def run_EM_missing(self):
try:
for class_no in range(self.min_class_label, self.max_class_label):
y_observed, experty_observed = self.binary_y_experty(class_no)
#random initializations for this class label
weights = np.random.random(self.F)
alpha = np.random.random(self.E) #expert sensitivity
beta = np.random.random(self.E) #expert specificity
l =0
iter = 0
while iter < self.MAXITER:
# First iteration
if not iter:
l_old = 0
expertcombined = np.array([])
for e in xrange(self.E):
experty_observed[e][experty_observed[e] == -1] = randrange(self.min_class_label,self.max_class_label+1)
#self.Training_experty[e] = self.experty[e][:self.Training_instances]
for e in experty_observed:
expertcombined = np.append(expertcombined,experty_observed[e], axis=0)
expertcombined = np.reshape(expertcombined, (self.E, self.Training_instances))
y_predicted = np.average( expertcombined, axis=0)
y_average = y_predicted.copy()
#acc_MV = np.size(np.where((y_average.round())==y_observed))/float(self.Training_instances)
self.results['weights_mv'][class_no] = NR.logistic_regression(self.Training_x[:,1:].T,np.asarray(y_average).reshape(-1),verbose=False, MAXIT=10000)
self.results['weights_at'][class_no] = NR.logistic_regression(self.Training_x[:,1:].T,np.asarray(y_observed).reshape(-1),verbose=False, MAXIT=10000)
else :
l_old = l
w_old = weights
alpha_old = alpha
beta_old = beta
experty_learnt = self.learn_experty_missing(alpha_old, beta_old, y_observed)
for e in experty_observed:
missing_ids = np.where(self.experty[e] == -1)
for m in missing_ids:
experty_observed[e][m] = experty_learnt[e][m]
#print "experty :"
#pprint(experty_observed)
a = Utils.a_calculations(alpha_old, experty_observed,self.y_shape)
b = Utils.b_calculations(beta_old, experty_observed, self.y_shape)
# E-step
y_predicted = EM.Estep(self.Training_x, w_old, a, b)
y_predicted = np.asarray(y_predicted).reshape(-1)
# M-step
weights, alpha, beta = EM.Mstep(self.Training_x, y_predicted, experty_observed)
a = Utils.a_calculations(alpha, experty_observed, self.y_shape)
b = Utils.b_calculations(beta, experty_observed, self.y_shape)
l = self.calculate_loglikelihood(y_predicted, weights, a, b)
#acc_EM = np.size(np.where(y_observed==y_predicted.round()))/float(self.Training_instances)
diff = np.fabs(l-l_old)
if diff <= self.CONV_THRESH and l>=l_old : break
iter = iter+1
if self.verbose:
print "EM algorithm :","diff:",diff,"log:", l, "iteration:", iter
self.results['weights'][class_no] = weights
self.results['alpha'][class_no] = alpha.round(1)
self.results['beta'][class_no] = beta.round(1)
"""self.results['loglikelihood'][class_no] = l
self.results['EM_perf']['f1_Score'][class_no] = Utils.calculate_F1score(y_observed, y_predicted)
self.results['MV_perf']['f1_Score'][class_no] = Utils.calculate_F1score(y_observed, y_average)
self.results['EM_perf']['rmse'][class_no] = Utils.calculate_RMSE(y_observed, y_predicted)
self.results['MV_perf']['rmse'][class_no] = Utils.calculate_RMSE(y_observed, y_average)
self.results['experty'] [class_no] = experty_observed
fig = plt.figure()
ax = fig.add_subplot(111)
ax.set_title('class :' + str(class_no))
ax.set_ylim(-1,2)
ax.plot(y_observed,'ro-',y_predicted,'-b.')"""
if self.verbose:
print "alphacap :"
pprint (alpha.round(1))
print "betacap :"
pprint (beta.round(1))
print "weights :"
pprint (weights)
print "f1_Score of EM approach :"
print self.results['EM_perf']['f1_Score'][class_no]
print "f1_Score of majority voting approach :"
print self.results['MV_perf']['f1_Score'][class_no]
print "y"
print y_observed
print "y maj"
print y_average.round(2)
print "y pred"
print y_predicted.round(2)
print "Expert wrong percentage"
print self.expert_wrong_percentage
print '--'*30
except Exception, e:
raise
def run(self):
try:
for class_no in range(self.min_class_label, self.max_class_label):
#for class_no in range(0,3):
#class_no = 1
#print class_no
y_observed, experty_observed = self.binary_y_experty(class_no)
#random initializations for this class label
weights = np.random.random(self.F)
alpha = np.random.random(self.E) #expert sensitivity
beta = np.random.random(self.E) #expert specificity
l = 0
iter = 0
while iter < self.MAXITER:
# First iteration
if not iter:
l_old = 0
expertcombined = np.array([])
for e in experty_observed:
expertcombined = np.append(expertcombined,experty_observed[e], axis=0)
expertcombined = np.reshape(expertcombined, (self.E, self.Training_instances))
y_average = np.average( expertcombined, axis=0)
#y_average = y_predicted.copy()
mv_expert_combined = np.reshape(expertcombined,expertcombined.size,order='F').reshape(np.shape(expertcombined)[1],np.shape(expertcombined)[0])
y_predicted = np.array([])
for emv in mv_expert_combined:
y_predicted = np.append(y_predicted, np.bincount(emv.astype(int)).argmax())
acc_MV = np.size(np.where((y_average.round())==y_observed))/float(self.Training_instances)
"""
Classifier with MV as input
"""
self.results['weights_mv'][class_no] = NR.logistic_regression(self.Training_x[:,1:].T,np.asarray(y_predicted).reshape(-1),verbose=False, MAXIT=10000)
self.results['weights_avg'][class_no] = NR.logistic_regression(self.Training_x[:,1:].T,np.asarray(y_average).reshape(-1),verbose=False, MAXIT=10000)
self.results['weights_at'][class_no] = NR.logistic_regression(self.Training_x[:,1:].T,np.asarray(y_observed).reshape(-1),verbose=False, MAXIT=10000)
else :
l_old = l
w_old = weights
alpha_old = alpha
beta_old = beta
a = Utils.a_calculations(alpha_old, experty_observed,self.y_shape)
b = Utils.b_calculations(beta_old, experty_observed, self.y_shape)
# E-step
y_predicted = EM.Estep(self.Training_x, w_old, a, b)
y_predicted = np.asarray(y_predicted).reshape(-1)
# M-step
weights, alpha, beta = EM.Mstep(self.Training_x, y_predicted, experty_observed)
a = Utils.a_calculations(alpha, experty_observed, self.y_shape)
b = Utils.b_calculations(beta, experty_observed, self.y_shape)
l = self.calculate_loglikelihood(y_predicted, weights, a, b)
acc_EM = np.size(np.where(y_observed==y_predicted.round()))/float(self.Training_instances)
diff = np.fabs(l-l_old)
if diff <= self.CONV_THRESH and l>=l_old : break
iter = iter+1
if self.verbose:
print "EM algorithm :","diff:",diff,"log:", l, "iteration:", iter
self.results['weights'][class_no] = weights
self.results['alpha'][class_no] = alpha.round(1)
self.results['beta'][class_no] = beta.round(1)
if self.verbose:
print "alphacap :"
pprint (alpha.round(1))
print "betacap :"
pprint (beta.round(1))
print "weights :"
pprint (weights)
print "Accuracy of EM approach :"
print str(acc_EM)
print "Accuracy of majority voting approach :"
print str(acc_MV)
print "y"
print y_observed
print "y maj"
print y_average.round(2)
print "y pred"
print y_predicted.round(2)
print "Expert wrong percentage"
print self.expert_wrong_percentage
print '--'*30
except:
#print "Running EM again --"
raise
