Daniel@0: %%% -- calculate TP/FP rates and totals -- %%%
Daniel@0: AUC = 0;
Daniel@0: faCnt = 0;
Daniel@0: tpCnt = 0;
Daniel@0: falseAlarms = zeros(1,size(thresholds,2));
Daniel@0: detections = zeros(1,size(thresholds,2));
Daniel@0: fPrev = -inf;
Daniel@0: faPrev = 0;
Daniel@0: tpPrev = 0;
Daniel@0: 
Daniel@0: P = max(size(find(testClass==1)));
Daniel@0: N = max(size(find(testClass==0)));
Daniel@0: 
Daniel@0: for i=1:length(thresholds)
Daniel@0:     if thresholds(i) ~= fPrev
Daniel@0:         falseAlarms(i) = faCnt;
Daniel@0:         detections(i) = tpCnt;
Daniel@0: 
Daniel@0:         AUC = AUC + polyarea([faPrev faPrev faCnt/N faCnt/N],[0 tpPrev tpCnt/P 0]);
Daniel@0: 
Daniel@0:         fPrev = thresholds(i);
Daniel@0:         faPrev = faCnt/N;
Daniel@0:         tpPrev = tpCnt/P;
Daniel@0:     end
Daniel@0:     
Daniel@0:     if testClass(i) == 1
Daniel@0:         tpCnt = tpCnt + 1;
Daniel@0:     else
Daniel@0:         faCnt = faCnt + 1;
Daniel@0:     end
Daniel@0: end
Daniel@0: 
Daniel@0: AUC = AUC + polyarea([faPrev faPrev 1 1],[0 tpPrev 1 0]);
Daniel@0: 
Daniel@0: FPrate = falseAlarms/N;
Daniel@0: TPrate = detections/P;