//Compute the regression data that will be stored at this node
bool RegressionTree::computeNodeRegressionData( const RegressionData &trainingData, VectorDouble ®ressionData ){
const UINT M = trainingData.getNumSamples();
const UINT N = trainingData.getNumInputDimensions();
const UINT T = trainingData.getNumTargetDimensions();
if( M == 0 ){
Regressifier::errorLog << "computeNodeRegressionData(...) - Failed to compute regression data, there are zero training samples!" << endl;
return false;
}
//Make sure the regression data is the correct size
regressionData.clear();
regressionData.resize( T, 0 );
//The regression data at this node is simply an average over all the training data at this node
for(unsigned int j=0; j<N; j++){
for(unsigned int i=0; i<M; i++){
regressionData[j] += trainingData[i].getTargetVector()[j];
}
regressionData[j] /= M;
}
return true;
}
void ClassificationBiasMatrix::PerformAdjustmnts (const VectorDouble& classifiedCounts,
VectorDouble& adjCounts,
VectorDouble& stdErrors
)
{
// For description of calc's see the paper:
// "Estimating the Taxonomic composition of a sample when individuals are classified with error"
// by Andrew Solow, Cabll Davis, Qiao Hu
// Woods Hole Oceanographic Institution, Woods Hole Massachusetts
// Marine Ecology Progress Series
// published 2006-july-06; vol 216:309-311
if (classifiedCounts.size () != (kkuint32)numClasses)
{
KKStr errMsg = "ClassificationBiasMatrix::PerformAdjustmnts ***ERROR*** Disagreement in length of classifiedCounts[" +
StrFormatInt ((kkint32)classifiedCounts.size (), "ZZZ0") +
"] and Prev Defined ClassList[" + StrFormatInt (numClasses, "ZZZ0") + "].";
runLog.Level (-1) << errMsg << endl;
valid = false;
throw KKException (errMsg);
}
kkint32 x = 0;
kkint32 i, j, k;
// We need to deal with the special case when one entry in the probability diagonal is zero.
{
for (x = 0; x < numClasses; x++)
{
if ((*probabilities)[x][x] == 0.0)
{
// This will cause the inversion of the diagonal matrix to fail. To deal
// with this situation; I will steal some probability from other buckets on
// same row.
double totalAmtStolen = 0.0;
double percentToSteal = 0.01;
for (i = 0; i < numClasses; i++)
{
if ((*probabilities)[x][i] != 0.0)
{
double amtToSteal = (*probabilities)[x][i] * percentToSteal;
(*probabilities)[x][i] = (*probabilities)[x][i] - amtToSteal;
totalAmtStolen += amtToSteal;
}
}
(*probabilities)[x][x] = totalAmtStolen;
}
}
}
Matrix m (numClasses, 1);
for (x = 0; x < numClasses; x++)
m[x][0] = classifiedCounts[x];
Matrix transposed = probabilities->Transpose ();
Matrix Q = transposed.Inverse ();
Matrix n = Q * m;
Matrix varM (numClasses, numClasses);
for (j = 0; j < numClasses; j++)
{
double varM_j = 0.0;
for (i = 0; i < numClasses; i++)
{
double p = (*probabilities)[i][j];
varM_j += n[i][0] * p * (1.0 - p);
}
varM[j][j] = varM_j;
}
for (j = 0; j < numClasses; j++)
{
for (k = 0; k < numClasses; k++)
{
if (j != k)
{
double covM_jk = 0.0;
for (i = 0; i < numClasses; i++)
covM_jk -= n[i][0] * (*probabilities)[i][j] * (*probabilities)[j][k];
varM[j][k] = covM_jk;
}
}
}
Matrix varN = Q * varM * Q.Transpose ();
adjCounts.clear ();
stdErrors.clear ();
for (x = 0; x < numClasses; x++)
{
adjCounts.push_back (n[x][0]);
stdErrors.push_back (sqrt (varN[x][x]));
}
return;
} /* PerformAdjustmnts */
