def solveForModeB(X, M, n, maxInner, epsilon, tol):
"""
Solve for the subproblem B = argmin (B >= 0) f(M)
Parameters
----------
X : the original tensor
M : the current CP factorization
n : the mode that we are trying to solve the subproblem for
epsilon : parameter to avoid dividing by zero
tol : the convergence tolerance (to
Returns
-------
M : the updated CP factorization
Phi : the last Phi(n) value
iter : the number of iterations before convergence / maximum
kktModeViolation : the maximum value of min(B(n), E - Phi(n))
"""
# Pi(n) = [A(N) kr A(N-1) kr ... A(n+1) kr A(n-1) kr .. A(1)]^T
Pi = tensorTools.calculatePi(X, M, n)
#print(M.U[n])
for iter in range(maxInner):
# Phi = (X(n) elem-div (B Pi)) Pi^T
Phi = tensorTools.calculatePhi(X, M.U[n], Pi, n, epsilon=epsilon)
#print(Phi)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(
np.minimum(M.U[n], 1 - Phi).flatten()))
if (kktModeViolation < tol):
break
# Do the multiplicative update
M.U[n] = np.multiply(M.U[n], Phi)
#print(" Mode={0}, Inner Iter={1}, KKT violation={2}".format(n, iter, kktModeViolation))
return M, Phi, iter, kktModeViolation
def solveForModeB(X, M, n, maxInner, epsilon, tol):
"""
Solve for the subproblem B = argmin (B >= 0) f(M)
Parameters
----------
X : the original tensor
M : the current CP factorization
n : the mode that we are trying to solve the subproblem for
epsilon : parameter to avoid dividing by zero
tol : the convergence tolerance (to
Returns
-------
M : the updated CP factorization
Phi : the last Phi(n) value
iter : the number of iterations before convergence / maximum
kktModeViolation : the maximum value of min(B(n), E - Phi(n))
"""
# Pi(n) = [A(N) kr A(N-1) kr ... A(n+1) kr A(n-1) kr .. A(1)]^T
Pi = tensorTools.calculatePi(X, M, n)
for iter in range(maxInner):
# Phi = (X(n) elem-div (B Pi)) Pi^T
Phi = tensorTools.calculatePhi(X, M.U[n], Pi, n, epsilon=epsilon)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(np.minimum(M.U[n], 1-Phi).flatten()))
if (kktModeViolation < tol):
break
# Do the multiplicative update
M.U[n] = np.multiply(M.U[n],Phi)
#print(" Mode={0}, Inner Iter={1}, KKT violation={2}".format(n, iter, kktModeViolation))
return M, Phi, iter, kktModeViolation
def solveForModeB1(X, M, n, maxInner, epsilon, tol, sita, Y1, lambta2):
"""
Solve for the subproblem B = argmin (B >= 0) f(M)
Parameters
----------
X : the original tensor
M : the current CP factorization
n : the mode that we are trying to solve the subproblem for
epsilon : parameter to avoid dividing by zero
tol : the convergence tolerance (to
Returns
-------
M : the updated CP factorization
Phi : the last Phi(n) value
iter : the number of iterations before convergence / maximum
kktModeViolation : the maximum value of min(B(n), E - Phi(n))
DemoU : the updated demographic U matrix
"""
# Pi(n) = [A(N) kr A(N-1) kr ... A(n+1) kr A(n-1) kr .. A(1)]^T
Pi = tensorTools.calculatePi(X, M, n)
#print 'Pi size', Pi.shape
#print 'pi='+str(Pi)
#print(M.U[n])
for iter in range(maxInner):
# Phi = (X(n) elem-div (B Pi)) Pi^T
#print X.vals.shape,X.shape
#print X.vals.flatten().shape
Phi = tensorTools.calculatePhi(X, M.U[n], Pi, n, epsilon=epsilon)
#print('phi'+str(Phi))
#print(Phi)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(
np.minimum(M.U[n], 1 - Phi).flatten()))
if (kktModeViolation < tol):
break
B = M.U[n]
#print B.shape
colNorm = np.apply_along_axis(np.linalg.norm, 0, B, 1)
zeroNorm = np.where(colNorm == 0)[0]
colNorm[zeroNorm] = 1
B = B / colNorm[np.newaxis, :]
tm = np.hstack((np.ones((B.shape[0], 1)), B))
Y1 = Y1.reshape((Y1.shape[0], 1))
derive = -1.0 * lambta2 / B.shape[0] * np.dot(
(Y1 - np.dot(tm, sita)), sita.T)
#print derive.shape
#print np.multiply(M.U[n],derive[:,1:]).shape
#print np.multiply(M.U[n],Phi).shape
M.U[n] = np.array(np.multiply(M.U[n], Phi)) - np.array(
(np.multiply(M.U[n], derive[:, 1:])))
#print 'after'
#print M.U[n][0]
#print(" Mode={0}, Inner Iter={1}, KKT violation={2}".format(n, iter, kktModeViolation))
return M, Phi, iter, kktModeViolation
def solveForModeB1(X, M, n, maxInner, epsilon, tol, Demog, DemoU, lambta1,
lambta4):
"""
Solve for the subproblem B = argmin (B >= 0) f(M)
Parameters
----------
X : the original tensor
M : the current CP factorization
n : the mode that we are trying to solve the subproblem for
epsilon : parameter to avoid dividing by zero
tol : the convergence tolerance (to
Demog : the demographic matrix shape : number of persons * fertures
DemoU : the estimated matrix decomposited from Demog matrix ,another matrix is X's factoe matrix U[0]
lambta1 is the parameter of docomposition of demographic information
lambta4 is the patameter of penalty item of demoU
Returns
-------
M : the updated CP factorization
Phi : the last Phi(n) value
iter : the number of iterations before convergence / maximum
kktModeViolation : the maximum value of min(B(n), E - Phi(n))
DemoU : the updated demographic U matrix
"""
# Pi(n) = [A(N) kr A(N-1) kr ... A(n+1) kr A(n-1) kr .. A(1)]^T
Pi = tensorTools.calculatePi(X, M, n)
#print 'pi='+str(Pi)
#print(M.U[n])
for iter in range(maxInner):
# Phi = (X(n) elem-div (B Pi)) Pi^T
Phi = tensorTools.calculatePhi(X, M.U[n], Pi, n, epsilon=epsilon)
#print('phi'+str(Phi))
#print(Phi)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(
np.minimum(M.U[n], 1 - Phi).flatten()))
if (kktModeViolation < tol):
break
llambta, tempB = __normalize_mode(M, n, 1)
#DemoU=__solveDemoU(M.U[n],100,epsilon,Demog,DemoU,lambta1)
#print 'demoU'
#print DemoU[0]
#M.U[n] = np.multiply(M.U[n],Phi)-(lambta1*np.multiply(M.U[n]/(np.sum(Pi,axis=0)),np.dot(np.dot(M.U[n],DemoU)-Demog,DemoU.T)))
# #M.U[n] = np.multiply(M.U[n],Phi)
M.U[n] = np.multiply(M.U[n], Phi) - (lambta1 * np.multiply(
M.U[n], np.dot(np.dot(tempB, DemoU) - Demog, DemoU.T)))
#print 'after'
#print M.U[n][0]
#print(" Mode={0}, Inner Iter={1}, KKT violation={2}".format(n, iter, kktModeViolation))
return M, Phi, iter, kktModeViolation
def __solveMode(self, Pi, B, C, n):
"""
Performs the inner iterations and checks for convergence.
"""
for innerI in range(self.maxInnerIters):
# Phi = (X(n) elem-div (B Pi)) Pi^T
Phi = tensorTools.calculatePhi(self.X, B, Pi, n, C=C)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(np.minimum(B, 1-Phi).flatten()))
if (kktModeViolation < self.convTol):
break
# Do the multiplicative update
B = np.multiply(B, Phi)
return B, (innerI+1), kktModeViolation
def solveForModeB1(X, M, n, maxInner, epsilon, tol,sita,Y1, lambta2):
"""
Solve for the subproblem B = argmin (B >= 0) f(M)
Parameters
----------
X : the original tensor
M : the current CP factorization
n : the mode that we are trying to solve the subproblem for
epsilon : parameter to avoid dividing by zero
tol : the convergence tolerance (to
Returns
-------
M : the updated CP factorization
Phi : the last Phi(n) value
iter : the number of iterations before convergence / maximum
kktModeViolation : the maximum value of min(B(n), E - Phi(n))
DemoU : the updated demographic U matrix
"""
# Pi(n) = [A(N) kr A(N-1) kr ... A(n+1) kr A(n-1) kr .. A(1)]^T
Pi = tensorTools.calculatePi(X, M, n)
#print 'pi='+str(Pi)
#print(M.U[n])
for iter in range(maxInner):
# Phi = (X(n) elem-div (B Pi)) Pi^T
Phi = tensorTools.calculatePhi(X, M.U[n], Pi, n, epsilon=epsilon)
#print('phi'+str(Phi))
#print(Phi)
# check for convergence that min(B(n), E - Phi(n)) = 0 [or close]
kktModeViolation = np.max(np.abs(np.minimum(M.U[n], 1-Phi).flatten()))
if (kktModeViolation < tol):
break
#DemoU=__solveDemoU(M.U[n],100,epsilon,Demog,DemoU,lambta1)
#print 'demoU'
#print DemoU[0]
#M.U[n] = np.multiply(M.U[n],Phi)-(lambta1*np.multiply(M.U[n]/(np.sum(Pi,axis=0)),np.dot(np.dot(M.U[n],DemoU)-Demog,DemoU.T)))
# #M.U[n] = np.multiply(M.U[n],Phi)
B=M.U[n]
colNorm = np.apply_along_axis(np.linalg.norm, 0, B, 1)
zeroNorm = np.where(colNorm == 0)[0]
colNorm[zeroNorm] = 1
B = B / colNorm[np.newaxis, :]
tm=np.hstack((np.ones((B.shape[0],1)),B))
Y1=Y1.reshape((Y1.shape[0],1))
derive=-lambta2*np.dot((Y1/(np.exp(np.multiply(np.dot(tm,sita),Y1))+np.ones((Y1.shape[0],1)))),sita.T)/tm.shape[0]
M.U[n] = np.multiply(M.U[n],Phi)-(np.multiply(M.U[n],derive[:,1:]))
return M, Phi, iter, kktModeViolation
