/// <summary>
/// Performs KMeans partitioning (clustering) on the data. This is done on all
/// parameters within the DataLines except for the dates, since they are not
/// easily normalized (without loosing the meaning).
/// </summary>
/// <param name="k">The number of partitions to use.</param>
/// <param name="data">The dataset that needs partitioning.</param>
/// <returns>A list of KMeanCluster together containing all the entries from
/// the original dataset.</returns>
public static List <KMeanCluster> KMeansPartition(int k, List <DataLine> data)
{
// select k random
KMeanCluster[] clusters = new KMeanCluster[k];
data.Shuffle(1337);
for (int i = 0; i < k; i++)
{
clusters[i] = new KMeanCluster(data[i]);
}
int iteration = 0;
while (iteration < 10000)
{ // upper cutoff
// beregn alle iris' til de k means
data.ForEach(a => clusters.OrderBy(c => Dissimilarity(a, c)).First().AddMember(a));
// beregn nye centroids
KMeanCluster[] newClusters = new KMeanCluster[k];
for (int c = 0; c < k; c++)
{
newClusters[c] = clusters[c].CalcCentroid();
}
// stop hvis ingen ændringer
if (Changed(clusters, newClusters))
{
clusters = newClusters;
}
else
{
break;
}
iteration++;
Console.WriteLine("Iteration " + iteration + " done.");
}
return(clusters.ToList());
}
private static double Dissimilarity(DataLine a, KMeanCluster c)
{
Dictionary <string, double> centroid = c.Centroid;
return(centroid.Sum(kv => Math.Abs(a.hashDoubles[kv.Key] ?? 0.0 - kv.Value))); // TODO: null == 0.0 might not be good
}
