bool DecisionTreeClusterNode::computeError( const ClassificationData &trainingData, MatrixFloat &data, const Vector< UINT > &classLabels, Vector< MinMax > ranges, Vector< UINT > groupIndex, const UINT featureIndex, Float &threshold, Float &error ){
error = 0;
threshold = 0;
const UINT M = trainingData.getNumSamples();
const UINT K = (UINT)classLabels.size();
Float giniIndexL = 0;
Float giniIndexR = 0;
Float weightL = 0;
Float weightR = 0;
VectorFloat groupCounter(2,0);
MatrixFloat classProbabilities(K,2);
//Use this data to train a KMeans cluster with 2 clusters
KMeans kmeans;
kmeans.setNumClusters( 2 );
kmeans.setComputeTheta( true );
kmeans.setMinChange( 1.0e-5 );
kmeans.setMinNumEpochs( 1 );
kmeans.setMaxNumEpochs( 100 );
//Disable the logging to clean things up
kmeans.setTrainingLoggingEnabled( false );
if( !kmeans.train_( data ) ){
errorLog << __GRT_LOG__ << " Failed to train KMeans model for feature: " << featureIndex << std::endl;
return false;
}
//Set the split threshold as the mid point between the two clusters
const MatrixFloat &clusters = kmeans.getClusters();
threshold = 0;
for(UINT i=0; i<clusters.getNumRows(); i++){
threshold += clusters[i][0];
}
threshold /= clusters.getNumRows();
//Iterate over each sample and work out if it should be in the lhs (0) or rhs (1) group based on the current threshold
groupCounter[0] = groupCounter[1] = 0;
classProbabilities.setAllValues(0);
for(UINT i=0; i<M; i++){
groupIndex[i] = trainingData[ i ][ featureIndex ] >= threshold ? 1 : 0;
groupCounter[ groupIndex[i] ]++;
classProbabilities[ getClassLabelIndexValue(trainingData[i].getClassLabel(),classLabels) ][ groupIndex[i] ]++;
}
//Compute the class probabilities for the lhs group and rhs group
for(UINT k=0; k<K; k++){
classProbabilities[k][0] = groupCounter[0]>0 ? classProbabilities[k][0]/groupCounter[0] : 0;
classProbabilities[k][1] = groupCounter[1]>0 ? classProbabilities[k][1]/groupCounter[1] : 0;
}
//Compute the Gini index for the lhs and rhs groups
giniIndexL = giniIndexR = 0;
for(UINT k=0; k<K; k++){
giniIndexL += classProbabilities[k][0] * (1.0-classProbabilities[k][0]);
giniIndexR += classProbabilities[k][1] * (1.0-classProbabilities[k][1]);
}
weightL = groupCounter[0]/M;
weightR = groupCounter[1]/M;
error = (giniIndexL*weightL) + (giniIndexR*weightR);
return true;
}
bool DecisionTreeThresholdNode::computeBestSplitBestRandomSplit( const UINT &numSplittingSteps, const ClassificationData &trainingData, const Vector< UINT > &features, const Vector< UINT > &classLabels, UINT &featureIndex, Float &minError ){
const UINT M = trainingData.getNumSamples();
const UINT N = (UINT)features.size();
const UINT K = (UINT)classLabels.size();
if( N == 0 ) return false;
minError = grt_numeric_limits< Float >::max();
UINT bestFeatureIndex = 0;
Float bestThreshold = 0;
Float error = 0;
Float giniIndexL = 0;
Float giniIndexR = 0;
Float weightL = 0;
Float weightR = 0;
Random random;
Vector< UINT > groupIndex(M);
VectorFloat groupCounter(2,0);
MatrixFloat classProbabilities(K,2);
//Loop over each feature and try and find the best split point
UINT m,n;
const UINT numFeatures = features.getSize();
for(m=0; m<numSplittingSteps; m++){
//Chose a random feature
n = random.getRandomNumberInt(0,numFeatures);
featureIndex = features[n];
//Randomly choose the threshold, the threshold is based on a randomly selected sample with some random scaling
threshold = trainingData[ random.getRandomNumberInt(0,M) ][ featureIndex ] * random.getRandomNumberUniform(0.8,1.2);
//Iterate over each sample and work out if it should be in the lhs (0) or rhs (1) group
groupCounter[0] = groupCounter[1] = 0;
classProbabilities.setAllValues(0);
for(UINT i=0; i<M; i++){
groupIndex[i] = trainingData[ i ][ featureIndex ] >= threshold ? 1 : 0;
groupCounter[ groupIndex[i] ]++;
classProbabilities[ getClassLabelIndexValue(trainingData[i].getClassLabel(),classLabels) ][ groupIndex[i] ]++;
}
//Compute the class probabilities for the lhs group and rhs group
for(UINT k=0; k<K; k++){
classProbabilities[k][0] = groupCounter[0]>0 ? classProbabilities[k][0]/groupCounter[0] : 0;
classProbabilities[k][1] = groupCounter[1]>0 ? classProbabilities[k][1]/groupCounter[1] : 0;
}
//Compute the Gini index for the lhs and rhs groups
giniIndexL = giniIndexR = 0;
for(UINT k=0; k<K; k++){
giniIndexL += classProbabilities[k][0] * (1.0-classProbabilities[k][0]);
giniIndexR += classProbabilities[k][1] * (1.0-classProbabilities[k][1]);
}
weightL = groupCounter[0]/M;
weightR = groupCounter[1]/M;
error = (giniIndexL*weightL) + (giniIndexR*weightR);
//Store the best threshold and feature index
if( error < minError ){
minError = error;
bestThreshold = threshold;
bestFeatureIndex = featureIndex;
}
}
//Set the best feature index that will be returned to the DecisionTree that called this function
featureIndex = bestFeatureIndex;
//Store the node size, feature index, best threshold and class probabilities for this node
set(M,featureIndex,bestThreshold,trainingData.getClassProbabilities(classLabels));
return true;
}
bool RegressionTree::computeBestSpiltBestIterativeSpilt( const RegressionData &trainingData, const Vector< UINT > &features, UINT &featureIndex, Float &threshold, Float &minError ){
const UINT M = trainingData.getNumSamples();
const UINT N = (UINT)features.size();
if( N == 0 ) return false;
minError = grt_numeric_limits< Float >::max();
UINT bestFeatureIndex = 0;
UINT groupID = 0;
Float bestThreshold = 0;
Float error = 0;
Float minRange = 0;
Float maxRange = 0;
Float step = 0;
Vector< UINT > groupIndex(M);
VectorFloat groupCounter(2,0);
VectorFloat groupMean(2,0);
VectorFloat groupMSE(2,0);
Vector< MinMax > ranges = trainingData.getInputRanges();
//Loop over each feature and try and find the best split point
for(UINT n=0; n<N; n++){
minRange = ranges[n].minValue;
maxRange = ranges[n].maxValue;
step = (maxRange-minRange)/Float(numSplittingSteps);
threshold = minRange;
featureIndex = features[n];
while( threshold <= maxRange ){
//Iterate over each sample and work out what group it falls into
for(UINT i=0; i<M; i++){
groupID = trainingData[i].getInputVector()[featureIndex] >= threshold ? 1 : 0;
groupIndex[i] = groupID;
groupMean[ groupID ] += trainingData[i].getInputVector()[featureIndex];
groupCounter[ groupID ]++;
}
groupMean[0] /= groupCounter[0] > 0 ? groupCounter[0] : 1;
groupMean[1] /= groupCounter[1] > 0 ? groupCounter[1] : 1;
//Compute the MSE for each group
for(UINT i=0; i<M; i++){
groupMSE[ groupIndex[i] ] += grt_sqr( groupMean[ groupIndex[i] ] - trainingData[ i ].getInputVector()[features[n]] );
}
groupMSE[0] /= groupCounter[0] > 0 ? groupCounter[0] : 1;
groupMSE[1] /= groupCounter[1] > 0 ? groupCounter[1] : 1;
error = sqrt( groupMSE[0] + groupMSE[1] );
//Store the best threshold and feature index
if( error < minError ){
minError = error;
bestThreshold = threshold;
bestFeatureIndex = featureIndex;
}
//Update the threshold
threshold += step;
}
}
//Set the best feature index and threshold
featureIndex = bestFeatureIndex;
threshold = bestThreshold;
return true;
}
bool DecisionTreeThresholdNode::computeBestSplitBestIterativeSplit( const UINT &numSplittingSteps, const ClassificationData &trainingData, const Vector< UINT > &features, const Vector< UINT > &classLabels, UINT &featureIndex, Float &minError ){
const UINT M = trainingData.getNumSamples();
const UINT N = features.getSize();
const UINT K = classLabels.getSize();
if( N == 0 ) return false;
minError = grt_numeric_limits< Float >::max();
UINT bestFeatureIndex = 0;
Float bestThreshold = 0;
Float error = 0;
Float minRange = 0;
Float maxRange = 0;
Float step = 0;
Float giniIndexL = 0;
Float giniIndexR = 0;
Float weightL = 0;
Float weightR = 0;
Vector< UINT > groupIndex(M);
VectorFloat groupCounter(2,0);
Vector< MinMax > ranges = trainingData.getRanges();
MatrixFloat classProbabilities(K,2);
//Loop over each feature and try and find the best split point
for(UINT n=0; n<N; n++){
minRange = ranges[n].minValue;
maxRange = ranges[n].maxValue;
step = (maxRange-minRange)/Float(numSplittingSteps);
threshold = minRange;
featureIndex = features[n];
while( threshold <= maxRange ){
//Iterate over each sample and work out if it should be in the lhs (0) or rhs (1) group
groupCounter[0] = groupCounter[1] = 0;
classProbabilities.setAllValues(0);
for(UINT i=0; i<M; i++){
groupIndex[i] = trainingData[ i ][ featureIndex ] >= threshold ? 1 : 0;
groupCounter[ groupIndex[i] ]++;
classProbabilities[ getClassLabelIndexValue(trainingData[i].getClassLabel(),classLabels) ][ groupIndex[i] ]++;
}
//Compute the class probabilities for the lhs group and rhs group
for(UINT k=0; k<K; k++){
classProbabilities[k][0] = groupCounter[0]>0 ? classProbabilities[k][0]/groupCounter[0] : 0;
classProbabilities[k][1] = groupCounter[1]>0 ? classProbabilities[k][1]/groupCounter[1] : 0;
}
//Compute the Gini index for the lhs and rhs groups
giniIndexL = giniIndexR = 0;
for(UINT k=0; k<K; k++){
giniIndexL += classProbabilities[k][0] * (1.0-classProbabilities[k][0]);
giniIndexR += classProbabilities[k][1] * (1.0-classProbabilities[k][1]);
}
weightL = groupCounter[0]/M;
weightR = groupCounter[1]/M;
error = (giniIndexL*weightL) + (giniIndexR*weightR);
//Store the best threshold and feature index
if( error < minError ){
minError = error;
bestThreshold = threshold;
bestFeatureIndex = featureIndex;
}
//Update the threshold
threshold += step;
}
}
//Set the best feature index that will be returned to the DecisionTree that called this function
featureIndex = bestFeatureIndex;
//Store the node size, feature index, best threshold and class probabilities for this node
set(M,featureIndex,bestThreshold,trainingData.getClassProbabilities(classLabels));
return true;
}
