/* private void recursiveBSP_WaveletsByConsts(List<GeoWave> geoWaveArr, int geoWaveId, int seed=0)
* {
* //CALC APPROX_SOLUTION FOR GEO WAVE
* double error = geoWaveArr[geoWaveId].calc_MeanValueReturnError(_trainingLabel, geoWaveArr[geoWaveId].pointsIdArray);
* if (error < _rc.approxThresh ||
* geoWaveArr[geoWaveId].pointsIdArray.Count() <= _rc.minWaveSize ||
* _rc.boundDepthTree <= geoWaveArr[geoWaveId].level)
* return;
*
* int dimIndex = -1;
* int Maingridindex = -1;
*
* bool IsPartitionOK = false;
* switch (_rc.split_type)
* {
* case 0:
* IsPartitionOK = getBestPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, _dime2Take);
* break;
* case 1:
* IsPartitionOK = getRandPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, seed);
* break;
* case 2:
* {
* Random ran1 = new Random(seed);
* Random ran2 = new Random(geoWaveId);
* int one = ran1.Next(0, int.MaxValue / 10);
* int two = ran2.Next(0, int.MaxValue / 10);
* bool[] Dim2TakeNode = getDim2Take(_rc, one + two);
* IsPartitionOK = getBestPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, Dim2TakeNode);
* }
* break;
* case 3:
* IsPartitionOK = getGiniPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, _dime2Take);
* break;
* case 4:
* {
* Random ran1 = new Random(seed);
* Random ran2 = new Random(geoWaveId);
* int one = ran1.Next(0, int.MaxValue / 10);
* int two = ran2.Next(0, int.MaxValue / 10);
* bool[] Dim2TakeNode = getDim2Take(_rc, one + two);
* IsPartitionOK = getGiniPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, Dim2TakeNode);
* }
* break;
*
*
* }
*
*
*
*
* if (!IsPartitionOK)
* return;
*
*
* GeoWave child0 = new GeoWave(geoWaveArr[geoWaveId].boubdingBox, _trainingLabel[0].Count(), geoWaveArr[geoWaveId].rc);
* GeoWave child1 = new GeoWave(geoWaveArr[geoWaveId].boubdingBox, _trainingLabel[0].Count(), geoWaveArr[geoWaveId].rc);
*
* //set partition
* child0.boubdingBox[1][dimIndex] = Maingridindex;
* child1.boubdingBox[0][dimIndex] = Maingridindex;
*
* //DOCUMENT ON CHILDREN
* child0.dimIndex = dimIndex;
* child0.Maingridindex = Maingridindex;
* child1.dimIndex = dimIndex;
* child1.Maingridindex = Maingridindex;
*
* child0.MaingridValue = Form1.MainGrid[dimIndex][Maingridindex];
* child1.MaingridValue = Form1.MainGrid[dimIndex][Maingridindex];
*
* //calc norm
* //calc mean value
*
* if (Form1.isBoxSingular(child0.boubdingBox, _rc.dim) || Form1.isBoxSingular(child1.boubdingBox, _rc.dim))
* return;
*
* //SHOULD I VERIFY THAT THE CHILD IS NOT ITS PARENT ? (IN CASES WHERE CAN'T MODEFY THE PARTITION)
*
* setChildrensPointsAndMeanValue(ref child0, ref child1, dimIndex, geoWaveArr[geoWaveId].pointsIdArray);
* //SET TWO CHILDS
* child0.parentID = child1.parentID = geoWaveId;
* child0.child0 = child1.child0 = -1;
* child0.child1 = child1.child1 = -1;
* child0.level = child1.level = geoWaveArr[geoWaveId].level + 1;
*
* child0.computeNormOfConsts(geoWaveArr[geoWaveId]);
* child1.computeNormOfConsts(geoWaveArr[geoWaveId]);
* geoWaveArr.Add(child0);
* geoWaveArr.Add(child1);
* geoWaveArr[geoWaveId].child0 = geoWaveArr.Count - 2;
* geoWaveArr[geoWaveId].child1 = geoWaveArr.Count - 1;
*
*
*
*
* //RECURSION STEP !!!
* recursiveBSP_WaveletsByConsts(geoWaveArr, geoWaveArr[geoWaveId].child0, seed);
* recursiveBSP_WaveletsByConsts(geoWaveArr, geoWaveArr[geoWaveId].child1, seed);
* }
*/
private SplitProps getTransformedPartitionAllDim(GeoWave parentNode, double error, SplitType splitType)
{
double[][] originalNodeData = parentNode.pointsIdArray.Select(id => _trainingDt[id]).ToArray();
double[][] transformedData;
//clean columns of categorical variables 2m0rr0w2
// originalNodeData = Helpers.copyAndRemoveCategoricalColumns(originalNodeData, _rc);
//result struct
SplitProps resultProps = new SplitProps();
switch (splitType)
{
case SplitType.LocalPca:
DimReduction.constructNodePcaByOriginalData(originalNodeData, parentNode);
transformedData = parentNode.localPca.Transform(originalNodeData);
break;
case SplitType.DiffMaps5Percent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.05);
break;
case SplitType.DiffMaps1Percent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.01);
break;
case SplitType.DiffMapsHalfPercent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.005);
break;
case SplitType.MainAxes:
transformedData = originalNodeData;
break;
case SplitType.Categorical:
transformedData = originalNodeData;
break;
default:
transformedData = null;
break;
}
if (transformedData == null)
{
//throw new Exception("******TRANSFORMATION ERROR!!!");
resultProps.isPartitionOk = false;
Debug.WriteLine("*********Failed transformation");
Debug.WriteLine("*********Failed node size: " + parentNode.pointsIdArray.Count);
return(resultProps);
}
parentNode.transformedDim = transformedData.First().Length;
//save dim of transformed data
int transformedDim = parentNode.transformedDim;
double[] errorEachDim = new double[transformedDim];
int[] partitionIdEachDim = new int[transformedDim];
// _rc.dim replaced by transformedDim dimention
Helpers.applyFor(0, transformedDim, dim =>
{
errorEachDim[dim] = getTransformedDataPartitionSingleDim(dim, transformedData, parentNode, partitionIdEachDim);
});
int bestDim = Enumerable.Range(0, transformedDim)
.Aggregate((a, b) => (errorEachDim[a] < errorEachDim[b]) ? a : b);
resultProps.splitId = partitionIdEachDim[bestDim];
//save id's order in transformed data at best dimention
resultProps.sortedIds = new List <int>(parentNode.pointsIdArray); // will be sorted at best split dimention
List <int> idsClone = new List <int>(resultProps.sortedIds); // id's in original position
resultProps.sortedIds.Sort((c1, c2) =>
transformedData[idsClone.IndexOf(c1)][bestDim].CompareTo(transformedData[idsClone.IndexOf(c2)][bestDim]));
//save partition value
int originalSplitLocation = idsClone.IndexOf(resultProps.splitId);
if (originalSplitLocation == -1)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
resultProps.isPartitionOk = (errorEachDim[bestDim] < error);
resultProps.splitValue = transformedData[originalSplitLocation][bestDim];
resultProps.error = errorEachDim[bestDim];
resultProps.type = splitType;
resultProps.dimIndex = bestDim;
//shift dimention if it was not categorical split 2m0rr0w2
/* foreach (int categoricalInd in _rc.indOfCategorical)
* {
* resultProps.dimIndex = (resultProps.dimIndex == categoricalInd)
* ? resultProps.dimIndex++
* : resultProps.dimIndex;
* }*/
return(resultProps);
}
// Transformed data decomposition
private void recursiveBSP_TransformedData(IList <GeoWave> geoWaveArr, int geoWaveId, List <SplitType> splitTypes)
{
GeoWave parentNode = geoWaveArr[geoWaveId];
double error = parentNode.calc_MeanValueReturnError(_trainingLabel, parentNode.pointsIdArray);
if (error < _rc.approxThresh ||
parentNode.pointsIdArray.Count() <= _rc.minWaveSize ||
_rc.boundDepthTree <= parentNode.level)
{
return;
}
List <SplitProps> resultSplitsProperties = (from splitType in splitTypes
select getTransformedPartitionAllDim(parentNode, error, splitType)).ToList();
resultSplitsProperties = resultSplitsProperties.Where(x => x.isPartitionOk).ToList();
//not exist split that may help
if (resultSplitsProperties.Count == 0)
{
return;
}
SplitProps bestSplit = resultSplitsProperties.Aggregate((a, b) => (a.error < b.error) ? a : b);
if (!bestSplit.isPartitionOk)
{
return;
}
parentNode.typeTransformed = bestSplit.type;
GeoWave child0 = new GeoWave(_rc.dim, _rc.labelDim, _rc);
GeoWave child1 = new GeoWave(_rc.dim, _rc.labelDim, _rc);
child0.dimIndex = bestSplit.dimIndex;
child1.dimIndex = bestSplit.dimIndex;
List <int> sortedIds = bestSplit.sortedIds;
int splitId = bestSplit.splitId;
//set childs id's
child0.pointsIdArray = sortedIds.GetRange(0, sortedIds.IndexOf(splitId));
child1.pointsIdArray = sortedIds.GetRange(sortedIds.IndexOf(splitId), sortedIds.Count - child0.pointsIdArray.Count);
// set upper split value only
child0.upperSplitValue = bestSplit.splitValue;
//set mean values
setTransformedChildMeanValue(ref child0);
setTransformedChildMeanValue(ref child1);
//set parent id
child0.parentID = geoWaveId;
child1.parentID = geoWaveId;
//set level
child0.level = parentNode.level + 1;
child1.level = parentNode.level + 1;
//debug writelines
Debug.WriteLine("************Parent Size:" + parentNode.pointsIdArray.Count);
Debug.WriteLine("************Level:" + (parentNode.level + 1));
Debug.WriteLine("************Type Splitted:" + bestSplit.type);
Debug.WriteLine("***********************************************************");
//!!! START DEBUG VISULIZE SPIRAL SPLIT
/* double[][] child0Data = child0.pointsIdArray.Select(id => _trainingDt[id]).ToArray();
* double[][] child1Data = child1.pointsIdArray.Select(id => _trainingDt[id]).ToArray();
* double[][] child0responce = child0.pointsIdArray.Select(id => _trainingLabel[id]).ToArray();
* double[][] child1responce = child1.pointsIdArray.Select(id => _trainingLabel[id]).ToArray();
* int level = child0.level;
* PrintEngine.debugVisualizeSpiralSplit(child0Data, child1Data,
* child0responce, child1responce,
* level, parentNode.typeTransformed, debugAnalysisFolderName);*/
//!!! END DEBUG VISUALIZE SPIRAL SPLIT'
//compute norms
child0.computeNormOfConsts(parentNode);
child1.computeNormOfConsts(parentNode);
child0.meanDiffFromParent = child0.MeanValue[0] - parentNode.MeanValue[0];
child1.meanDiffFromParent = child1.MeanValue[0] - parentNode.MeanValue[0];
geoWaveArr.Add(child0);
geoWaveArr.Add(child1);
parentNode.child0 = geoWaveArr.IndexOf(child0);
parentNode.child1 = geoWaveArr.IndexOf(child1);
//RECURSION STEP !!!
recursiveBSP_TransformedData(geoWaveArr, parentNode.child0, splitTypes);
recursiveBSP_TransformedData(geoWaveArr, parentNode.child1, splitTypes);
}
