public Split getBestSplit(ExampleSet inputSet, Attribute attribute, String labelName) {
SortedExampleSet exampleSet =
new SortedExampleSet((ExampleSet) inputSet.clone(), attribute, SortedExampleSet.INCREASING);
Attribute labelAttribute = exampleSet.getAttributes().getLabel();
int labelIndex = labelAttribute.getMapping().mapString(labelName);
double oldLabel = Double.NaN;
double bestSplit = Double.NaN;
double lastValue = Double.NaN;
double bestBenefit = Double.NEGATIVE_INFINITY;
double bestTotalWeight = 0;
int bestSplitType = Split.LESS_SPLIT;
// initiating online counting of benefit: only 2 Datascans needed then
criterion.reinitOnlineCounting(exampleSet);
for (Example e : exampleSet) {
double currentValue = e.getValue(attribute);
double label = e.getValue(labelAttribute);
if ((Double.isNaN(oldLabel)) || (oldLabel != label) && (lastValue != currentValue)) {
double splitValue = (lastValue + currentValue) / 2.0d;
double[] benefits;
if (labelName == null) {
benefits = criterion.getOnlineBenefit(e);
} else {
benefits = criterion.getOnlineBenefit(e, labelIndex);
}
// online method returns both possible relations in one array(greater / smaller) in one
// array
if ((benefits[0] > minValue)
&& (benefits[0] > 0)
&& (benefits[1] > 0)
&& ((benefits[0] > bestBenefit)
|| ((benefits[0] == bestBenefit) && (benefits[1] > bestTotalWeight)))) {
bestBenefit = benefits[0];
bestSplit = splitValue;
bestTotalWeight = benefits[1];
bestSplitType = Split.LESS_SPLIT;
}
if ((benefits[2] > minValue)
&& (benefits[2] > 0)
&& (benefits[3] > 0)
&& ((benefits[2] > bestBenefit)
|| ((benefits[2] == bestBenefit) && (benefits[3] > bestTotalWeight)))) {
bestBenefit = benefits[2];
bestSplit = splitValue;
bestTotalWeight = benefits[3];
bestSplitType = Split.GREATER_SPLIT;
}
oldLabel = label;
}
lastValue = currentValue;
criterion.update(e);
}
return new Split(bestSplit, new double[] {bestBenefit, bestTotalWeight}, bestSplitType);
}
