public static HdbscanResult Run <T>(HdbscanParameters <T> parameters)
{
var numPoints = parameters.DataSet?.Length ?? parameters.Distances.Length;
PrecomputeNormalMatrixDistancesIfApplicable(parameters, numPoints);
var sparseDistance = PrecomputeSparseMatrixDistancesIfApplicable(parameters, numPoints);
var internalDistanceFunc = DetermineInternalDistanceFunc(parameters, sparseDistance, numPoints);
// Compute core distances
var coreDistances = HdbscanAlgorithm.CalculateCoreDistances(
internalDistanceFunc,
numPoints,
parameters.MinPoints);
// Calculate minimum spanning tree
var mst = HdbscanAlgorithm.ConstructMst(
internalDistanceFunc,
numPoints,
coreDistances,
true);
mst.QuicksortByEdgeWeight();
var pointNoiseLevels = new double[numPoints];
var pointLastClusters = new int[numPoints];
var hierarchy = new List <int[]>();
// Compute hierarchy and cluster tree
var clusters = HdbscanAlgorithm.ComputeHierarchyAndClusterTree(
mst,
parameters.MinClusterSize,
parameters.Constraints,
hierarchy,
pointNoiseLevels,
pointLastClusters);
// Propagate clusters
var infiniteStability = HdbscanAlgorithm.PropagateTree(clusters);
// Compute final flat partitioning
var prominentClusters = HdbscanAlgorithm.FindProminentClusters(
clusters,
hierarchy,
numPoints);
// Compute outlier scores for each point
var scores = HdbscanAlgorithm.CalculateOutlierScores(
clusters,
pointNoiseLevels,
pointLastClusters,
coreDistances);
return(new HdbscanResult
{
Labels = prominentClusters,
OutliersScore = scores,
HasInfiniteStability = infiniteStability
});
}
public static HdbscanResult Run(HdbscanParameters parameters)
{
var numPoints = parameters.DataSet != null
? parameters.DataSet.Length
: parameters.Distances.Length;
if (parameters.Distances == null)
{
// Precompute distances.
var distances = new double[numPoints][];
for (var i = 0; i < distances.Length; i++)
{
distances[i] = new double[numPoints];
}
if (parameters.UseMultipleThread)
{
var size = numPoints * numPoints;
var maxDegreeOfParallelism = parameters.MaxDegreeOfParallelism;
if (maxDegreeOfParallelism == 0)
{
// Not specified. Use all threads.
maxDegreeOfParallelism = Environment.ProcessorCount;
}
var option = new ParallelOptions {
MaxDegreeOfParallelism = Math.Max(1, maxDegreeOfParallelism)
};
Parallel.For(0, size, option, index =>
{
var i = index % numPoints;
var j = index / numPoints;
if (i < j)
{
var distance = parameters.DistanceFunction.ComputeDistance(
parameters.DataSet[i],
parameters.DataSet[j]);
distances[i][j] = distance;
distances[j][i] = distance;
}
});
}
else
{
for (var i = 0; i < numPoints; i++)
{
for (var j = 0; j < i; j++)
{
var distance = parameters.DistanceFunction.ComputeDistance(
parameters.DataSet[i],
parameters.DataSet[j]);
distances[i][j] = distance;
distances[j][i] = distance;
}
}
}
parameters.Distances = distances;
}
// Compute core distances
var coreDistances = HdbscanAlgorithm.CalculateCoreDistances(
parameters.Distances,
parameters.MinPoints);
// Calculate minimum spanning tree
var mst = HdbscanAlgorithm.ConstructMst(
parameters.Distances,
coreDistances,
true);
mst.QuicksortByEdgeWeight();
var pointNoiseLevels = new double[numPoints];
var pointLastClusters = new int[numPoints];
var hierarchy = new List <int[]>();
// Compute hierarchy and cluster tree
var clusters = HdbscanAlgorithm.ComputeHierarchyAndClusterTree(
mst,
parameters.MinClusterSize,
parameters.Constraints,
hierarchy,
pointNoiseLevels,
pointLastClusters);
// Propagate clusters
var infiniteStability = HdbscanAlgorithm.PropagateTree(clusters);
// Compute final flat partitioning
var prominentClusters = HdbscanAlgorithm.FindProminentClusters(
clusters,
hierarchy,
numPoints);
// Compute outlier scores for each point
var scores = HdbscanAlgorithm.CalculateOutlierScores(
clusters,
pointNoiseLevels,
pointLastClusters,
coreDistances);
return(new HdbscanResult
{
Labels = prominentClusters,
OutliersScore = scores,
HasInfiniteStability = infiniteStability
});
}
