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