def getClusters(self):
PASS = 1
matrix = csr_matrix(self.similarityMatrixBuilder)
continu = True
realClusters = []
clusters = []
nextMatrix = []
firstRound = True
while continu:
print "-" * 40
print "PASS #{0}".format(PASS)
PASS += 1
clustersNumber, clusters = LouvainClusterer.getOnePassLouvainCommunities(
matrix)
if firstRound:
firstRound = False
realClusters = clusters
else:
LouvainClusterer.updateRealClusters(realClusters, clusters)
if (clustersNumber == matrix.shape[0]):
continu = False
else:
nextMatrix = LouvainClusterer.buildNewSimilarityMatrix(
matrix, clusters, clustersNumber)
matrix = nextMatrix
return np.array(realClusters)
def getClusters(self) :
PASS=1
matrix=csr_matrix(self.similarityMatrixBuilder)
continu=True
realClusters=[]
clusters=[]
nextMatrix=[]
firstRound=True
while continu :
print "-"*40
print "PASS #{0}".format(PASS)
PASS+=1
clustersNumber,clusters=LouvainClusterer.getOnePassLouvainCommunities(matrix)
if firstRound :
firstRound=False
realClusters=clusters
else :
LouvainClusterer.updateRealClusters(realClusters,clusters)
if (clustersNumber==matrix.shape[0]) :
continu=False
else :
nextMatrix=LouvainClusterer.buildNewSimilarityMatrix(matrix,clusters,clustersNumber)
matrix=nextMatrix
return np.array(realClusters)
def getClusters(self):
PASS = 1
matrix = csr_matrix(self.similarityMatrixBuilder)
continu = True
realClusters = []
clusters = []
nextMatrix = []
firstRound = True
while continu:
print "-" * 40
print "PASS #{0}".format(PASS)
PASS += 1
# on recupere le nombre de clusters et le vecteur de mapping obtenus a l'aide de la matrice de similarité
clustersNumber, clusters = LouvainClusterer.getOnePassLouvainCommunities(
matrix)
# on charge nos donnees de clusters dans realClusters
if firstRound:
firstRound = False
realClusters = clusters
else:
LouvainClusterer.updateRealClusters(realClusters, clusters)
# on regarde si on a reussi a reduire le nombre de clusters
# si oui on réitère en modifiant la matrice de travail
# et ce jusqu'à ne plus arriver à réduire le nbr de clusters
if (clustersNumber == matrix.shape[0]):
continu = False
else:
nextMatrix = LouvainClusterer.buildNewSimilarityMatrix(
matrix, clusters, clustersNumber)
matrix = nextMatrix
return np.array(realClusters)
def getOnePassLouvainCommunities(matrix):
ITER = 1
matrixSize = matrix.shape[0]
sumsOfWeights = np.array([
sum(matrix.getrow(i).data) + sum(matrix.getcol(i).data)
for i in range(matrixSize)
])
totalSumOfWeight = sum(sumsOfWeights)
clusters = np.array(range(0, matrixSize))
modified = True
while modified:
print " ITER #{0}".format(ITER)
ITER += 1
modified = False
for i in range(0, matrixSize):
newCluster = LouvainClusterer.getArgMaxModularity(
matrix, clusters, sumsOfWeights, totalSumOfWeight, i)
if (newCluster != -1):
modified = True
clusters[i] = newCluster
newClusterIdentifiers = {}
clustersNumber = 0
for i in range(0, len(clusters)):
if (clusters[i] in newClusterIdentifiers):
clusters[i] = newClusterIdentifiers[clusters[i]]
else:
newClusterIdentifiers[clusters[i]] = clustersNumber
clusters[i] = clustersNumber
clustersNumber += 1
print "End of a Pass"
print "-" * 40
return clustersNumber, clusters
def getArgMaxModularity(matrix, clusters, sumsOfWeights, totalSumOfWeight,
i):
"""
This function returns -1 if there is no optimization of modularity
"""
DMCE = LouvainClusterer.getDeltaModularityCalculElements(
matrix, clusters, sumsOfWeights, totalSumOfWeight, i)
DMCEI = DMCE[clusters[i]]
maxDelta = 0
maxJ = -1
for j in matrix.getrow(i).indices:
delta = DMCE[clusters[j]] - DMCEI
if (delta > maxDelta):
maxDelta, maxJ = delta, j
for j in matrix.getcol(i).indices:
delta = DMCE[clusters[j]] - DMCEI
if (delta > maxDelta):
maxDelta, maxJ = delta, j
if (maxDelta == 0):
return -1
return clusters[maxJ]
def getArgMaxModularity(matrix,clusters,sumsOfWeights,totalSumOfWeight,i) :
"""
This function returns -1 if there is no optimization of modularity
"""
DMCE = LouvainClusterer.getDeltaModularityCalculElements(matrix,clusters,sumsOfWeights,totalSumOfWeight,i)
DMCEI=DMCE[clusters[i]]
maxDelta=0
maxJ=-1
for j in matrix.getrow(i).indices :
delta=DMCE[clusters[j]]-DMCEI
if (delta>maxDelta) :
maxDelta,maxJ=delta,j
for j in matrix.getcol(i).indices :
delta=DMCE[clusters[j]]-DMCEI
if (delta>maxDelta) :
maxDelta,maxJ=delta,j
if (maxDelta==0) :
return -1
return clusters[maxJ]
def getOnePassLouvainCommunities(matrix) :
ITER=1
matrixSize=matrix.shape[0]
sumsOfWeights=np.array([sum(matrix.getrow(i).data)+sum(matrix.getcol(i).data) for i in range(matrixSize)])
totalSumOfWeight=sum(sumsOfWeights)
clusters=np.array(range(0,matrixSize))
modified=True
while modified :
print " ITER #{0}".format(ITER)
ITER+=1
modified=False
for i in range(0,matrixSize) :
newCluster=LouvainClusterer.getArgMaxModularity(matrix,clusters,sumsOfWeights,totalSumOfWeight,i)
if (newCluster!=-1) :
modified=True
clusters[i]=newCluster
newClusterIdentifiers={}
clustersNumber=0
for i in range(0,len(clusters)) :
if (clusters[i] in newClusterIdentifiers) :
clusters[i]=newClusterIdentifiers[clusters[i]]
else :
newClusterIdentifiers[clusters[i]]=clustersNumber
clusters[i]=clustersNumber
clustersNumber+=1
print "End of a Pass"
print "-"*40
return clustersNumber, clusters
def getSimilarityOf2Clusters(matrix, clusters, clusterI, clusterJ):
iVertices = np.where(clusters == clusterI)[0]
jVertices = np.where(clusters == clusterJ)[0]
sumOfWeights = 0
for k in iVertices:
for l in jVertices:
sumOfWeights += LouvainClusterer.getFromMatrix(matrix, k, l)
return sumOfWeights
def buildNewSimilarityMatrix(matrix, clusters, clustersNumber):
nextMatrix = dok_matrix((clustersNumber, clustersNumber),
dtype=np.float)
for i in range(0, clustersNumber):
for j in range(i, clustersNumber):
nextMatrix[i, j] = LouvainClusterer.getSimilarityOf2Clusters(
matrix, clusters, i, j)
return csr_matrix(nextMatrix)
def getSimilarityOf2Clusters(matrix,clusters,clusterI,clusterJ) :
iVertices=np.where(clusters == clusterI)[0]
jVertices=np.where(clusters == clusterJ)[0]
sumOfWeights=0
for k in iVertices :
for l in jVertices :
sumOfWeights+=LouvainClusterer.getFromMatrix(matrix,k,l)
return sumOfWeights
def getOnePassLouvainCommunities(matrix):
ITER = 1
matrixSize = matrix.shape[0]
# on somme les valeurs en colonnes et en lignes pour chaque cluster de la matrice
# sumsOfWeights -> liste de matrixSize valeurs
sumsOfWeights = np.array([
sum(matrix.getrow(i).data) + sum(matrix.getcol(i).data)
for i in range(matrixSize)
])
# totalSumOfWeight -> somme de toutes les valeurs
totalSumOfWeight = sum(sumsOfWeights)
# on crée une liste de matrixSize valeurs pour representer les clusters
# au début on a autant de clusteur que d'elements de la matrice
clusters = np.array(range(0, matrixSize))
modified = True
# boucle pour agglomérer les elements entre eux
while modified:
print " ITER #{0}".format(ITER)
ITER += 1
modified = False
for i in range(0, matrixSize):
newCluster = LouvainClusterer.getArgMaxModularity(
matrix, clusters, sumsOfWeights, totalSumOfWeight, i)
if (newCluster != -1):
modified = True
clusters[i] = newCluster
# on aplique un reduce sur le vecteur clusters pour avoir les vrais clusters
# newClusterIdentifiers est une map d'association entre les clusters du vecteur cluster et les nouveaux clusters
newClusterIdentifiers = {}
clustersNumber = 0
for i in range(0, len(clusters)):
if (clusters[i] in newClusterIdentifiers):
clusters[i] = newClusterIdentifiers[clusters[i]]
else:
newClusterIdentifiers[clusters[i]] = clustersNumber
clusters[i] = clustersNumber
clustersNumber += 1
print "End of a Pass"
print "-" * 40
return clustersNumber, clusters
def buildNewSimilarityMatrix(matrix,clusters,clustersNumber) :
nextMatrix=dok_matrix((clustersNumber,clustersNumber),dtype=np.float)
for i in range(0,clustersNumber) :
for j in range(i,clustersNumber) :
nextMatrix[i,j]=LouvainClusterer.getSimilarityOf2Clusters(matrix,clusters,i,j)
return csr_matrix(nextMatrix)
