def estimate(self, iterations=300, corrupt_value=0.):
cost, updates = self.get_cost_ranking_updates(
corruption_level=corrupt_value)
train_da = theano.function([],
cost,
updates=updates,
givens={
self.x: self.train,
self.o: self.train
})
for iteration in range(iterations):
train_cost = train_da()
print ("iterations: %3d, cost: %.6f" % \
(iteration + 1, train_cost))
train_preds = self.predict(self.train)
#####################################
ns = [5, 10, 15]
precisions = np.zeros(len(ns))
recalls = np.zeros(len(ns))
ndcgs = np.zeros(len(ns))
maps = np.zeros(len(ns))
auc = 0
numberOfUsers = 0
for i in range(self._num_user):
train_data = self.train[i].eval()
missing_index = np.where(train_data == 0)
test_data = self.test[i].eval()
#groundTruth = np.where(test_data == 1)
groundTruth = np.where(test_data > 0.5)
if len(groundTruth[0]) == 0:
continue
if len(np.where(train_data > 0)) == 0:
continue
numberOfUsers += 1
train_predicted = train_preds[i].eval()
predicted = train_preds[i].eval()
missing_index = missing_index[0]
predicted_missingdata = predicted[missing_index]
index = np.argsort(predicted_missingdata)
missing_index = np.take(missing_index, index)
rankedList = missing_index[::-1] # reverse
precisions += PrecAt(rankedList, groundTruth[0], ns)
maps += MapAt(rankedList, groundTruth[0], ns)
ndcgs += nDCGAt(rankedList, groundTruth[0], ns)
numDropped = self._num_item - len(rankedList)
auc += AUC(rankedList, groundTruth[0], numDropped)
precisions /= float(numberOfUsers)
ndcgs /= float(numberOfUsers)
maps /= float(numberOfUsers)
auc /= float(numberOfUsers)
print ("iterations: %3d, pre5: %.6f, pre10: %.6f, pre15: %.6f, ndcg5: %.6f, ndcg10: %.6f, ndcg15: %.6f, map5: %.6f, map10: %.6f, map15: %.6f, auc: %.6f" % \
(iteration + 1, precisions[0], precisions[1], precisions[2], ndcgs[0], ndcgs[1], ndcgs[2], maps[0], maps[1], maps[2], auc))
f = open("./result/resultAE.txt", 'a')
data = ("iterations: %3d, pre5: %.6f, pre10: %.6f, pre15: %.6f, ndcg5: %.6f, ndcg10: %.6f, ndcg15: %.6f, map5: %.6f, map10: %.6f, map15: %.6f, auc: %.6f" % \
(iteration + 1, precisions[0], precisions[1], precisions[2], ndcgs[0], ndcgs[1], ndcgs[2], maps[0], maps[1], maps[2], auc))
f.write(data)
f.close()
def estimate(self, iterations=300, alpha_value=1, batch_size=943):
index = T.lscalar()
n_train_batches = self._num_user // batch_size
output_matrix = np.asarray(self.output.eval())
alpha_matrix = np.asarray(self.train.eval())
alpha_matrix = output_matrix - alpha_matrix # remains only imputated items
alpha_matrix = alpha_matrix * (1 - alpha_value) * (
-1) + 1 # imtutated items change to alpha and the others are 1
alpha_matrix = shared(np.matrix(alpha_matrix,
dtype=theano.config.floatX),
borrow=True)
cost, updates = self.get_cost_ranking_updates(corruption_level=0.)
train_da = theano.function(
[index],
cost,
updates=updates,
givens={
self.x:
self.train[index * batch_size:(index + 1) * batch_size],
self.o:
self.output[index * batch_size:(index + 1) * batch_size],
self.alpha:
alpha_matrix[index * batch_size:(index + 1) * batch_size]
})
for iteration in range(iterations):
train_cost = 0
for batch_index in range(n_train_batches):
train_cost = train_da(batch_index)
print ("iterations: %3d, cost: %.6f" % \
(iteration + 1, train_cost))
train_preds = self.predict(self.train)
#####################################
ns = [5, 10, 15]
precisions = np.zeros(len(ns))
recalls = np.zeros(len(ns))
ndcgs = np.zeros(len(ns))
maps = np.zeros(len(ns))
auc = 0
numberOfUsers = 0
for i in range(self._num_user):
train_data = self.train[i].eval()
missing_index = np.where(train_data == 0)
test_data = self.test[i].eval()
#groundTruth = np.where(test_data == 1)
groundTruth = np.where(test_data > 0.5)
if len(groundTruth[0]) == 0:
continue
if len(np.where(train_data > 0)) == 0:
continue
numberOfUsers += 1
train_predicted = train_preds[i].eval()
predicted = train_preds[i].eval()
missing_index = missing_index[0]
predicted_missingdata = predicted[missing_index]
index = np.argsort(predicted_missingdata)
missing_index = np.take(missing_index, index)
rankedList = missing_index[::-1] # reverse
precisions += PrecAt(rankedList, groundTruth[0], ns)
maps += MapAt(rankedList, groundTruth[0], ns)
ndcgs += nDCGAt(rankedList, groundTruth[0], ns)
numDropped = self._num_item - len(rankedList)
auc += AUC(rankedList, groundTruth[0], numDropped)
precisions /= float(numberOfUsers)
ndcgs /= float(numberOfUsers)
maps /= float(numberOfUsers)
auc /= float(numberOfUsers)
print ("iterations: %3d, pre5: %.6f, pre10: %.6f, pre15: %.6f, ndcg5: %.6f, ndcg10: %.6f, ndcg15: %.6f, map5: %.6f, map10: %.6f, map15: %.6f, auc: %.6f" % \
(iteration + 1, precisions[0], precisions[1], precisions[2], ndcgs[0], ndcgs[1], ndcgs[2], maps[0], maps[1], maps[2], auc))
f = open("./result/resultAE.txt", 'a')
data = ("iterations: %3d, pre5: %.6f, pre10: %.6f, pre15: %.6f, ndcg5: %.6f, ndcg10: %.6f, ndcg15: %.6f, map5: %.6f, map10: %.6f, map15: %.6f, auc: %.6f" % \
(iteration + 1, precisions[0], precisions[1], precisions[2], ndcgs[0], ndcgs[1], ndcgs[2], maps[0], maps[1], maps[2], auc))
f.write(data)
f.close()