/**
* Recreate a tree node content with the specified data based on the tree node's existing content.
*
* @param first The start(inclusive) index of the train data used for creating the tree node
* content.
* @param last The end(exclusive) index of the train data used for creating the tree node content.
* @param node The tree node whose content need to be recreated.
* @return The recreated tree node content.
*/
private TreeNodeContent createContent(int first, int last, TreeNode node) {
// Compute the total weight and its class distribution of [first
// last) prune cases
float totalWeight = 0;
AttributeDelegate classAttributeDelegate =
this.attributeDelegates[dataSet.getClassAttributeIndex()];
float[] totalClassDistri = new float[this.dataSet.getClassCount()];
Arrays.fill(totalClassDistri, 0);
for (int i = first; i < last; i++) {
int classLabel = classAttributeDelegate.getClassBranch(this.cases[i]);
totalClassDistri[classLabel] += this.weight[cases[i]];
}
// Find the original classification of the tree node
String nodeClassification = node.getContent().getClassification();
String[] classValues = this.dataSet.getClassValues();
int maxClassIndex = indexOf(nodeClassification, classValues);
// Find the most probable classification of the prune data on
// the current tree node
for (int i = 0; i < totalClassDistri.length; i++) {
totalWeight += totalClassDistri[i];
if (totalClassDistri[i] > totalClassDistri[maxClassIndex]) maxClassIndex = i;
}
String classification = classValues[maxClassIndex];
// Estimate the leafError of the tree node with the [first last)
// prune data
float basicLeafError = totalWeight - totalClassDistri[maxClassIndex];
float extraLeafError = Estimator.getExtraError(totalWeight, basicLeafError);
float estimatedLeafError = basicLeafError + extraLeafError;
return new TreeNodeContent(totalWeight, totalClassDistri, classification, estimatedLeafError);
}
/**
* Prune the decision tree from top to bottom with EBP strategy.
*
* @param node The current tree node to be pruned
* @param first The start(inclusive) index of the train data used for pruning.
* @param last The end(exclusive) index of the train data used for pruning.
* @param update whether the current pruning is a trial to retrieve the error after pruning
* (update = false) or an actual pruning (update = true).
* @return The estimated error after completing pruning the subtree started from the current tree
* node.
*/
private float ebpPrune(TreeNode node, int first, int last, boolean update) {
TreeNodeContent content = createContent(first, last, node);
float estimatedLeafError = content.getErrorAsLeafNode();
// If this is an actual pruning instead of an error-estimation,
// reset the tree node information
if (update) node.setContent(content);
InternalNode internalNode;
// If the current tree node is a Leaf, its pruning is finished
if (node instanceof LeafNode) {
return estimatedLeafError;
} else {
internalNode = (InternalNode) node;
}
/*
* Begin to estimate the errors of each branch to get the
* errorAsInternalNode of the tree node
*/
// The estimated test error of the tree node as an InternalNode
float estimatedTreeError = 0;
// The branch index with the maximal weight distribution
int maxBranch = -1;
// The maximal branch weight
float maxBranchWeight = 0;
// The index of the test attribute on the tree node
int testAttributeIndex =
indexOf(
internalNode.getTestAttribute().getName(),
this.dataSet.getMetaData().getAttributeNames());
AttributeDelegate testAttributeDelegate = this.attributeDelegates[testAttributeIndex];
int testBranchCount = testAttributeDelegate.getBranchCount();
// Record the class weight distribution of the selected test
// attribute
float[] branchDistri = new float[testBranchCount + 1];
/*
* 'missingBegin' records the begin index of the missing data if
* there is any, otherwise it coordinates with beginIndex;
* 'groupBegin' records the begin index to group the cases for
* one branch 'nextGroupBegin' records the begin index group the
* cases for next branch
*/
int missingBegin = first;
int groupBegin = first;
// Group the missing data to the most front
if (testAttributeDelegate.hasMissingData()) {
groupBegin = testAttributeDelegate.groupForward(first, last, -1, branchDistri);
}
// Classify the [first last) cases to the branches of the test
// attribute
// except for the last branch, to construct the children tree
// nodes
for (int index = 0; index < testBranchCount; index++) {
// For a continuous attribute, the group criterion is
// cutRank;
// For a discrete attribute, the group criterion is the
// branch value(or index)
int split =
testAttributeDelegate instanceof ContinuousAttributeDelegate
? internalNode.getCutRank()
: index;
// For the first several branches, we need to group the
// specified branch values forward
// near "groupBegin" and compute its branch weight
int nextGroupBegin;
if (index < testBranchCount - 1) {
nextGroupBegin = testAttributeDelegate.groupForward(groupBegin, last, split, branchDistri);
}
// For the last branch, the "nextGroupBegin" must be last
// and its branch weight must be
// the rest weight of the total weight.
else {
nextGroupBegin = last;
float lastWeight = content.getTrainWeight();
for (int j = 0; j < branchDistri.length - 1; j++) {
lastWeight -= branchDistri[j];
}
branchDistri[branchDistri.length - 1] = lastWeight;
}
// If there is no cases distributed in this branch, omit
if (groupBegin == nextGroupBegin) {
continue;
}
// If there is missing data
else if (groupBegin > missingBegin) {
// Compute the weight ratio of this branch
float ratio = branchDistri[index + 1] / (content.getTrainWeight() - branchDistri[0]);
// split the weight of the missing data with by
// multiplying the ratio
for (int i = missingBegin; i < groupBegin; i++) this.weight[this.cases[i]] *= ratio;
// Accumulate the estimated errorAsInternalNode
estimatedTreeError +=
ebpPrune(internalNode.getChildAt(index), missingBegin, nextGroupBegin, update);
// Restore the original sequence of the cases after the
// recursive construction
missingBegin = testAttributeDelegate.groupBackward(missingBegin, nextGroupBegin);
// Restore the weight of the missing data with by
// dividing the ratio
for (int i = missingBegin; i < nextGroupBegin; i++) this.weight[this.cases[i]] /= ratio;
} else {
estimatedTreeError +=
ebpPrune(internalNode.getChildAt(index), missingBegin, nextGroupBegin, update);
// When there is no missing data, missingBegin moves
// together with groupBegin
missingBegin = nextGroupBegin;
}
// For next branch, group from nextGroupBegin index
groupBegin = nextGroupBegin;
// Select the biggest branch with maximal weight for
// branchError estimation
if (branchDistri[index + 1] > maxBranchWeight) {
maxBranchWeight = branchDistri[index + 1];
maxBranch = index;
}
}
// If this sentence is not present, it will lead to significant
// pruning!
// Do not evaluate doubled subtree raising (i.e. subtree-raising
// of subtree-raising)
if (!update) return estimatedTreeError;
// Estimate the subtree-raising error
float estimatedBranchError = ebpPrune(internalNode.getChildAt(maxBranch), first, last, false);
TreeNode parent = (InternalNode) internalNode.getParent();
// Select a strategy with the minimal error
if (estimatedLeafError <= estimatedBranchError + 0.1
&& estimatedLeafError <= estimatedTreeError + 0.1) {
LeafNode newNode = new LeafNode(content);
if (parent != null) {
int childIndex = parent.indexOfChild(internalNode);
parent.setChildAt(childIndex, newNode);
} else this.dt.setRoot(newNode);
node = newNode;
} else if (estimatedBranchError <= estimatedTreeError + 0.1) {
ebpPrune(internalNode.getChildAt(maxBranch), first, last, true);
TreeNode newNode = node.getChildAt(maxBranch);
if (parent != null) {
int childIndex = parent.indexOfChild(internalNode);
parent.setChildAt(childIndex, newNode);
} else this.dt.setRoot(newNode);
node = newNode;
} else {
node.setTrainError(estimatedTreeError);
}
return node.getTrainError();
}