public void GetXDistanceFromSample_GiveCorrectDistance()
{
Sample sample1 = new Sample(new double[] { 0, 3, 4, 5}, 1, 0);
Sample sample2 = new Sample(new double[] { 7, 6, 3, -1}, 1, 0);
double distance = Math.Round(sample1.GetXDistanceFromSample(sample2), 3);
Assert.AreEqual(distance, 9.747);
}
public void ClusterXCreate_CreateCorrectClusterX()
{
// set params
Params.inputDataDimension = 3;
Sample sample = new Sample(new double[] { 1.0, 2.0, 3.0 }, 1.0, 0);
ClusterX clusterX = new ClusterX(sample, null);
Assert.IsTrue(clusterX.Mean.EqualsToVector(new Vector(new double[] { 1.0, 2.0, 3.0 })));
}
public ClusterY(Sample sample, Node parent)
: base(sample, parent)
{
this.dimension = Params.outputDataDimension;
this.items.Add(new Vector(sample.Y.Values.ToArray(), sample.Label, this.items.Count + 1));
this.mean = new Vector(sample.Y.Values.ToArray());
}
public ClusterPair(ClusterX cX, ClusterY cY, Sample sample)
{
clusterX = cX;
clusterY = cY;
this.PreviousCenter = 0;
this.CurrentCenter = 0;
this.samples = new List<Sample>() { sample };
}
public ClusterX(Sample sample, Node parent)
: base(sample, parent)
{
this.child = null;
this.dimension = Params.inputDataDimension;
items.Add(new Vector(sample.X.Values.ToArray(), sample.Label, 1));
this.mean = new Vector(sample.X.Values.ToArray());
//#warning TODO count covariance matrix
//this.covarianceMatrix = ILMath.zeros(Params.inputDataDimension, Params.inputDataDimension);
}
public void CountGSOManifold_ProvideCorrectCounting()
{
Params.inputDataDimension = 4;
Node node = new Node(0.0, 0.0, 0.0, 0.0, "");
Sample s1 = new Sample(new double[] { 1, 0, 2, 1 }, 1, 0);
Sample s2 = new Sample(new double[] { -1, 1, 0, -1 }, 1, 0);
Sample s3 = new Sample(new double[] { 2, 1, 1, 1 }, 1, 0);
node.ClustersX.Add(new ClusterX(s1, null));
node.ClustersX.Add(new ClusterX(s2, null));
node.ClustersX.Add(new ClusterX(s3, null));
List<Vector> scatterVectors = new List<Vector>();
scatterVectors.Add(new Vector(new double[] { 1, 0, 2, 1 }));
scatterVectors.Add(new Vector(new double[] { -1, 1, 0, -1 }));
scatterVectors.Add(new Vector(new double[] { 2, 1, 1, 1 }));
ILArray<double> array = node.GetManifold(scatterVectors);
}
public Cluster(Sample sample, Node parent)
{
this.items = new List<Vector>();
this.parent = parent;
}
/// <summary>
/// Create new clusers X and Y and their cluster pair
/// </summary>
/// <param name="sample">new sample</param>
private void CreateNewClusters(Sample sample, Node parent)
{
ClusterX newClusterX = new ClusterX(sample, parent);
this.clustersX.Add(newClusterX);
ClusterY newClusterY = new ClusterY(sample, parent);
this.clustersY.Add(newClusterY);
ClusterPair clusterPair = new ClusterPair(newClusterX, newClusterY, sample);
newClusterX.SetClusterPair(clusterPair);
newClusterY.SetClusterPair(clusterPair);
clusterPair.Id = clusterPairs.Count;
clusterPair.Samples.Add(sample);
this.clusterPairs.Add(clusterPair);
}
/// <summary>
/// get nearest cluster pair by most dicrimnating features vector
/// </summary>
/// <param name="sample"></param>
/// <param name="distance"></param>
/// <returns></returns>
private ClusterPair GetNearestClusterPairXBySDNLL_MDF(Sample sample, out double distance, out int index)
{
// convert vector to mdf vector
ILArray<double> x = sample.X.Values.ToArray();
ILArray<double> scaterPart = x - this.c;
ILArray<double> mdfVector = ILMath.multiply(this.gSOManifold.T, scaterPart);
distance = double.MaxValue;
ClusterPair closestPair = clusterPairs[0];
int i = 0;
index = -1;
foreach (ClusterPair item in clusterPairs)
{
double newDistance = item.X.GetSDNLL_MDF(mdfVector);
if (newDistance < distance)
{
distance = newDistance;
closestPair = item;
index = i;
}
i++;
}
//Console.WriteLine("MDF" + sample.Id.ToString() + " : " + index.ToString());
return closestPair;
}
private double GetDistanceFromClosestCenter(List<Sample> centers, Sample sample)
{
double minDistance = Double.MaxValue;
foreach (var item in centers)
{
double distance = item.X.GetDistance(sample.X);
if (distance < minDistance)
{
minDistance = distance;
}
}
return minDistance;
}
// for k-means
private void UpdateCenterY(Sample center, List<Sample> listOfSamples)
{
center.Y = Vector.GetMeanOfVectors(listOfSamples.Select(sample => sample.Y).ToList());
}
private List<Tuple<double, ClusterPair>> GetDistancesAndClusterPairsY(Sample sample)
{
List<Tuple<double, ClusterPair>> result = new List<Tuple<double, ClusterPair>>();
foreach (ClusterPair item in clusterPairs)
{
if (sample.Y.Values.Count() < 2)
{
throw new InvalidOperationException("Bad operation");
}
double newDistance = item.Y.Mean.GetDistance(sample.Y);
result.Add(new Tuple<double,ClusterPair>(newDistance, item));
}
result = result.OrderBy(i => i.Item1).ToList();
return result;
}
public TestResult GetLabelOfCategory(Sample item)
{
return this.root.GetLabelOfCategory(item);
}
public ClusterPair GetNearestClusterPairX(Sample sample, out double distance, out int index)
{
distance = double.MaxValue;
ClusterPair closestPair = clusterPairs[0];
int i = 0;
index = -1;
foreach (ClusterPair item in clusterPairs)
{
double newDistance = item.X.Mean.GetDistance(sample.X);
if (newDistance < distance)
{
distance = newDistance;
closestPair = item;
index = i;
}
i++;
}
return closestPair;
}
public Node GetNextAByEuclidean(Sample sample)
{
if (children.Count > 0)
{
double distance = double.MaxValue;
Node result = children[0];
foreach (var node in children)
{
int clusterIndex = 1;
foreach (var cluster in node.ClustersX)
{
double tmpDistance = cluster.Mean.GetDistance(sample.X);
//Console.WriteLine(String.Format("E : N : {0}, C:{1}, D:{2}", node.Id, clusterIndex, tmpDistance.ToString()));
clusterIndex++;
if (tmpDistance < distance)
{
distance = tmpDistance;
result = node;
}
}
}
return result;
}
return null;
}
//public void CountClosestClusterPairByWidthSearch(Sample item, ClusterPairTestResult result)
//{
// if (this.children != null && this.children.Count != 0 && this.children[0].IsLeafNode)
// {
// double distance = double.MaxValue;
// int index = 0;
// ClusterPair clPair = this.GetNearestClusterPairXBySDNLL_MDF(item, out distance, out index);
// if (distance < result.Distance)
// {
// result.ClusterPair = clPair;
// result.Distance = distance;
// }
// }
// else
// {
// var nodesToSearch = this.GetNodesToSearch(item);
// foreach (var node in nodesToSearch)
// {
// node.CountClosestClusterPairByWidthSearch(item, result);
// }
// }
//}
public void CountClosestClusterPairByWidthSearch(Sample item, ClusterPairTestResult result)
{
if (this.IsLeafNode)
{
if (this.Parent != null)
{
double distance = double.MaxValue;
int index = 0;
ClusterPair clPair = this.Parent.GetNearestClusterPairXBySDNLL_MDF(item, out distance, out index);
Console.WriteLine(distance.ToString());
if (distance < result.Distance)
{
//Console.WriteLine("Store");
result.ClusterPair = clPair;
result.Distance = distance;
}
}
}
else
{
var nodesToSearch = this.GetNodesToSearch(item);
foreach (var node in nodesToSearch)
{
node.CountClosestClusterPairByWidthSearch(item, result);
}
}
}
public TestResult GetLabelOfCategory(Sample item)
{
if (this.isLeafNode)
{
double distance = double.MinValue;
int index = int.MinValue;
ClusterPair clPair = this.GetNearestClusterPairX(item, out distance, out index);
TestResult tr = new TestResult()
{
ClusterMeanX = clPair.X.Mean,
ClusterMeanY = clPair.Y.Mean,
Label = clPair.X.Label,
Samples = clPair.Samples
};
return tr;
}
double distance2 = double.MinValue;
int index2 = int.MinValue;
ClusterPair nearestClPair = this.GetNearestClusterPairXBySDNLL_MDF(item, out distance2, out index2);
Node next = nearestClPair.CorrespondChild;
return next.GetLabelOfCategory(item);
}
private void UpdateClusters(Sample sample)
{
// parameters bly bound of number of y clusters in node, dy resolution
// find nearest xj cluster using euclidean distance
ClusterPair nearestCluster = null;
double distance = double.MaxValue;
int index = -1;
//nearestCluster = this.GetNearestClusterPairY(sample, out distance, out index);
List<Tuple<double, ClusterPair>> orderedClusterPairs = this.GetDistancesAndClusterPairsY(sample);
// if is count < like bly and distance > deltay create new cluster
// add new cluster pair (x,y), increment n
if (this.clusterPairs.Count < Params.bly && orderedClusterPairs[0].Item1 > this.deltaY)
{
if (!Params.StoreItems)
{
throw new InvalidOperationException("Not possible create new cluster. No items available for MDF counting.");
}
this.CreateNewClusters(sample, this);
// update MDF space
this.CountC();
this.CountGSOManifold();
this.CountMDFOfVectors();
this.CountMDFMeans();
this.CountCovarianceMatricesMDF();
//count cov matrix mean
this.CountCovarianceMatrixMeanMDF();
}
// else update p percents of xj cluster and yj cluster using amnesic average
else
{
#warning TODO update ceratin portion
int countOfClusters = orderedClusterPairs.Count;
int countOfClustersToUpdate = (int)((orderedClusterPairs.Count - 1 ) * Params.p) + 1;
//Update a certain portion p (e.g., p = 0:2, i.e., pulling top 20%) of nearest clusters using the amnesic average
//explained in Section III-F and return the index j
for (int i = 0; i < countOfClustersToUpdate; i++)
{
//Console.WriteLine("Y");
orderedClusterPairs[i].Item2.Y.AddItem(sample.Y, sample.Label);
}
Vector newItem = new Vector(sample.X.Values.ToArray());
newItem.Label = sample.Label;
newItem.CountMDF(this.gSOManifold, this.c);
// add sample to clusters, update statistics of clusters
orderedClusterPairs[0].Item2.X.AddItemNonLeaf(newItem);
if (Params.StoreSamples)
{
orderedClusterPairs[0].Item2.Samples.Add(sample);
}
// update meanMDF and varianceMDF
this.UpdateMeanAndVarianceMdf(newItem);
//count cov matrix mean
this.CountCovarianceMatrixMeanMDF();
}
}
private void UpdateClusterPairsX_ForSwapping(Sample sample)
{
double distance = 0.0;
int index = 0;
ClusterPair nearestCluster = this.GetNearestClusterPairX(sample, out distance, out index);
//Console.WriteLine(distance.ToString());
// if is count < like bl and distance > delta create new cluster
// add new cluster pair (x,y), increment n
if (clusterPairs.Count < this.blx && distance > this.deltaX)
{
this.CreateNewClusters(sample, this);
}
// else update xj cluster and yj cluster using amnesic average
else
{
// add sample to clusters, update statistics of clusters
nearestCluster.AddItem(sample);
}
}
private void UpdateClusterPairsX(Sample sample)
{
// update cluster pairs
// parameters bl bound of number of microclusters in node, dx resolution
// find nearest xj cluster using euclidean distance
double distance = 0.0;
int index = 0;
ClusterPair nearestCluster = this.GetNearestClusterPairX(sample, out distance, out index);
//Console.WriteLine(distance.ToString());
// if is count < like bl and distance > delta create new cluster
// add new cluster pair (x,y), increment n
if (clusterPairs.Count < this.blx && distance > this.deltaX)
{
this.CreateNewClusters(sample, this);
}
// else update xj cluster and yj cluster using amnesic average
else
{
// add sample to clusters, update statistics of clusters
nearestCluster.AddItem(sample);
}
// spawn if necessary
// if 2(n - q)/q2 > bs spawn to q children
// use k-means alg
if (this.GetNSPP() > Params.bs)
{
#region Swapping evaluation log
//for (int i = 0; i < 10; i++)
//{
// this.EvaluateSwap();
// Console.WriteLine("Round" + i.ToString());
// for (int j = 0; j < Params.q; j++)
// {
// Node node = new Node(this);
// List<ClusterPair> clPairs = this.clusterPairs.Where(cp => cp.CurrentCenter == j).ToList();
// Console.WriteLine("Region" + j.ToString() + ": " + clPairs.Count.ToString());
// }
//}
#endregion
if (Params.SwapType == 1)
{
this.EvaluateSwap();
this.Swap();
}
if (Params.SwapType == 2)
{
this.EvaluateSwap();
this.Swap_Modified();
}
if (Params.SwapType == 3)
{
for (int i = 0; i < 10; i++)
{
if (this.KMeansPlusPlusClustering())
{
break;
}
if (i == 9)
{
throw new InvalidOperationException("Not successfull keans clustering.");
}
}
this.EvaluateSwap();
this.Swap();
}
if (Params.SwapType == 4)
{
for (int i = 0; i < 10; i++)
{
if (this.KMeansClusteringY())
{
break;
}
if (i == 9)
{
throw new InvalidOperationException("Not successfull keans clustering.");
}
}
this.EvaluateSwap();
this.Swap();
}
// count most discriminating features space etc.
this.CountC();
this.CountGSOManifold();
this.CountMDFOfVectors();
this.CountMDFMeans();
this.CountCovarianceMatricesMDF();
//count cov matrix mean
this.CountCovarianceMatrixMeanMDF();
this.CountMeanAndVarianceMDF();
// dispose cluster items
if (!Params.StoreItems)
{
this.DisposeClustersItems();
}
}
}
public Vector GetOutputFromKnownSamples(Sample sample)
{
if (!this.outputs.ContainsKey(sample.Label))
{
this.outputs[sample.Label] = new MappedValue(){ Mean = new Vector(sample.X.Values.ToArray()), Count = 1};
}
else
{
this.outputs[sample.Label].Count++;
int count = this.outputs[sample.Label].Count;
this.outputs[sample.Label].Mean.Multiply(((double)count-1.0)/(double)count);
Vector addPart = new Vector(sample.X.Values.ToArray());
addPart.Multiply(1 / (double)count);
this.outputs[sample.Label].Mean.Add(addPart);
}
return new Vector(this.outputs[sample.Label].Mean.Values.ToArray());
}
/// <summary>
/// get nearest cluster pair by most dicrimnating features vector
/// </summary>
/// <param name="sample"></param>
/// <param name="distance"></param>
/// <returns></returns>
private ClusterPair GetNearestClusterPairXMDF(Sample sample, out double distance, out int index)
{
ILArray<double> thisVector = sample.X.Values.ToArray();
ILArray<double> scaterPart = thisVector - C;
ILArray<double> vector = ILMath.multiply(this.gSOManifold.T, scaterPart.ToArray());
distance = double.MaxValue;
ClusterPair closestPair = clusterPairs[0];
int i = 0;
index = -1;
foreach (ClusterPair item in clusterPairs)
{
double newDistance = item.X.GetMDFDistanceFromMDFMean(vector);
//Console.WriteLine("Distance in MDF : " + newDistance.ToString());
if (newDistance < distance)
{
distance = newDistance;
closestPair = item;
index = i;
}
i++;
}
return closestPair;
}
public void AddItem(Sample s)
{
this.X.AddItem(s.X, s.Label);
this.Y.AddItem(s.Y, s.Label);
this.samples.Add(s);
}
public ClusterPair GetTestResultByWidthSearch(Sample item)
{
ClusterPairTestResult result = new ClusterPairTestResult() { Distance = double.MaxValue, ClusterPair = new ClusterPair()};
this.CountClosestClusterPairByWidthSearch(item, result);
return result.ClusterPair;
}
/// <summary>
/// Update tree with sample
/// </summary>
/// <param name="sample">added sample to tree</param>
public void UpdateTree(Sample sample)
{
this.root.UpdateNode(sample);
}
public void UpdateNode(Sample sample)
{
//Console.WriteLine("Add sample " + count.ToString());
// add sample (because of counting of output)
if (Params.StoreSamples)
{
this.samples.Add(sample);
}
// count y of sample, if it is null
if (sample.Y == null)
{
throw new InvalidOperationException("Output Y of sample is null");
}
this.countOfSamples++;
if (this.isLeafNode)
{
// do leaf node staff
if (this.countOfSamples == 1)
{
// create new clusters and cluster pair
this.CreateNewClusters(sample, this);
}
else
{
// update cluster pairs
this.UpdateClusterPairsX(sample);
}
}
else
{
// update y clusters
// q count of cluster, dy resolution
// 1. find nearest y cluster, euclidean distance
// 2. if n < q and dy > distance, increment n, add new cluster y
// else update p ( e.g p = 0,2 -> 20% ) nearest cluster using amnesic average
// return nearest cluster
// update x cluster associated with returned y, mean with amnesic average
//if (this.isPlastic)
//{
this.UpdateClusters(sample);
//}
double distance = 0;
int index = 0;
ClusterPair nearestClPair = this.GetNearestClusterPairXBySDNLL_MDF(sample, out distance, out index);
Node next = nearestClPair.CorrespondChild;
next.UpdateNode(sample);
}
}
public TestResult GetTestResultByWidthSearch(Sample item)
{
ClusterPair resultClusterPair = root.GetTestResultByWidthSearch(item);
return new TestResult()
{
ClusterMeanX = resultClusterPair.X.Mean,
ClusterMeanY = resultClusterPair.Y.Mean,
Label = resultClusterPair.X.Label
};
}
public void UpdateNode_ForSwapping(Sample sample)
{
this.countOfSamples++;
if (this.isLeafNode)
{
// do leaf node staff
if (this.countOfSamples == 1)
{
// create new clusters and cluster pair
this.CreateNewClusters(sample, this);
}
else
{
// update cluster pairs
this.UpdateClusterPairsX_ForSwapping(sample);
}
}
}
public List<Node> GetNodesToSearch(Sample sample)
{
List<ClusterPair> clusterPairs = this.GetClosestClusterPairsAndSDNLLDistances_MDF(sample);
return clusterPairs.GroupBy(item => item.CorrespondChild.Id).Select(item => item.First().CorrespondChild).ToList();
}
private ClusterPair GetNearestClusterPairY(Sample sample, out double distance)
{
distance = double.MaxValue;
ClusterPair closestPair = clusterPairs[0];
foreach (ClusterPair item in clusterPairs)
{
double newDistance = item.Y.Mean.GetDistance(sample.Y);
if (item.Y.Mean.GetDistance(sample.Y) < distance)
{
distance = newDistance;
closestPair = item;
}
}
return closestPair;
}
