public boolean tryToExpand(double splitConfidence, double tieThreshold) {
// splitConfidence. Hoeffding Bound test parameter.
// tieThreshold. Hoeffding Bound test parameter.
// SplitCriterion splitCriterion = new SDRSplitCriterionAMRules();
// SplitCriterion splitCriterion = new SDRSplitCriterionAMRulesNode();//JD for assessing only
// best branch
AMRulesSplitCriterion splitCriterion =
(AMRulesSplitCriterion)
((AMRulesSplitCriterion)
((AMRulesRegressorOld) this.amRules)
.splitCriterionOption.getPreMaterializedObject())
.copy();
// Using this criterion, find the best split per attribute and rank the results
AttributeSplitSuggestion[] bestSplitSuggestions = this.getBestSplitSuggestions(splitCriterion);
Arrays.sort(bestSplitSuggestions);
// Declare a variable to determine if any of the splits should be performed
boolean shouldSplit = false;
// If only one split was returned, use it
if (bestSplitSuggestions.length < 2) {
shouldSplit = ((bestSplitSuggestions.length > 0) && (bestSplitSuggestions[0].merit > 0));
bestSuggestion = bestSplitSuggestions[bestSplitSuggestions.length - 1];
} // Otherwise, consider which of the splits proposed may be worth trying
else {
// Determine the hoeffding bound value, used to select how many instances should be used to
// make a test decision
// to feel reasonably confident that the test chosen by this sample is the same as what would
// be chosen using infinite examples
double hoeffdingBound = computeHoeffdingBound(1, splitConfidence, getWeightSeen());
debug("Hoeffding bound " + hoeffdingBound, 4);
// Determine the top two ranked splitting suggestions
bestSuggestion = bestSplitSuggestions[bestSplitSuggestions.length - 1];
AttributeSplitSuggestion secondBestSuggestion =
bestSplitSuggestions[bestSplitSuggestions.length - 2];
debug("Merits: " + secondBestSuggestion.merit + " " + bestSuggestion.merit, 4);
// If the upper bound of the sample mean for the ratio of SDR(best suggestion) to SDR(second
// best suggestion),
// as determined using the hoeffding bound, is less than 1, then the true mean is also less
// than 1, and thus at this
// particular moment of observation the bestSuggestion is indeed the best split option with
// confidence 1-delta, and
// splitting should occur.
// Alternatively, if two or more splits are very similar or identical in terms of their
// splits, then a threshold limit
// (default 0.05) is applied to the hoeffding bound; if the hoeffding bound is smaller than
// this limit then the two
// competing attributes are equally good, and the split will be made on the one with the
// higher SDR value.
if (bestSuggestion.merit > 0) {
// if ((((secondBestSuggestion.merit / bestSuggestion.merit) + hoeffdingBound) < 1) //ratio
if ((((bestSuggestion.merit - secondBestSuggestion.merit))
> hoeffdingBound) // if normalized
|| (hoeffdingBound < tieThreshold)) {
debug("Expanded ", 5);
shouldSplit = true;
}
}
}
if (shouldSplit == true) {
AttributeSplitSuggestion splitDecision =
bestSplitSuggestions[bestSplitSuggestions.length - 1];
double minValue = Double.MAX_VALUE;
double[] branchMerits =
splitCriterion.computeBranchSplitMerits(bestSuggestion.resultingClassDistributions);
for (int i = 0; i < bestSuggestion.numSplits(); i++) {
double value = branchMerits[i];
if (value < minValue) {
minValue = value;
splitIndex = i;
statisticsNewRuleActiveLearningNode =
bestSuggestion.resultingClassDistributionFromSplit(i);
}
}
statisticsBranchSplit = splitDecision.resultingClassDistributionFromSplit(splitIndex);
statisticsOtherBranchSplit =
bestSuggestion.resultingClassDistributionFromSplit(splitIndex == 0 ? 1 : 0);
}
return shouldSplit;
}
@Override
protected void attemptToSplit(ActiveLearningNode node, SplitNode parent, int parentIndex) {
if (!node.observedClassDistributionIsPure()) {
SplitCriterion splitCriterion =
(SplitCriterion) getPreparedClassOption(this.splitCriterionOption);
AttributeSplitSuggestion[] bestSplitSuggestions =
node.getBestSplitSuggestions(splitCriterion, this);
Arrays.sort(bestSplitSuggestions);
boolean shouldSplit = false;
if (bestSplitSuggestions.length < 2) {
shouldSplit = bestSplitSuggestions.length > 0;
} else {
double hoeffdingBound =
computeHoeffdingBound(
splitCriterion.getRangeOfMerit(node.getObservedClassDistribution()),
this.splitConfidenceOption.getValue(),
node.getWeightSeen());
AttributeSplitSuggestion bestSuggestion =
bestSplitSuggestions[bestSplitSuggestions.length - 1];
AttributeSplitSuggestion secondBestSuggestion =
bestSplitSuggestions[bestSplitSuggestions.length - 2];
if ((bestSuggestion.merit - secondBestSuggestion.merit > hoeffdingBound)
|| (hoeffdingBound < this.tieThresholdOption.getValue())) {
shouldSplit = true;
}
// }
if ((this.removePoorAttsOption != null) && this.removePoorAttsOption.isSet()) {
Set<Integer> poorAtts = new HashSet<Integer>();
// scan 1 - add any poor to set
for (int i = 0; i < bestSplitSuggestions.length; i++) {
if (bestSplitSuggestions[i].splitTest != null) {
int[] splitAtts = bestSplitSuggestions[i].splitTest.getAttsTestDependsOn();
if (splitAtts.length == 1) {
if (bestSuggestion.merit - bestSplitSuggestions[i].merit > hoeffdingBound) {
poorAtts.add(new Integer(splitAtts[0]));
}
}
}
}
// scan 2 - remove good ones from set
for (int i = 0; i < bestSplitSuggestions.length; i++) {
if (bestSplitSuggestions[i].splitTest != null) {
int[] splitAtts = bestSplitSuggestions[i].splitTest.getAttsTestDependsOn();
if (splitAtts.length == 1) {
if (bestSuggestion.merit - bestSplitSuggestions[i].merit < hoeffdingBound) {
poorAtts.remove(new Integer(splitAtts[0]));
}
}
}
}
for (int poorAtt : poorAtts) {
node.disableAttribute(poorAtt);
}
}
}
if (shouldSplit) {
AttributeSplitSuggestion splitDecision =
bestSplitSuggestions[bestSplitSuggestions.length - 1];
if (splitDecision.splitTest == null) {
// preprune - null wins
deactivateLearningNode(node, parent, parentIndex);
} else {
SplitNode newSplit =
newSplitNode(splitDecision.splitTest, node.getObservedClassDistribution());
for (int i = 0; i < splitDecision.numSplits(); i++) {
// Unique Change of HoeffdingTree
Node newChild =
newLearningNode(
splitDecision.resultingClassDistributionFromSplit(i),
((LearningNodeClassifier) node).getClassifier());
newSplit.setChild(i, newChild);
}
this.activeLeafNodeCount--;
this.decisionNodeCount++;
this.activeLeafNodeCount += splitDecision.numSplits();
if (parent == null) {
this.treeRoot = newSplit;
} else {
parent.setChild(parentIndex, newSplit);
}
}
// manage memory
enforceTrackerLimit();
}
}
}
