public double[][] GetPrincipleFeatures(double[][] trainingData,
string[] attributeHeaders,
ref double[] standardDeviations,
ref double[][] rotationMatrix,
int maxParallelThreads = -1,
int indexTargetAttribute = -1)
{
if (trainingData.Length !=
attributeHeaders.Length)
{
throw new AttributesCountMismatchException();
}
bool ignoreLastColumn = indexTargetAttribute >= 0?true:false;
double[][] zeroCenteredData = Dispersion.GetZeroCenteredData(trainingData,
ignoreLastColumn,
maxParallelThreads);
if (Rank > zeroCenteredData.Length)
{
Rank = zeroCenteredData.Length;
}
//Generate Covariance Matrix from NewData
double[][] coVarMatrix = Dispersion.CovarianceMatrixSample(zeroCenteredData, maxParallelThreads, true);
//Compute EigenValue/Vectors of coVarMatrix
double[][] eigenVectors = null;
double[] eigenValues = null;
MatrixOperations mo = new MatrixOperations();
mo.EigenValues(coVarMatrix, ref eigenValues,
ref eigenVectors);
//Form Feature Vector
//Sort indexes by EigenValues
//List<int> A = new List<int>() { 3, 2, 1 };
var sorted = eigenValues
.Select((x, i) => new KeyValuePair <double, int>(x, i))
.OrderByDescending(x => x.Key)
.ToList();
//List<int> B = sorted.Select(x => x.Key).ToList();
List <int> idxSorted = sorted.Select(x => x.Value).ToList();
//Convert SD Matrix
List <double> standardDevList = new List <double>();
for (int idx = 0; idx < Rank; idx++)
{
if (eigenValues[idxSorted[idx]] >= 0)
{
double sd = System.Math.Sqrt(eigenValues[idxSorted[idx]]);
if (sd >= Tolerance)
{
standardDevList.Add(sd);
}
}
}
standardDeviations = standardDevList.ToArray();
//EigenValues are already sorted in decreasing order
//Hence idxSorted[0] already has highest eigenvalue/std dev
rotationMatrix = new double[standardDeviations.Length][];
for (int idx = 0; idx < standardDeviations.Length; idx++)
{
rotationMatrix[idx] = eigenVectors[idxSorted[idx]];
}
//Derive New Data Set
double[][] finalPCAData = mo.Multiply(mo.Transpose(rotationMatrix),
mo.Transpose(zeroCenteredData));
finalPCAData = mo.Transpose(finalPCAData);
return(finalPCAData);
}