/**
* Finds the best parameter combination. (recursive for each parameter being optimised).
*
* @param depth the index of the parameter to be optimised at this level
* @param trainData the data the search is based on
* @param random a random number generator
* @throws Exception if an error occurs
*/
protected void findParamsByCrossValidation(int depth, Instances trainData, Random random)
throws Exception {
if (depth < m_CVParams.size()) {
CVParameter cvParam = (CVParameter) m_CVParams.elementAt(depth);
double upper;
switch ((int) (cvParam.m_Lower - cvParam.m_Upper + 0.5)) {
case 1:
upper = m_NumAttributes;
break;
case 2:
upper = m_TrainFoldSize;
break;
default:
upper = cvParam.m_Upper;
break;
}
double increment = (upper - cvParam.m_Lower) / (cvParam.m_Steps - 1);
for (cvParam.m_ParamValue = cvParam.m_Lower;
cvParam.m_ParamValue <= upper;
cvParam.m_ParamValue += increment) {
findParamsByCrossValidation(depth + 1, trainData, random);
}
} else {
Evaluation evaluation = new Evaluation(trainData);
// Set the classifier options
String[] options = createOptions();
if (m_Debug) {
System.err.print("Setting options for " + m_Classifier.getClass().getName() + ":");
for (int i = 0; i < options.length; i++) {
System.err.print(" " + options[i]);
}
System.err.println("");
}
((OptionHandler) m_Classifier).setOptions(options);
for (int j = 0; j < m_NumFolds; j++) {
// We want to randomize the data the same way for every
// learning scheme.
Instances train = trainData.trainCV(m_NumFolds, j, new Random(1));
Instances test = trainData.testCV(m_NumFolds, j);
m_Classifier.buildClassifier(train);
evaluation.setPriors(train);
evaluation.evaluateModel(m_Classifier, test);
}
double error = evaluation.errorRate();
if (m_Debug) {
System.err.println("Cross-validated error rate: " + Utils.doubleToString(error, 6, 4));
}
if ((m_BestPerformance == -99) || (error < m_BestPerformance)) {
m_BestPerformance = error;
m_BestClassifierOptions = createOptions();
}
}
}
/**
* Generates the classifier.
*
* @param instances set of instances serving as training data
* @throws Exception if the classifier has not been generated successfully
*/
public void buildClassifier(Instances instances) throws Exception {
// can classifier handle the data?
getCapabilities().testWithFail(instances);
// remove instances with missing class
Instances trainData = new Instances(instances);
trainData.deleteWithMissingClass();
if (!(m_Classifier instanceof OptionHandler)) {
throw new IllegalArgumentException("Base classifier should be OptionHandler.");
}
m_InitOptions = ((OptionHandler) m_Classifier).getOptions();
m_BestPerformance = -99;
m_NumAttributes = trainData.numAttributes();
Random random = new Random(m_Seed);
trainData.randomize(random);
m_TrainFoldSize = trainData.trainCV(m_NumFolds, 0).numInstances();
// Check whether there are any parameters to optimize
if (m_CVParams.size() == 0) {
m_Classifier.buildClassifier(trainData);
m_BestClassifierOptions = m_InitOptions;
return;
}
if (trainData.classAttribute().isNominal()) {
trainData.stratify(m_NumFolds);
}
m_BestClassifierOptions = null;
// Set up m_ClassifierOptions -- take getOptions() and remove
// those being optimised.
m_ClassifierOptions = ((OptionHandler) m_Classifier).getOptions();
for (int i = 0; i < m_CVParams.size(); i++) {
Utils.getOption(((CVParameter) m_CVParams.elementAt(i)).m_ParamChar, m_ClassifierOptions);
}
findParamsByCrossValidation(0, trainData, random);
String[] options = (String[]) m_BestClassifierOptions.clone();
((OptionHandler) m_Classifier).setOptions(options);
m_Classifier.buildClassifier(trainData);
}
