private PointClusters GetBestKMeans(List <Case> allPoints)
{
Dictionary <int, PointClusters> AllClusters = new Dictionary <int, PointClusters>();
PointClusters cluster; //= new PointClusters();
List <Case> seedPoints = new List <Case>();
double Sk = 0;
// Number of Attributes
int Nd = allPoints[0].GetFeatures().Count - 2;//except id , cluster
seedPoints = GetSeedPoints(allPoints, 1);
cluster = GetKMeans(allPoints, seedPoints, 1);
foreach (KeyValuePair <Case, double> pair in cluster.dist_error)
{
Sk += pair.Value;
}
cluster.Sk = Sk;
cluster.Fk = 100;
AllClusters.Add(1, cluster);
for (int i = 2; i <= 19; i++)//
{
List <KmeansPlus.PointClusters> Allmyclusters = new List <KmeansPlus.PointClusters>();
double min_fk = Double.MaxValue;
int min_i = 0;
for (int j = 0; j < 1; j++)
{
cluster = new PointClusters();
seedPoints = GetSeedPoints(allPoints, i);
iter = 0;
cluster = GetKMeans(allPoints, seedPoints, i);
if (AllClusters[i - 1].Sk == 0)
{
cluster.Fk = 1;
}
else
{
double ak = 1;
if (Nd > 1)
{
ak = Convert.ToDouble(Ak(i, Nd));
}
cluster.Fk = cluster.Sk / (ak * AllClusters[i - 1].Sk);
}
Allmyclusters.Add(cluster);
if (cluster.Fk < min_fk)
{
min_i = j; min_fk = cluster.Fk;
}
}
cluster = Allmyclusters[min_i];
AllClusters.Add(i, cluster);
}
double min = double.MaxValue;
int kk = 0;
foreach (KeyValuePair <int, PointClusters> pair in AllClusters)
{
if (pair.Value.Fk < min)
{
kk = pair.Key; cluster = pair.Value; min = pair.Value.Fk;
}
}
return(cluster);
}
/// <summary>
/// Get the centroid of a set of points
/// cf. http://en.wikipedia.org/wiki/Centroid
/// Consider also: Metoid cf. http://en.wikipedia.org/wiki/Medoids
/// </summary>
/// <param name="pcs"></param>
/// <returns></returns>
private List <Case> GetCentroid(PointClusters pcs)
{
List <Case> newSeeds = new List <Case>(pcs.PC.Count);
Case newSeed;
int sumf = 0; double sumd = 0; bool sumb = false;
List <string> sumst = new List <string>();
List <int> sums_count = new List <int>();
Feature f = new Feature();
int t = 0;
foreach (List <Case> cluster in pcs.PC.Values)
{
newSeed = new Case(cluster[0].GetCaseID(), cluster[0].GetCaseName(), cluster[0].GetCaseDescription());
for (int j = 0; j < cluster[0].GetFeatures().Count; j++)
{
foreach (Case p in cluster)
{
f = (Feature)p.GetFeatures()[j];
if (f.GetFeatureName() == "id" || f.GetFeatureName() == "cluster")
{
break;
}
if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_INT)
{
sumf += Convert.ToInt32(f.GetFeatureValue());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_FLOAT)
{
sumd += Convert.ToDouble(f.GetFeatureValue());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_BOOL)
{
sumb = (sumb || Convert.ToBoolean(f.GetFeatureValue()));
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_STRING || f.GetFeatureType() == FeatureType.TYPE_FEATURE_CATEGORICAL)
{
int position = sumst.IndexOf(f.GetFeatureValue().ToString());
if (position >= 0)
{
sums_count[position] = sums_count[position] + 1;
}
else
{
sumst.Add(f.GetFeatureValue().ToString());
sums_count.Add(1);
}
}
}
if (f.GetFeatureName() == "cluster")
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), -1, f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
else if (f.GetFeatureName() == "id")
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), t, f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_INT)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumf / cluster.Count(), f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_FLOAT)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumd / cluster.Count(), f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_BOOL)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumb, f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_STRING || f.GetFeatureType() == FeatureType.TYPE_FEATURE_CATEGORICAL)
{
int max = sums_count.IndexOf(sums_count.Max());
string string_value = sumst[max];
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), string_value, f.GetWeight(), f.GetIsKey(), f.GetIsIndex(), f.GetFeatureUnit());
}
sumf = 0;
sumd = 0;
sumb = false;
sumst.Clear();
sums_count.Clear();
} // end feature
newSeeds.Add(newSeed);
sumf = 0;
sumd = 0;
sumb = false;
sumst.Clear();
sums_count.Clear();
t++;
}// end clusters
return(newSeeds);
}
//Bog standard k-means.
private PointClusters GetKMeans(List <Case> allPoints, List <Case> seedPoints, int k)
{
begin : PointClusters cluster = new PointClusters();
double[] Distances = new double[k];
double minD = double.MaxValue;
List <Case> sameDPoint = new List <Case>();
bool exit = true;
//Cycle thru all points in ensemble and assign to nearest centre
foreach (Case p in allPoints)
{
foreach (Case sPoint in seedPoints)
{
double dist = GetEuclideanD(p, sPoint);
if (dist < minD)
{
sameDPoint.Clear();
minD = dist;
sameDPoint.Add(sPoint);
}
else if (dist == minD)
{
if (!sameDPoint.Contains(sPoint))
{
sameDPoint.Add(sPoint);
}
}
}
//Extract nearest central point.
Case keyPoint;
if (sameDPoint.Count > 1)
{
int index = GetRandNumCrypto(0, sameDPoint.Count);
keyPoint = sameDPoint[index];
}
else
{
keyPoint = sameDPoint[0];
}
//Assign ensemble point to correct central point cluster
if (!cluster.PC.ContainsKey(keyPoint)) //New
{
List <Case> newCluster = new List <Case>();
newCluster.Add(p);
cluster.PC.Add(keyPoint, newCluster);
}
else
{ //Existing cluster centre
cluster.PC[keyPoint].Add(p);
}
//Reset
sameDPoint.Clear();
minD = double.MaxValue;
}
//Bulletproof check - it it come out of the wash incorrect then re-seed.
if (cluster.PC.Count != k)
{
cluster.PC.Clear();
seedPoints = GetSeedPoints(allPoints, k);
goto begin;
}
List <Case> newSeeds = GetCentroid(cluster);
//Determine exit
foreach (Case newSeed in newSeeds)
{
if (!cluster.PC.ContainsKey(newSeed))
{
exit = false;
}
}
if ((exit) || (iter == 1000))
{
return(cluster);
}
else
{
iter++;
return(GetKMeans(allPoints, newSeeds, k));
}
}
private PointClusters GetKMeans(List <Case> allPoints, List <Case> seedPoints, int k)
{
begin : PointClusters cluster = new PointClusters();
double[] Distances = new double[k];
double minD = double.MaxValue;
List <Case> sameDPoint = new List <Case>();
bool exit = true;
//Cycle thru all points in ensemble and assign to nearest centre
foreach (Case p in allPoints)
{
foreach (Case sPoint in seedPoints)
{
double dist = GetEuclideanD(p, sPoint);
if (dist < minD)
{
sameDPoint.Clear();
minD = dist;
sameDPoint.Add(sPoint);
}
else if (dist == minD)
{
if (!sameDPoint.Contains(sPoint))
{
sameDPoint.Add(sPoint);
}
}
}
//Extract nearest central point.
Case keyPoint;
if (sameDPoint.Count > 1)
{
int index = GetRandNumCrypto(0, sameDPoint.Count);
keyPoint = sameDPoint[index];
}
else
{
keyPoint = sameDPoint[0];
}
//Assign ensemble point to correct central point cluster
if (!cluster.PC.ContainsKey(keyPoint)) //New
{
List <Case> newCluster = new List <Case>();
newCluster.Add(p);
cluster.PC.Add(keyPoint, newCluster);
cluster.dist_error.Add(keyPoint, minD);
}
else
{ //Existing cluster centre
cluster.PC[keyPoint].Add(p);
double value = cluster.dist_error[keyPoint];
cluster.dist_error[keyPoint] += minD;
}
//Reset
sameDPoint.Clear();
minD = double.MaxValue;
}
//Bulletproof check - it it come out of the wash incorrect then re-seed.
if (cluster.PC.Count != k)
{
cluster.PC.Clear();
cluster.dist_error.Clear();
seedPoints = GetSeedPoints(allPoints, k);
goto begin;
}
List <Case> newSeeds = GetCentroid(cluster);
List <Case> n = new List <Case>();
bool found = true;
//Determine exit
// check if centers don't change, equality between cases
foreach (Case newSeed in newSeeds) // last centers
{
foreach (KeyValuePair <Case, List <Case> > item in cluster.PC) //current centers
{
found = true;
foreach (Feature f in item.Key.GetFeatures())
{
Feature seedf = newSeed.GetFeature(f.GetFeatureName());
if (!(f.GetFeatureValue().ToString() == seedf.GetFeatureValue().ToString()))
{
found = false; break;
}
}
if (found)
{
break;
}
}
if (!found)
{
exit = false;
}
// if (!cluster.PC.ContainsKey(newSeed)) wrong for equal objects
// exit = false;
}
if ((exit) || (iter == 1000))
{
return(cluster);
}
else
{
iter++;
return(GetKMeans(allPoints, newSeeds, k));
}
}
/// <summary>
/// Get the centroid of a set of points
/// cf. http://en.wikipedia.org/wiki/Centroid
/// Consider also: Metoid cf. http://en.wikipedia.org/wiki/Medoids
/// </summary>
/// <param name="pcs"></param>
/// <returns></returns>
private List <Case> GetCentroid(PointClusters pcs)
{
List <Case> newSeeds = new List <Case>(pcs.PC.Count);
Case newSeed;
int sumf = 0; double sumd = 0; bool sumb = false; List <string> sumst = new List <string>();
Feature f = new Feature();
foreach (List <Case> cluster in pcs.PC.Values)
{
newSeed = new Case(cluster[0].GetCaseID(), cluster[0].GetCaseName(), cluster[0].GetCaseDescription());
for (int j = 0; j < cluster[0].GetFeatures().Count; j++)
{
foreach (Case p in cluster)
{
f = (Feature)p.GetFeatures()[j];
if (f.GetFeatureName() == "id")
{
break;
}
if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_INT)
{
sumf += Convert.ToInt32(f.GetFeatureValue());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_FLOAT)
{
sumd += Convert.ToDouble(f.GetFeatureValue());
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_BOOL)
{
sumb = (sumb || Convert.ToBoolean(f.GetFeatureValue()));
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_STRING)
{
sumst.Add(f.GetFeatureValue().ToString());
}
}
if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_INT)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumf / cluster.Count(), 1.0, false, false, "");
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_FLOAT)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumd / cluster.Count(), 1.0, false, false, "");
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_BOOL)
{
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumb, 1.0, false, false, "");
}
else if (f.GetFeatureType() == FeatureType.TYPE_FEATURE_STRING)
{
Random rd = new Random(); // for calculating random numbers
int rnd = rd.Next(sumst.Count);
newSeed.AddFeature(f.GetFeatureName(), f.GetFeatureType(), sumst[rnd].ToString(), 1.0, false, false, "");
}
// newSeed = new Case(sumX / cluster.Count, sumY / cluster.Count);
// newSeeds.Add(newSeed);
}
newSeeds.Add(newSeed);
sumf = 0;
sumd = 0;
sumb = false;
sumst.Clear();
}
return(newSeeds);
}
