private AttributeSelection AttributeSelectionFilter(int opcao)
{
AttributeSelection filter = new AttributeSelection();
// CfSubsetEval e BestFirst
if (opcao == 1)
{
weka.attributeSelection.CfsSubsetEval evaluator = new weka.attributeSelection.CfsSubsetEval();
filter.setEvaluator(evaluator);
weka.attributeSelection.BestFirst search = new weka.attributeSelection.BestFirst();
filter.setSearch(search);
}
// InfoGainAttributeEval e Ranker
else if (opcao == 2)
{
// Evaluator
weka.attributeSelection.InfoGainAttributeEval evaluator = new weka.attributeSelection.InfoGainAttributeEval();
//evaluator.setMissingSeparate(true);
filter.setEvaluator(evaluator);
// Search strategy: best first (default values)
weka.attributeSelection.Ranker search = new weka.attributeSelection.Ranker();
search.setThreshold(0);
filter.setSearch(search);
}
return(filter);
}
private int[] ReduceByGreedySupervised(weka.core.Instances insts)
{
int[] rang = null;
AttributeSelection filter = new AttributeSelection(); // package weka.filters.supervised.attribute!
CfsSubsetEval eval = new CfsSubsetEval();
GreedyStepwise search = new GreedyStepwise();
search.setSearchBackwards(true);
filter.setEvaluator(eval);
filter.setSearch(search);
filter.SelectAttributes(insts);
rang = filter.selectedAttributes();
return rang;
}
//public static T[] SubArray<T>(this T[] data, int index, int length)
//{
// T[] result = new T[length];
// Array.Copy(data, index, result, 0, length);
// return result;
//}
public static void useLowLevelInformationGainFeatureSelection(Instances data)
{
AttributeSelection attsel = new AttributeSelection();
InfoGainAttributeEval eval = new InfoGainAttributeEval();
Ranker search = new Ranker();
//1000 features >0, should be equal to -1 features and threshold 0.0. anyway i use 0.01.
//int numtoselect = 1000;// 1520;
//search.setThreshold(-1.7976931348623157E308d);
int numtoselect = 0;
float threshold = 0;
if (GuiPreferences.Instance.IgSelectionType == IGType.Threshold)
{
numtoselect = -1;
threshold = Convert.ToSingle(GuiPreferences.Instance.NudIGThreshold);
GuiPreferences.Instance.setLog("Filtering using IG threshold: " + GuiPreferences.Instance.NudIGThreshold.ToString());
}
else if (GuiPreferences.Instance.IgSelectionType == IGType.Voxels)
{
//num of vox plus above 0 cut off.
numtoselect = Convert.ToInt32(GuiPreferences.Instance.NudIGVoxelAmount);
//search.setThreshold(-1.7976931348623157E308d);
GuiPreferences.Instance.setLog("Filtering using IG Voxel Amount of: " + GuiPreferences.Instance.NudIGVoxelAmount.ToString());
}
else
{
GuiPreferences.Instance.setLog("error wrong IG type");
}
search.setNumToSelect(numtoselect);
search.setThreshold(threshold);
attsel.setEvaluator(eval);
attsel.setSearch(search);
attsel.SelectAttributes(data);
//reeturned back to the global instance
Preferences.Instance.attsel = attsel;
//hIstogram saving indices and ranked to preferences for easy access
SortedDictionary <double, int> Histogram = new SortedDictionary <double, int>();
double[][] blah = attsel.rankedAttributes();
for (double i = -0.05; i < 1.2; i += 0.05)
{
if (!Histogram.ContainsKey(i))
{
Histogram.Add(i, 0);
}
for (int j = 0; j < blah.Length; j++)
{
if (blah[j][1] > i - 0.05 && blah[j][1] <= i)
{
Histogram[i] += 1;
}
}
}
GuiPreferences.Instance.setLog("Histogram:");
for (double i = -0.05; i < 1.2; i += 0.05)
{
GuiPreferences.Instance.setLog("Threshold: " + i.ToString() + ": " + Histogram[i].ToString());
}
//--------------
if (GuiPreferences.Instance.IgSelectionType == IGType.Voxels)
{
//SELECT K BIGGER THAN ZERO.
int IgVoxAboveZero = 0;
for (int j = 0; j < GuiPreferences.Instance.NudIGVoxelAmount; j++)
{
if (blah[j][1] > 0)
{
IgVoxAboveZero++;
}
}
//this is a bit redundant, to replace this redundancy, create two vectors, selected() & ranked()
search.setNumToSelect(IgVoxAboveZero);
search.setThreshold(threshold);
attsel.setEvaluator(eval);
attsel.setSearch(search);
attsel.SelectAttributes(data);
Preferences.Instance.attsel = attsel;
GuiPreferences.Instance.NudIGVoxelAmount = IgVoxAboveZero;
GuiPreferences.Instance.setLog("Filtering using IG Voxel Amount of (above zero): " + GuiPreferences.Instance.NudIGVoxelAmount.ToString());
GuiPreferences.Instance.setLog("Changing NudIGVoxelAmount to the above figure!");
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
///
//NOTE: uncommenting this proves that ranked attributes are 0-based. this sorts them out from 0 to 204799 (Also threshold above should be 0
/*double[][] blah = attsel.rankedAttributes();
* double[] blat = new double[blah.Length];
* for (int i = 0; i < blah.Length;i++ )
* {
* blat[i] = blah[i][0];
* }
* Array.Sort(blat);*/
//////////////////////////////////////////////////////////////////////////////////////////////////////
//if this code is used it will need to be ammended for numtoselect which is -1 (All) in this case
/*double[][] ranked = attsel.rankedAttributes();
* int indexBiggerThanZero = 0;
* int classColumn = indices[indices.Length - 1];
* while ((ranked[indexBiggerThanZero][1] > threshold) && (indexBiggerThanZero < numtoselect))
* indexBiggerThanZero++;
*
* //return K features whose IG value is bigger than zero
* int[] indicesFinal = new int[] {};
* //less than K features, we dynamically resize an array and return it + class
* if (indexBiggerThanZero < numtoselect)
* {
* Array.Resize(ref indicesFinal, indexBiggerThanZero+1);
* Array.Copy(indices, 0, indicesFinal, 0, indexBiggerThanZero);
*
* //Array.Copy(indices, 0, indices, 0, indices.Length+1);
* indicesFinal[indicesFinal.Length - 1] = classColumn;
*
* return indicesFinal;
* }
* else
* {
* //if indexBiggerThanZero is the same, all features ig values are aboce zero, we return the original int array
* return indices;
* }*/
}
