}
Arrays.sort(wrep.BGS_univariate_ROC);
// remove-naive bayes loop
Remove rm = new Remove();
Classifier bn = new BayesNet();
// Classifier bnCopy;
Classifier bnCopyEv;
Evaluation ev;
K2 xK2 = new K2();
xK2.setInitAsNaiveBayes(true);
xK2.setMaxNrOfParents(1);
((BayesNet) bn).setSearchAlgorithm(xK2);
int winningNaiveRemove = -1;
double[] accuracyRes = new double[learnPostSel.numAttributes()];
double[] ROCRes = new double[learnPostSel.numAttributes()];
for (int removeNaive_counter = learnPostSel.numAttributes() -2;removeNaive_counter > 1 ; removeNaive_counter--){
StringBuilder sb = new StringBuilder("");
sb.append(removeNaive_counter).append("-").append(learnPostSel.numAttributes()-1);
rm.setAttributeIndices(sb.toString());
rm.setInputFormat(learnPostSel);
Instances learnPostSelNaive = rm.useFilter(learnPostSel, rm);
// TODO: make a chart here.
// bnCopy = Classifier.makeCopy(bn);
// bnCopyEv = Classifier.makeCopy(bn);
ev = new Evaluation(learnPostSelNaive);
ev.crossValidateModel(Classifier.makeCopy(bn), learnPostSelNaive,Math.min(CVfoldNum,learn.numInstances()),new Random(removeNaive_counter));
// bnCopy.buildClassifier(learnPostSelNaive); //this only gives the bayes score.. is that interesting?
accuracyRes[removeNaive_counter] = ev.pctCorrect();
ROCRes[removeNaive_counter] = ev.weightedAreaUnderROC();
}
wrep.BGS_multivariate_ROC_ACC = new double[2][learn.numAttributes()-3];
// for (int mv_counter = 0; mv_counter<accuracyRes.length-2; mv_counter++){
System.arraycopy(ROCRes, 2, wrep.BGS_multivariate_ROC_ACC[0], 0, learn.numAttributes()-3);
System.arraycopy(accuracyRes, 2, wrep.BGS_multivariate_ROC_ACC[1], 0, learn.numAttributes()-3);
// wrep.BGS_multivariate_ROC_ACC[1] = accuracyRes;
// }
winningNaiveRemove = pickWinner(accuracyRes,ROCRes);
StringBuilder sb = new StringBuilder("");
sb.append(winningNaiveRemove).append("-").append(learnPostSel.numAttributes()-1);
rm.setAttributeIndices(sb.toString());
rm.setInputFormat(learnPostSel);
Instances learnPreRandRemove = rm.useFilter(learnPostSel, rm);
Instances holdPreRandRemove = rm.useFilter(holdPostSel, rm);
wrep.BGS_midAttributeSelectionNumberOfFeatures = learnPreRandRemove.numAttributes();
ev = new Evaluation(learnPreRandRemove);
ev.crossValidateModel(Classifier.makeCopy(bn), learnPreRandRemove,Math.min(CVfoldNum,learn.numInstances()),new Random(winningNaiveRemove));
// bnCopy.buildClassifier(learnPostSelNaive); //this only gives the bayes score.. is that interesting?
double midAttACC = ev.pctCorrect();
double midAttROC = ev.weightedAreaUnderROC();
wrep.BGS_midAttributeACC = midAttACC;
wrep.BGS_midAttributeROC = midAttROC;
// random removal loop
int numCats = Math.min(10, learnPreRandRemove.numAttributes()-1);
int[] numToRem = new int[numCats];
numToRem[0]= 1;
for (int setup_counter = 1 ; setup_counter<numCats; setup_counter++){
numToRem[setup_counter] = Math.round(learnPreRandRemove.numAttributes()/numCats)*setup_counter;
}
double[][] accPostRandRemove = new double[numCats][removeIterations];
double[][] rocPostRandRemove = new double[numCats][removeIterations];
Classifier yBN = new BayesNet();
K2 yK2 = new K2();
yK2.setMaxNrOfParents(1);
yK2.setInitAsNaiveBayes(true);
SimpleEstimator ySE = new SimpleEstimator();
// ySE.setAlpha(5);
((BayesNet) yBN).setSearchAlgorithm(yK2);
((BayesNet) yBN).setEstimator(ySE);
for (int numToRem_counter = 0; numToRem_counter<numToRem.length;numToRem_counter++){
for (int iteration_counter = 0; iteration_counter < removeIterations; iteration_counter++){
Instances learnPostRandRemove = new Instances(learnPreRandRemove);
// Instances holdPostRandRemove = new Instances(holdPreRandRemove);
Random rnd = new Random(1+numToRem_counter*10000+iteration_counter*10);
// StringBuilder remString = new StringBuilder("");
for (int rmv_counter = 0 ; rmv_counter < numToRem[numToRem_counter]; rmv_counter++){
String theInt = String.valueOf(1+rnd.nextInt(learnPreRandRemove.numAttributes()-1-rmv_counter));
Remove rmv = new Remove();
// System.out.println(1+numToRem_counter*10000+iteration_counter*10);
// System.out.println(theInt);
// remString.append(theInt).append(";");
rmv.setAttributeIndices(theInt);
try{
rmv.setInputFormat(learnPostRandRemove);
learnPostRandRemove = rmv.useFilter(learnPostRandRemove, rmv);
// holdPostRandRemove = rmv.useFilter(holdPostRandRemove, rmv);
} catch (Exception ex) {System.out.println(ex.toString());}
} // rmv_counter
// now the postRandRemove instances are configured
Classifier yBNcopy = Classifier.makeCopy(yBN);
Evaluation evPostRand = new Evaluation(learnPostRandRemove);
evPostRand.crossValidateModel(yBNcopy, learnPostRandRemove,Math.min(CVfoldNum,learn.numInstances()),new Random(winningNaiveRemove));
accPostRandRemove[numToRem_counter][iteration_counter] = evPostRand.pctCorrect();
rocPostRandRemove[numToRem_counter][iteration_counter] = evPostRand.weightedAreaUnderROC();
// if ((numToRem_counter % 100)==0){System.out.print(numToRem_counter);System.out.println("RandomRemovalLoop: ".concat(String.valueOf(evPostRand.weightedAreaUnderROC())));}
}
}
//pick a winner
wrep.BGS_removes_ROC_HI_LO_AVG = new double[3][numCats];
for (int numToRem_counter = 0; numToRem_counter<numToRem.length;numToRem_counter++){
Arrays.sort(rocPostRandRemove[numToRem_counter]);
wrep.BGS_removes_ROC_HI_LO_AVG[0][numToRem_counter] = rocPostRandRemove[numToRem_counter][rocPostRandRemove[numToRem_counter].length-1];
wrep.BGS_removes_ROC_HI_LO_AVG[1][numToRem_counter] = rocPostRandRemove[numToRem_counter][0];
wrep.BGS_removes_ROC_HI_LO_AVG[2][numToRem_counter] = rocPostRandRemove[numToRem_counter][Math.round((float) Math.ceil((rocPostRandRemove[numToRem_counter].length-1)/2))];
}
int[] postRandWinnerArray = new int[numCats];
for (int numToRem_counter = 0; numToRem_counter<numToRem.length;numToRem_counter++){
postRandWinnerArray[numToRem_counter] = pickWinner(accPostRandRemove[numToRem_counter], rocPostRandRemove[numToRem_counter]);
}
// TODO: make a chart here.
double[] accWinners = new double[numCats];
double[] rocWinners = new double[numCats];
for (int numToRem_counter = 0; numToRem_counter<numToRem.length;numToRem_counter++){
accWinners[numToRem_counter] = accPostRandRemove[numToRem_counter][postRandWinnerArray[numToRem_counter]];
rocWinners[numToRem_counter] = rocPostRandRemove[numToRem_counter][postRandWinnerArray[numToRem_counter]];
}
int winningPostRandRemove = pickWinner(accWinners, rocWinners);
double[] accFinalRound = new double[removeIterations*10];
double[] rocFinalRound = new double[removeIterations*10];
double accFinal = -1;
double rocFinal = -1;
Classifier zBN = new BayesNet();
K2 zK2 = new K2();
zK2.setMaxNrOfParents(1);
zK2.setInitAsNaiveBayes(true);
SimpleEstimator zSE = new SimpleEstimator();
// ySE.setAlpha(5);
((BayesNet) zBN).setSearchAlgorithm(zK2);
((BayesNet) zBN).setEstimator(zSE);
for (int iteration_counter = 0; iteration_counter < removeIterations; iteration_counter++){
Instances learnFinalRound = new Instances(learnPreRandRemove);
Instances holdFinalRound = new Instances(holdPreRandRemove);
rndd.setSeed(iteration_counter);
StringBuilder remString = new StringBuilder("");
for (int rmv_counter = 0 ; rmv_counter < numToRem[winningPostRandRemove]; rmv_counter++){
String theInt = String.valueOf(1+rndd.nextInt(learnPreRandRemove.numAttributes()-1-rmv_counter));
Remove rmv = new Remove();
remString.append(theInt).append(";");
rmv.setAttributeIndices(theInt);
try{
rmv.setInputFormat(learnFinalRound);
learnFinalRound = rmv.useFilter(learnFinalRound, rmv);
holdFinalRound = rmv.useFilter(holdFinalRound, rmv);
} catch (Exception ex) {System.out.println(ex.toString());}
} // rmv_counter
// now the postRandRemove instances are configured