public ClusteringResult Cluster(IUnlabeledExampleCollection <SparseVector <double> > dataset)
{
Utils.ThrowException(dataset == null ? new ArgumentNullException("dataset") : null);
Utils.ThrowException(dataset.Count < NumLeaves ? new ArgumentValueException("dataset") : null);
ClusteringResult clusters = mKMeansClustering.Cluster(dataset);
UnlabeledDataset <SparseVector <double> > centroids = new UnlabeledDataset <SparseVector <double> >();
foreach (Cluster cluster in clusters.Roots)
{
SparseVector <double> centroid = ModelUtils.ComputeCentroid(cluster.Items, dataset, CentroidType.NrmL2);
centroids.Add(centroid);
centroid = Trim(centroid, 1000, 0.8);
cluster.ClusterInfo = 1; // cluster level
}
SparseMatrix <double> simMtx = ModelUtils.GetDotProductSimilarity(centroids, /*thresh=*/ 0, /*fullMatrix=*/ false);
SparseMatrix <double> clustMtxTr = ModelUtils.GetTransposedMatrix(centroids);
int iter = 1;
while (clusters.Roots.Count > 1)
{
Console.WriteLine("Iteration {0} ...", iter++);
int idx1, idx2;
FindMaxSim(simMtx, out idx1, out idx2);
Update(simMtx, clustMtxTr, clusters.Roots.Count, idx1, idx2, clusters.Roots.Inner, dataset, /*damping=*/ 0.9);
Console.WriteLine(simMtx.ToString("E0.00"));
Console.WriteLine();
}
return(clusters);
}
private void Update(SparseMatrix <double> simMtx, SparseMatrix <double> clustMtxTr, int numClusters, int idx1, int idx2, ArrayList <Cluster> clusters,
IUnlabeledExampleCollection <SparseVector <double> > dataset, double damping)
{
Debug.Assert(idx1 < idx2);
// create new parent
Cluster c1 = clusters[idx1];
Cluster c2 = clusters[idx2];
Cluster parent = new Cluster();
parent.Items.AddRange(c1.Items);
parent.Items.AddRange(c2.Items);
parent.ClusterInfo = Math.Max((int)c1.ClusterInfo, (int)c2.ClusterInfo) + 1;
c1.Parent = parent;
c2.Parent = parent;
parent.AddChild(c1);
parent.AddChild(c2);
SparseVector <double> centroid = ModelUtils.ComputeCentroid(parent.Items, dataset, CentroidType.NrmL2);
centroid = Trim(centroid, 1000, 0.8);
// remove clusters
clusters.RemoveAt(idx2);
clusters.RemoveAt(idx1);
// add new parent
clusters.Add(parent);
// remove rows at idx1 and idx2
simMtx.PurgeRowAt(idx2);
simMtx.PurgeRowAt(idx1);
// remove cols at idx1 and idx2
simMtx.PurgeColAt(idx2);
simMtx.PurgeColAt(idx1);
clustMtxTr.PurgeColAt(idx2);
clustMtxTr.PurgeColAt(idx1);
// update matrices
numClusters -= 2;
foreach (IdxDat <double> item in centroid)
{
if (clustMtxTr[item.Idx] == null)
{
clustMtxTr[item.Idx] = new SparseVector <double>(new IdxDat <double>[] { new IdxDat <double>(numClusters, item.Dat) });
}
else
{
clustMtxTr[item.Idx].InnerIdx.Add(numClusters);
clustMtxTr[item.Idx].InnerDat.Add(item.Dat);
}
}
double[] simVec = ModelUtils.GetDotProductSimilarity(clustMtxTr, numClusters + 1, centroid);
for (int i = 0; i < simVec.Length; i++)
{
simVec[i] *= Math.Pow(damping, (double)((int)parent.ClusterInfo + (int)clusters[i].ClusterInfo) / 2.0);
}
SparseMatrix <double> col = new SparseMatrix <double>();
col[0] = new SparseVector <double>(simVec);
simMtx.AppendCols(col.GetTransposedCopy(), numClusters);
}
public Prediction <LblT> Predict(SparseVector <double> example)
{
Utils.ThrowException(mDatasetMtx == null ? new InvalidOperationException() : null);
Utils.ThrowException(example == null ? new ArgumentNullException("example") : null);
ArrayList <KeyDat <double, LblT> > tmp = new ArrayList <KeyDat <double, LblT> >(mLabels.Count);
double[] dotProdSimVec = ModelUtils.GetDotProductSimilarity(mDatasetMtx, mLabels.Count, example);
for (int i = 0; i < mLabels.Count; i++)
{
tmp.Add(new KeyDat <double, LblT>(dotProdSimVec[i], mLabels[i]));
}
tmp.Sort(DescSort <KeyDat <double, LblT> > .Instance);
Dictionary <LblT, double> voting = new Dictionary <LblT, double>(mLblCmp);
int n = Math.Min(mK, tmp.Count);
double value;
if (mSoftVoting) // "soft" voting
{
for (int i = 0; i < n; i++)
{
KeyDat <double, LblT> item = tmp[i];
if (!voting.TryGetValue(item.Dat, out value))
{
voting.Add(item.Dat, item.Key);
}
else
{
voting[item.Dat] = value + item.Key;
}
}
}
else // normal voting
{
for (int i = 0; i < n; i++)
{
KeyDat <double, LblT> item = tmp[i];
if (!voting.TryGetValue(item.Dat, out value))
{
voting.Add(item.Dat, 1);
}
else
{
voting[item.Dat] = value + 1.0;
}
}
}
Prediction <LblT> classifierResult = new Prediction <LblT>();
foreach (KeyValuePair <LblT, double> item in voting)
{
classifierResult.Inner.Add(new KeyDat <double, LblT>(item.Value, item.Key));
}
classifierResult.Inner.Sort(DescSort <KeyDat <double, LblT> > .Instance);
return(classifierResult);
}
public Prediction <LblT> Predict(SparseVector <double> example)
{
Utils.ThrowException(mCentroidMtxTr == null ? new InvalidOperationException() : null);
Utils.ThrowException(example == null ? new ArgumentNullException("example") : null);
Prediction <LblT> result = new Prediction <LblT>();
double[] dotProdSimVec = ModelUtils.GetDotProductSimilarity(mCentroidMtxTr, mLabels.Count, example);
for (int i = 0; i < dotProdSimVec.Length; i++)
{
result.Inner.Add(new KeyDat <double, LblT>(dotProdSimVec[i], mLabels[i]));
}
result.Inner.Sort(DescSort <KeyDat <double, LblT> > .Instance);
return(result);
}
private double GetClusterQuality(IUnlabeledExampleCollection <SparseVector <double> > dataset, out SparseVector <double> centroid)
{
// compute centroid
centroid = ModelUtils.ComputeCentroid(dataset, CentroidType.NrmL2);
// compute intra-cluster similarities
double[] simData = ModelUtils.GetDotProductSimilarity(dataset, centroid);
// compute cluster quality
double quality = 0;
for (int i = 0; i < simData.Length; i++)
{
quality += simData[i];
}
quality /= (double)simData.Length;
return(quality);
}
internal void Assign(ArrayList <CentroidData> centroids, SparseMatrix <double> dataMtx, int instCount, int offs, out double clustQual)
{
int k = centroids.Count;
double[][] dotProd = new double[k][];
clustQual = 0;
int i = 0;
foreach (CentroidData cen in centroids)
{
SparseVector <double> cenVec = cen.GetSparseVector();
dotProd[i++] = ModelUtils.GetDotProductSimilarity(dataMtx, instCount, cenVec);
}
for (int instIdx = 0; instIdx < instCount; instIdx++)
{
double maxSim = double.MinValue;
ArrayList <int> candidates = new ArrayList <int>();
for (int cenIdx = 0; cenIdx < k; cenIdx++)
{
double sim = dotProd[cenIdx][instIdx];
if (sim > maxSim)
{
maxSim = sim;
candidates.Clear();
candidates.Add(cenIdx);
}
else if (sim == maxSim)
{
candidates.Add(cenIdx);
}
}
if (candidates.Count > 1)
{
candidates.Shuffle(mRnd);
}
centroids[candidates[0]].Items.Add(instIdx + offs);
clustQual += maxSim;
}
clustQual /= (double)instCount;
}
public void Train(ILabeledExampleCollection <LblT, SparseVector <double> > dataset)
{
Utils.ThrowException(dataset == null ? new ArgumentNullException("dataset") : null);
Utils.ThrowException(dataset.Count == 0 ? new ArgumentValueException("dataset") : null);
Dictionary <LblT, CentroidData> centroids = new Dictionary <LblT, CentroidData>(mLblCmp);
foreach (LabeledExample <LblT, SparseVector <double> > labeledExample in dataset)
{
if (!centroids.ContainsKey(labeledExample.Label))
{
CentroidData centroidData = new CentroidData();
centroidData.AddToSum(labeledExample.Example);
centroids.Add(labeledExample.Label, centroidData);
}
else
{
CentroidData centroidData = centroids[labeledExample.Label];
centroidData.AddToSum(labeledExample.Example);
}
}
foreach (CentroidData cenData in centroids.Values)
{
cenData.UpdateCentroidLen();
}
double learnRate = 1;
double[][] dotProd = null;
SparseMatrix <double> dsMtx = null;
if (mIterations > 0)
{
dotProd = new double[centroids.Count][];
dsMtx = ModelUtils.GetTransposedMatrix(ModelUtils.ConvertToUnlabeledDataset(dataset));
}
for (int iter = 1; iter <= mIterations; iter++)
{
mLogger.Info("Train", "Iteration {0} / {1} ...", iter, mIterations);
// compute dot products
mLogger.Info("Train", "Computing dot products ...");
int j = 0;
foreach (KeyValuePair <LblT, CentroidData> labeledCentroid in centroids)
{
mLogger.ProgressNormal(Logger.Level.Info, /*sender=*/ this, "Train", "Centroid {0} / {1} ...", j + 1, centroids.Count);
SparseVector <double> cenVec = labeledCentroid.Value.GetSparseVector();
dotProd[j] = ModelUtils.GetDotProductSimilarity(dsMtx, dataset.Count, cenVec);
j++;
}
// classify training examples
mLogger.Info("Train", "Classifying training examples ...");
int errCount = 0;
for (int instIdx = 0; instIdx < dataset.Count; instIdx++)
{
mLogger.ProgressFast(Logger.Level.Info, /*sender=*/ this, "Train", "Example {0} / {1} ...", instIdx + 1, dataset.Count);
double maxSim = double.MinValue;
CentroidData assignedCentroid = null;
CentroidData actualCentroid = null;
LabeledExample <LblT, SparseVector <double> > labeledExample = dataset[instIdx];
SparseVector <double> vec = labeledExample.Example;
int cenIdx = 0;
foreach (KeyValuePair <LblT, CentroidData> labeledCentroid in centroids)
{
double sim = dotProd[cenIdx][instIdx];
if (sim > maxSim)
{
maxSim = sim; assignedCentroid = labeledCentroid.Value;
}
if (labeledCentroid.Key.Equals(labeledExample.Label))
{
actualCentroid = labeledCentroid.Value;
}
cenIdx++;
}
if (assignedCentroid != actualCentroid)
{
assignedCentroid.AddToDiff(-learnRate, vec);
actualCentroid.AddToDiff(learnRate, vec);
errCount++;
}
}
mLogger.Info("Train", "Training set error rate: {0:0.00}%", (double)errCount / (double)dataset.Count * 100.0);
// update centroids
int k = 0;
foreach (CentroidData centroidData in centroids.Values)
{
mLogger.ProgressNormal(Logger.Level.Info, /*sender=*/ this, "Train", "Centroid {0} / {1} ...", ++k, centroids.Count);
centroidData.Update(mPositiveValuesOnly);
centroidData.UpdateCentroidLen();
}
learnRate *= mDamping;
}
mCentroidMtxTr = new SparseMatrix <double>();
mLabels = new ArrayList <LblT>();
int rowIdx = 0;
foreach (KeyValuePair <LblT, CentroidData> labeledCentroid in centroids)
{
mCentroidMtxTr[rowIdx++] = labeledCentroid.Value.GetSparseVector();
mLabels.Add(labeledCentroid.Key);
}
mCentroidMtxTr = mCentroidMtxTr.GetTransposedCopy();
}
private double[][] GetKernel(int rmvFeatIdx)
{
int numSv = SvmLightLib.GetSupportVectorCount(mModelId);
// initialize matrix
double[][] kernel = new double[numSv][];
// compute linear kernel
SparseMatrix <double> m = new SparseMatrix <double>();
for (int i = 0; i < numSv; i++)
{
SparseVector <double> sv = GetSupportVector(i);
m[i] = sv;
}
if (rmvFeatIdx >= 0)
{
m.RemoveColAt(rmvFeatIdx);
}
SparseMatrix <double> mTr = m.GetTransposedCopy();
for (int i = 0; i < numSv; i++)
{
double[] innerProd = ModelUtils.GetDotProductSimilarity(mTr, numSv, m[i]);
kernel[i] = innerProd;
}
// compute non-linear kernel
switch (mKernelType)
{
case SvmLightKernelType.Polynomial:
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Pow(mKernelParamS * kernel[row][col] + mKernelParamC, mKernelParamD);
}
}
break;
case SvmLightKernelType.RadialBasisFunction:
double[] diag = new double[kernel.Length];
for (int i = 0; i < kernel.Length; i++)
{
diag[i] = kernel[i][i];
} // save diagonal
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Exp(-mKernelParamGamma * (diag[row] + diag[col] - 2.0 * kernel[row][col]));
}
}
break;
case SvmLightKernelType.Sigmoid:
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Tanh(mKernelParamS * kernel[row][col] + mKernelParamC);
}
}
break;
}
return(kernel);
}
public ClusteringResult Cluster(IUnlabeledExampleCollection <SparseVector <double> > dataset)
{
Utils.ThrowException(dataset == null ? new ArgumentNullException("dataset") : null);
Utils.ThrowException(dataset.Count < mK ? new ArgumentValueException("dataset") : null);
ClusteringResult clustering = null;
double globalBestClustQual = 0;
for (int trial = 1; trial <= mTrials; trial++)
{
mLogger.Info("Cluster", "Clustering trial {0} of {1} ...", trial, mTrials);
ArrayList <CentroidData> centroids = new ArrayList <CentroidData>(mK);
ArrayList <int> bestSeeds = null;
for (int i = 0; i < mK; i++)
{
centroids.Add(new CentroidData());
}
// select seed items
double minSim = double.MaxValue;
ArrayList <int> tmp = new ArrayList <int>(dataset.Count);
for (int i = 0; i < dataset.Count; i++)
{
tmp.Add(i);
}
for (int k = 0; k < 3; k++)
{
ArrayList <SparseVector <double> > seeds = new ArrayList <SparseVector <double> >(mK);
tmp.Shuffle(mRnd);
for (int i = 0; i < mK; i++)
{
seeds.Add(dataset[tmp[i]]);
}
// assess quality of seed items
double simAvg = 0;
foreach (SparseVector <double> seed1 in seeds)
{
foreach (SparseVector <double> seed2 in seeds)
{
if (seed1 != seed2)
{
simAvg += DotProductSimilarity.Instance.GetSimilarity(seed1, seed2);
}
}
}
simAvg /= (double)(mK * mK - mK);
//Console.WriteLine(simAvg);
if (simAvg < minSim)
{
minSim = simAvg;
bestSeeds = new ArrayList <int>(mK);
for (int i = 0; i < mK; i++)
{
bestSeeds.Add(tmp[i]);
}
}
}
for (int i = 0; i < mK; i++)
{
centroids[i].Items.Add(bestSeeds[i]);
centroids[i].Update(dataset);
centroids[i].UpdateCentroidLen();
}
double[][] dotProd = new double[mK][];
SparseMatrix <double> dsMtx = ModelUtils.GetTransposedMatrix(dataset);
// main loop
int iter = 0;
double bestClustQual = 0;
double clustQual;
while (true)
{
iter++;
mLogger.Info("Cluster", "Iteration {0} ...", iter);
clustQual = 0;
// assign items to clusters
int j = 0;
foreach (CentroidData cen in centroids)
{
SparseVector <double> cenVec = cen.GetSparseVector();
dotProd[j] = ModelUtils.GetDotProductSimilarity(dsMtx, dataset.Count, cenVec);
j++;
}
for (int instIdx = 0; instIdx < dataset.Count; instIdx++)
{
double maxSim = double.MinValue;
ArrayList <int> candidates = new ArrayList <int>();
for (int cenIdx = 0; cenIdx < mK; cenIdx++)
{
double sim = dotProd[cenIdx][instIdx];
if (sim > maxSim)
{
maxSim = sim;
candidates.Clear();
candidates.Add(cenIdx);
}
else if (sim == maxSim)
{
candidates.Add(cenIdx);
}
}
if (candidates.Count > 1)
{
candidates.Shuffle(mRnd);
}
if (candidates.Count > 0) // *** is this always true?
{
centroids[candidates[0]].Items.Add(instIdx);
clustQual += maxSim;
}
}
clustQual /= (double)dataset.Count;
mLogger.Info("Cluster", "Quality: {0:0.0000}", clustQual);
// check if done
if (iter > 1 && clustQual - bestClustQual <= mEps)
{
break;
}
bestClustQual = clustQual;
// compute new centroids
for (int i = 0; i < mK; i++)
{
centroids[i].Update(dataset);
centroids[i].UpdateCentroidLen();
}
}
if (trial == 1 || clustQual > globalBestClustQual)
{
globalBestClustQual = clustQual;
// save the result
clustering = new ClusteringResult();
for (int i = 0; i < mK; i++)
{
clustering.AddRoot(new Cluster());
clustering.Roots.Last.Items.AddRange(centroids[i].Items);
}
}
}
return(clustering);
}