public void CorrectlyCalculateCosineSimilarity(double[] vectorA, double[] vectorB, double expectedSimilarity)
{
var similarityFunction =
DistanceFunctionResolver.ResolveDistanceFunction(DistanceFunctionType.Cosine);
var calculatedSimilarity = similarityFunction.Invoke(vectorA, vectorB);
Assert.Equal(expectedSimilarity, calculatedSimilarity, 8);
}
/// <summary>
/// Calculates a matrix which organises the distances between data points in the embeddings.
/// </summary>
private static Dictionary <string, Dictionary <string, double> > CalculateDistanceMatrix(
IReadOnlyCollection <IEmbedding> embeddings,
DistanceFunctionType distanceFunctionType)
{
var distanceFunction = DistanceFunctionResolver.ResolveDistanceFunction(distanceFunctionType);
var matrix = new Dictionary <string, Dictionary <string, double> >();
foreach (var embedding in embeddings)
{
var distances = embeddings.ToDictionary(
otherIEmbedding => otherIEmbedding.Label,
otherIEmbedding => distanceFunction.Invoke(embedding.Vector, otherIEmbedding.Vector));
matrix.Add(embedding.Label, distances);
}
return(matrix);
}
public static Dictionary <string, int> GetLabelClusterMap(
IEnumerable <IEmbedding> embeddings,
double epsilon = 0.5,
int minimumSamples = 5,
DistanceFunctionType distanceFunctionType = DistanceFunctionType.Euclidean,
int concurrentThreads = 4)
{
var embeddingsList = embeddings.ToList();
var distanceFunction = DistanceFunctionResolver.ResolveDistanceFunction(distanceFunctionType);
var clusterLabels = new ConcurrentDictionary <string, int>();
var clusterRelationships = new ConcurrentBag <ConcurrentBag <int> >();
var clusterIndex = 0;
var sampleSize = (int)Math.Ceiling((double)embeddingsList.Count / concurrentThreads);
Parallel.For(0, concurrentThreads, threadIndex =>
{
foreach (var embedding in embeddingsList.Skip(threadIndex * sampleSize).Take(sampleSize))
{
if (clusterLabels.ContainsKey(embedding.Label))
{
continue;
}
var neighbors = GetNeighborsAndWeight(
embedding,
embeddingsList,
distanceFunction,
epsilon);
if (neighbors.Count < minimumSamples)
{
clusterLabels.AddOrUpdate(
embedding.Label,
-1,
(key, existingClusterIndex) => existingClusterIndex);
continue;
}
var localClusterIndex = clusterIndex++;
clusterLabels.AddOrUpdate(
embedding.Label,
(key) =>
{
clusterRelationships.Add(new ConcurrentBag <int> {
localClusterIndex
});
return(localClusterIndex);
},
(key, existingClusterIndex) =>
{
clusterRelationships.First(r => r.Contains(existingClusterIndex)).Add(localClusterIndex);
return(localClusterIndex);
});
for (var i = 0; i < neighbors.Count; i++)
{
var currentNeighbor = neighbors[i];
if (clusterLabels.TryGetValue(currentNeighbor.Label, out var existingClusterId))
{
if (existingClusterId != -1 && existingClusterId != localClusterIndex)
{
clusterRelationships.First(r => r.Contains(existingClusterId)).Add(localClusterIndex);
}
clusterLabels[currentNeighbor.Label] = localClusterIndex;
continue;
}
clusterLabels.AddOrUpdate(
currentNeighbor.Label,
localClusterIndex,
(key, existingClusterIndex) =>
{
clusterRelationships.First(r => r.Contains(existingClusterIndex)).Add(localClusterIndex);
return(localClusterIndex);
});
var currentNeighborsNeighbors = GetNeighborsAndWeight(
currentNeighbor,
embeddingsList,
distanceFunction,
epsilon);
if (currentNeighborsNeighbors.Count >= minimumSamples)
{
neighbors = neighbors.Union(currentNeighborsNeighbors).ToList();
}
}
}
});
var clusterIndexMap = GetClusterIndexMap(clusterRelationships);
return(clusterLabels.ToDictionary(
x => x.Key,
x => clusterIndexMap[x.Value]));
}