def trainAutorec(testData, trainData, trainMask, unratedItemsMask, numOfRatings, unrated_items, GroundTruth,
itemCount, topN, m, h, r, o, l, p, s=False):
# prepare training
bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, lastRecall, \
lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = initializeEvaluations(topN)
userIdList_Train, allTrainData, allTestData, unratedTrainMask = dataprocess.prepareTrainAndTest(trainData, unratedItemsMask, testData)
with tf.Graph().as_default():
data, mask, y, cost = models.Autorec(itemCount, h, r)
# define optimizer
if o == 1:
optimizer = tf.train.GradientDescentOptimizer(l).minimize(cost)
elif o == 2:
optimizer = tf.train.RMSPropOptimizer(l).minimize(cost)
else:
optimizer = tf.train.AdamOptimizer(l).minimize(cost)
printTrigger = True
# Start training
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
for epoch in range(epochCount):
random.shuffle(userIdList_Train)
for batchId in range(int(len(userIdList_Train) / batchSize)):
start = batchId * batchSize
end = start + batchSize
batchData = []
batchMask = []
for i in range(start, end):
userId = userIdList_Train[i]
batchData.append(trainData[userId])
if s:
# random negative sampling
unrated = np.zeros(itemCount)
tmp = np.random.choice(unrated_items[userId], numOfRatings[userId], replace=False)
for j in tmp:
unrated[j] = 1
batchMask.append(unrated + trainMask[userId])
else:
batchMask.append(trainMask[userId])
batchData = np.array(batchData)
batchMask = np.array(batchMask)
c, _ = sess.run([cost, optimizer], feed_dict={data: batchData, mask: batchMask})
print("[epoch %d/%d]\tcost : %.4f" % (epoch + 1, epochCount, c))
userID_list = []
for userId in testData:
userID_list.append(userId)
# calculate accuracy measures
preidictedValues, predictedIndices = sess.run(tf.nn.top_k(y * mask, itemCount),
feed_dict={data: allTrainData,
mask: unratedTrainMask})
printTrigger, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, \
bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = \
evaluation.evaluation_topN (GroundTruth, predictedIndices, topN, bestPrecision, bestRecall, bestNDCG,
bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures,
lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, epoch, c)
if printTrigger == True:
print("%.4f %.4f %.4f %.4f %.4f" % (bestPrecision[0], bestRecall[0], bestNDCG[0], bestMRR[0], bestGlobalMRR))
else:
print("[" + p + " / sampling:" + str(s) + " Done...")
sess.close()
be="ML100k"
d="d1"
st=False
de=False
ran=0
a=0
b=0
# be d m p s st de r h o l ran a b precision recall ndcg mrr globalmrr cost epoch LAprecision LArecall LAndcg LAmrr LAglobalmrr
printResultsToFile(topN, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount,
bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, be, d, m, p,
s, st, de, r, h, o, l, ran, a, b)
def trainDAS(testData, trainData, trainData_i, trainMask, trainMask_i, unratedItemsMask, positiveMask, numOfRatings, unrated_items, GroundTruth, itemCount, topN, m, h, r, o, l, ran, a, b):
bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = initializeEvaluations(topN)
with tf.Graph().as_default():
data, mask, data_i, mask_i, _positiveMask, length_positive, y, cost = models.DualAPR_sigmoid (itemCount, h, r, ran, a, b)
if o == 1:
optimizer = tf.train.GradientDescentOptimizer(l).minimize(cost)
elif o == 2:
optimizer = tf.train.RMSPropOptimizer(l).minimize(cost)
else:
optimizer = tf.train.AdamOptimizer(l).minimize(cost)
# prepare training
userIdList_Train, allTrainData, allTestData, unratedTrainMask = dataprocess.prepareTrainAndTest(trainData, unratedItemsMask, testData)
printTrigger = True
# Start training
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer()) #initialize_all_variables
for epoch in range(epochCount):
random.shuffle(userIdList_Train)
for batchId in range(int(len(userIdList_Train) / batchSize)):
start = batchId * batchSize
end = start + batchSize
batchData = []
batchMask = []
batchData_i = []
batchMask_i = []
batchPositiveMask = []
len_positive = []
for i in range(start, end):
userId = userIdList_Train[i]
batchData.append(trainData[userId])
batchMask.append(trainMask[userId])
batchData_i.append(trainData_i[userId])
batchMask_i.append(trainMask_i[userId])
batchPositiveMask.append(positiveMask[userId])
len_positive.append(numOfRatings[userId])
batchData = np.array(batchData)
batchMask = np.array(batchMask)
batchData_i = np.array(batchData_i)
batchMask_i = np.array(batchMask_i)
batchPositiveMask = np.array(batchPositiveMask)
len_positive = np.array(len_positive)
c, _ = sess.run([cost, optimizer], feed_dict={data: batchData, mask: batchMask, data_i: batchData_i, mask_i: batchMask_i, _positiveMask: batchPositiveMask, length_positive: len_positive})
print("[epoch %d/%d]\tcost : %.4f" % (epoch + 1, epochCount, c))
# calculate accuracy measures
preidictedValues, predictedIndices = sess.run(tf.nn.top_k(y * mask, itemCount),
feed_dict={data: allTrainData,
mask: unratedTrainMask})
printTrigger, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = evaluation.evaluation_topN (GroundTruth, predictedIndices, topN, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, epoch, c)
if printTrigger == True:
print("%.4f %.4f %.4f %.4f %.4f" % (bestPrecision[0], bestRecall[0], bestNDCG[0], bestMRR[0], bestGlobalMRR))
else:
print("[a: %d, b: %.2f, rankingparam: %.2f] Done..." % (a, b, ran))
sess.close()
# be d m p s st de r h o l ran a b precision recall ndcg mrr globalmrr cost epoch LAprecision LArecall LAndcg LAmrr LAglobalmrr
be="ML100k"
d="d1"
p="None"
s=False
st=False
de=False
printResultsToFile(topN, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, be, d, m, p, s, st, de, r, h, o, l, ran, a, b)
def trainSVDpp(testData, trainData, trainMask, unratedItemsMask, numOfRatings, unrated_items,
GroundTruth, itemCount, topN, m, h, r, o, l, p, userCount, s = False):
bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures, \
lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = initializeEvaluations(topN)
userIdList_Train, allTrainData, allTestData, unratedTrainMask, test_ulist = dataprocess.prepareTrainAndTest(
trainData,unratedItemsMask,testData)
# print(userCount, itemCount)
with tf.Graph().as_default():
user_batch, item_batch, rate_batch, rating_list_batch, userImplicit_batch,\
embd_y, infer, cost = models.SVDpp(user_num=userCount, item_num=itemCount, r=r,
dim=DIM)
if o == 1:
optimizer = tf.train.GradientDescentOptimizer(l).minimize(cost)
elif o == 2:
optimizer = tf.train.RMSPropOptimizer(l).minimize(cost)
else:
optimizer = tf.train.AdamOptimizer(l).minimize(cost)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
for epoch in range(epochCount):
random.shuffle(userIdList_Train)
for batchId in range(int(len(userIdList_Train) / batchSize)):
start = batchId * batchSize
end = start + batchSize
batchUserId = []
batchItemId = []
batchRate = []
rated_itemsIndex = []
numOfRatings_list = []
userImplicit_list = []
for i in range(start, end):
userId = userIdList_Train[i]
for j in range(itemCount):
if (trainMask[userId][j] != 0):
batchUserId.append(userId)
batchItemId.append(j)
batchRate.append(trainData[userId][j])
rated_itemsIndex.append(np.nonzero(trainData[userId]))
numOfRatings_list.append(numOfRatings[userId])
for j in range(len(batchUserId)):
userImplicit = sess.run(embd_y, feed_dict={rating_list_batch:
rated_itemsIndex[j]})
userImplicit = np.reshape(userImplicit, (len(userImplicit[0]),
len(userImplicit[0][0])))
userImplicit = np.sum(userImplicit, axis=0)
for k in range(DIM):
userImplicit[k] = (userImplicit[k]/math.sqrt(numOfRatings_list[j]))
userImplicit_list.append(userImplicit)
c, _, predict = sess.run([cost, optimizer, infer], feed_dict={user_batch: batchUserId,
item_batch: batchItemId,
rate_batch: batchRate,
userImplicit_batch: userImplicit_list})
print("[epoch %d/%d]\tcost : %.4f" % (epoch + 1, epochCount, c))
predictRate = []
for u in test_ulist:
predictNumOfRate = numOfRatings[u]
predictRateitemId = np.nonzero(testData[u])
predictBatchUserId = []
predictBatchItemId = []
predictBatchUserImplicit = []
for i in range(itemCount):
predictBatchUserId.append(u)
predictBatchItemId.append(i)
tmp = sess.run(embd_y, feed_dict={rating_list_batch: predictRateitemId})
tmp = np.reshape(tmp, (len(tmp[0]), len(tmp[0][0])))
tmp = np.sum(tmp, axis=0)
for k in range(DIM):
tmp[k] = (tmp[k] / math.sqrt(predictNumOfRate))
for j in range(len(predictBatchUserId)):
predictBatchUserImplicit.append(tmp)
rate = sess.run([infer], feed_dict={user_batch: predictBatchUserId,
item_batch: predictBatchItemId,
userImplicit_batch: predictBatchUserImplicit
})
predictRate.append(list(rate[0]))
predictRate = np.asarray(predictRate)
predictedValues, predictedIndices = sess.run(tf.nn.top_k(predictRate * unratedTrainMask, itemCount))
printTrigger, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount, \
bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR = \
evaluation.evaluation_topN(GroundTruth, predictedIndices, topN, bestPrecision, bestRecall, bestNDCG,
bestMRR, bestGlobalMRR, bestCost, bestEpochCount, bestMeanForMeasures,
lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, epoch, c)
if printTrigger == True:
print("%.4f %.4f %.4f %.4f %.4f" % (
bestPrecision[0], bestRecall[0], bestNDCG[0], bestMRR[0], bestGlobalMRR))
else:
print("[" + p + " / sampling:" + str(s) + " Done...")
sess.close()
be = "ML100k"
d = "d1"
st = False
de = False
ran = 0
a = 0
b = 0
# be d m p s st de r h o l ran a b precision recall ndcg mrr globalmrr cost epoch LAprecision LArecall LAndcg LAmrr LAglobalmrr
printResultsToFile(topN, bestPrecision, bestRecall, bestNDCG, bestMRR, bestGlobalMRR, bestCost, bestEpochCount,
bestMeanForMeasures, lastRecall, lastPrecision, lastNDCG, lastMRR, lastGlobalMRR, be, d, m,
p, s, st, de, r, h, o, l, ran, a, b)