def _buildRandomWalkScore(self,d):
epsilon = 0.00001
#TO CHECK : Should P be computed based on the cosine similarity matrix ??? I don't think so!
#P = matrix(toolBox.randomWalkTransitionProbability(self.UI))
#P = matrix(toolBox.AtoP(toolBox.cosineSimilarity(self.UI)))
P = matrix(toolBox.transitionProbability(self.UI))
Phi = self.ItemPrior
W = repeat(array([Phi]), Phi.shape[0], axis=0)
#TO BE TESTED
#W = 1.0/W
W = exp(-500*W)
P = matrix(array(P)*W)
R = matrix(zeros((self.nbItems, self.nbItems)), dtype=float64)
U = matrix(eye(self.nbItems, self.nbItems,dtype=float64))
#U = U / self.nbItems
diff = 1000
#count = 20
print 'Convergence ...'
while (diff > epsilon) :#and (count > 0):
sys.stdout.write('.')
oldRSum= sum(R)
R = d*P*R + (1-d)*U
diff = abs(oldRSum - sum(R)) / sum(R)
#count-=1
print diff
#for i in arange(R.shape[0]):
# R[i,i] = 1
print R
print sum(R,0)
self.R = array(R)
def _buildRandomWalkScore(self, d):
epsilon = 0.00001
#TO CHECK : Should P be computed based on the cosine similarity matrix ??? I don't think so!
#P = matrix(toolBox.randomWalkTransitionProbability(self.UI))
#P = matrix(toolBox.AtoP(toolBox.cosineSimilarity(self.UI)))
P = matrix(toolBox.transitionProbability(self.UI))
Phi = self.ItemPrior
W = repeat(array([Phi]), Phi.shape[0], axis=0)
#TO BE TESTED
#W = 1.0/W
W = exp(-500 * W)
P = matrix(array(P) * W)
R = matrix(zeros((self.nbItems, self.nbItems)), dtype=float64)
U = matrix(eye(self.nbItems, self.nbItems, dtype=float64))
#U = U / self.nbItems
diff = 1000
#count = 20
print 'Convergence ...'
while (diff > epsilon): #and (count > 0):
sys.stdout.write('.')
oldRSum = sum(R)
R = d * P * R + (1 - d) * U
diff = abs(oldRSum - sum(R)) / sum(R)
#count-=1
print diff
#for i in arange(R.shape[0]):
# R[i,i] = 1
print R
print sum(R, 0)
self.R = array(R)
def train(self):
#Split this methods into two parts to be able to quickly output Weight matrix for different values of alpha and teta
#Cosine computation
#self.computeSim('cos')
#Compute cost matrix
self.Ctot = array(self.computeCostMatrix())
#Compute afinity matrix
self.P = toolBox.transitionProbability(self.UI)
def train(self):
#Split this methods into two parts to be able to quickly output Weight matrix for different values of alpha and teta
#Cosine computation
#self.computeSim('cos')
#Compute cost matrix
self.Ctot = array(self.computeCostMatrix())
#Compute afinity matrix
self.P = toolBox.transitionProbability(self.UI)
def train(self): '''Different ways of computing the transition probability matrix can be considered''' #TO FIX : Should P be computed from the similarity matrix or directly as transition probability from the graph ? if (self.sim == 'cos'): self.Sim = toolBox.cosineSimilarity(self.UI) self.P = toolBox.AtoP(self.Sim) if (self.sim == 'cp'): self.Sim = toolBox.cosineSimilarity(self.UI) self.P = toolBox.AtoP(self.Sim) if (self.sim == 'bn'): self.P = toolBox.transitionProbability(self.UI)
def train(self):
'''Different ways of computing the transition probability matrix can be considered'''
#TO FIX : Should P be computed from the similarity matrix or directly as transition probability from the graph ?
if (self.sim == 'cos'):
self.Sim = toolBox.cosineSimilarity(self.UI)
self.P = toolBox.AtoP(self.Sim)
if (self.sim == 'cp'):
self.Sim = toolBox.cosineSimilarity(self.UI)
self.P = toolBox.AtoP(self.Sim)
if (self.sim == 'bn'):
self.P = toolBox.transitionProbability(self.UI)
self.P = transpose(self.P)
def train(self): print "Start training..." #debug_here() #self.Cost = repeat(reshape(self.ItemPrior, (self.ItemPrior.shape[0],1)), self.ItemPrior.shape[0], axis=0) self.Cost = repeat(reshape(self.ItemPrior/mean(self.ItemPrior), (1,self.ItemPrior.shape[0])), self.ItemPrior.shape[0], axis=0) print self.Cost.shape #debug_here() if (self.sim == 'cos'): self.Sim = toolBox.cosineSimilarity(self.UI) self.Pref = toolBox.AtoP(self.Sim) if (self.sim == 'bn'): self.Pref = toolBox.transitionProbability(self.UI) #self.P = toolBox.BRWWR_Comp(self.Pref, ones(self.Pref.shape), self.theta) self.P = toolBox.BRWWR_Comp(self.Pref, self.Cost, self.theta) #Transpose P to get a correct random walk with restart self.P = transpose(self.P)
def train(self):
print "Start training..."
#debug_here()
#self.Cost = repeat(reshape(self.ItemPrior, (self.ItemPrior.shape[0],1)), self.ItemPrior.shape[0], axis=0)
self.Cost = repeat(reshape(self.ItemPrior / mean(self.ItemPrior),
(1, self.ItemPrior.shape[0])),
self.ItemPrior.shape[0],
axis=0)
print self.Cost.shape
#debug_here()
if (self.sim == 'cos'):
self.Sim = toolBox.cosineSimilarity(self.UI)
self.Pref = toolBox.AtoP(self.Sim)
if (self.sim == 'bn'):
self.Pref = toolBox.transitionProbability(self.UI)
#self.P = toolBox.BRWWR_Comp(self.Pref, ones(self.Pref.shape), self.theta)
self.P = toolBox.BRWWR_Comp(self.Pref, self.Cost, self.theta)
#Transpose P to get a correct random walk with restart
self.P = transpose(self.P)
