private Cluster[] GetCentroids(short[] docIds)
{
Cluster[] clusters = GetInitialClusters(docIds, getSampleSize(docIds.Length));
DocTermMatrix matrix = new DocTermMatrix(docIds, GetIndex);
do
{
int clusterToMerge1 = 0;
int clusterToMerge2 = 0;
float sim = 0;
float newSim = 0;
for (int i = 0; i < clusters.Length; i++)
{
for (int j = 0; j < clusters.Length; j++)
{
if (i != j)
{
newSim = getSimilarity(clusters[i], clusters[j], matrix);
if (newSim > sim)
{
clusterToMerge1 = i;
clusterToMerge2 = j;
sim = newSim;
}
}
}
}
clusters = mergeClusters(clusterToMerge1, clusterToMerge2, clusters);
}while (clusters.Length > GetK);
return(clusters);
}
private float getSimilarity(Cluster c1, Cluster c2, DocTermMatrix matrix)
{
float centroid1 = matrix.GetCentroid(c1);
float centroid2 = matrix.GetCentroid(c2);
float norm1 = matrix.GetNorm(c1);
float norm2 = matrix.GetNorm(c2);
return((centroid1 * centroid2) / (norm1 * norm2));
}
public override Cluster[] CalculateClusters(short[] docIds, int commonTerms)
{
DocTermMatrix matrix = new DocTermMatrix(docIds, GetIndex);
Cluster[] prevClusters = GetEmptyClusters(GetK);
Cluster[] newClusters;
if (centroids == null)
{
newClusters = GetInitialClusters(docIds, GetK);
}
else
{
newClusters = centroids;
}
int iterations = 0;
int maxIterations = 10;
do
{
prevClusters = newClusters; //save previous clusters
newClusters = GetEmptyClusters(GetK); //empty new clusters and re-populate, use prevClusters as means
for (int i = 0; i < docIds.Length; i++) //for each doc
{
short docId = docIds[i];
float sim = 0;
int assignedCluster = -1;
for (int j = 0; j < GetK; j++) //calculate similarity with each clustern
{
float newSim = getAvgSimilarity(docId, prevClusters[j], matrix);
if (newSim > sim) //if new similarity is higher - update similarity and asign to new cluster
{
sim = newSim;
assignedCluster++;
}
}
if (assignedCluster > -1) //assign to cluster ONLY if sim to any cluster is > 0
{
newClusters[assignedCluster].AddDoc(docId);
}
}
} while (iterations++ < maxIterations && !areConverged(prevClusters, newClusters));
if (getTerms)
{
return(LoadCommonTerms(removeEmptyClusters(newClusters), commonTerms));
}
else
{
return(removeEmptyClusters(newClusters));
}
}
private float getAvgSimilarity(short docId, Cluster c, DocTermMatrix matrix)
{
float sim = 0;
Hashtable docIds = c.DocIds;
IDictionaryEnumerator en = docIds.GetEnumerator();
while (en.MoveNext())
{
sim += matrix.GetSimilarity(docId, Convert.ToInt16(en.Key));
}
return(sim / c.DocumentCount);
}
public ClusteringAnalyzer(Index index, short[] allDocIds)
{
this.index = index;
this.allDocIds = allDocIds;
matrix = new DocTermMatrix(allDocIds, index);
}
