def init_reviews(self):
print('init_reviews', time.strftime("%H:%M:%S"))
self.reviews = []
self.specific_reviews = []
self.generic_reviews = []
# for text_review in self.text_reviews:
# self.reviews.append(Review(text_review))
my_file = '/Users/fpena/tmp/reviews_hotel.pkl'
# my_file = '/Users/fpena/tmp/reviews_restaurant.pkl'
# my_file = '/Users/fpena/tmp/sentences_hotel.pkl'
# with open(my_file, 'wb') as write_file:
# pickle.dump(self.reviews, write_file, pickle.HIGHEST_PROTOCOL)
with open(my_file, 'rb') as read_file:
self.reviews = pickle.load(read_file)
# self.reviews = self.reviews
# for review in self.reviews:
# print(review)
cluster_labels = reviews_clusterer.cluster_reviews(self.reviews)
review_clusters =\
reviews_clusterer.split_list_by_labels(self.reviews, cluster_labels)
# print(cluster_labels)
self.specific_reviews = review_clusters[0]
self.generic_reviews = review_clusters[1]
self.all_nouns = context_utils.get_all_nouns(self.reviews)
context_utils.generate_stats(self.specific_reviews, self.generic_reviews)
def test_split_list_by_labels(self):
lst = ['a', 'b', 'c', 'd', 'e', 'f']
labels = [2, 0, 0, 1, 1, 0]
expected_matrix = [['b', 'c', 'f'], ['d', 'e'], ['a']]
actual_matrix = reviews_clusterer.split_list_by_labels(lst, labels)
self.assertItemsEqual(actual_matrix, expected_matrix)
def test_split_list_by_labels(self):
lst = ['a', 'b', 'c', 'd', 'e', 'f']
labels = [2, 0, 0, 1, 1, 0]
expected_matrix = [
['b', 'c', 'f'],
['d', 'e'],
['a']
]
actual_matrix = reviews_clusterer.split_list_by_labels(lst, labels)
self.assertItemsEqual(actual_matrix, expected_matrix)
def separate_reviews(self):
"""
Separates the reviews into specific and generic. The separation is done
by clustering
"""
# print('separating reviews', time.strftime("%H:%M:%S"))
cluster_labels = reviews_clusterer.cluster_reviews(self.reviews)
review_clusters =\
reviews_clusterer.split_list_by_labels(self.reviews, cluster_labels)
self.specific_reviews = review_clusters[0]
self.generic_reviews = review_clusters[1]
def init_reviews(self):
print('init_reviews', time.strftime("%H:%M:%S"))
# self.reviews = reviews
self.specific_reviews = []
self.generic_reviews = []
# for text_review in self.text_reviews:
# self.reviews.append(Review(text_review))
# my_file = '/Users/fpena/UCC/Thesis/projects/yelp/source/python/topicmodeling/context/reviews_hotel.pkl'
records_file = '/Users/fpena/UCC/Thesis/datasets/context/stuff/reviews_hotel_shuffled.json'
reviews_file = '/Users/fpena/UCC/Thesis/datasets/context/stuff/reviews_hotel_shuffled.pkl'
# my_file = '/Users/fpena/UCC/Thesis/datasets/context/stuff/reviews_restaurant_shuffled.pkl'
# my_file = '/Users/fpena/tmp/reviews_restaurant.pkl'
# my_file = '/Users/fpena/tmp/sentences_hotel.pkl'
# with open(my_file, 'wb') as write_file:
# pickle.dump(self.reviews, write_file, pickle.HIGHEST_PROTOCOL)
# self.records = ETLUtils.load_json_file(records_file)
#
# with open(reviews_file, 'rb') as read_file:
# self.reviews = pickle.load(read_file)[:100]
#
# print(self.records[50]['text'])
# print(self.reviews[50].text)
# self.reviews = self.reviews
# for review in self.reviews:
# print(review)
cluster_labels = reviews_clusterer.cluster_reviews(self.reviews)
review_clusters =\
reviews_clusterer.split_list_by_labels(self.reviews, cluster_labels)
# print(cluster_labels)
self.specific_reviews = review_clusters[0]
self.generic_reviews = review_clusters[1]
self.all_nouns = context_utils.get_all_nouns(self.reviews)
context_utils.generate_stats(self.specific_reviews, self.generic_reviews)
def calculate_recall_in_top_n(records,
recommender,
n,
num_folds,
split=None,
min_score=5.0,
cache_reviews=None,
reviews_type=None):
start_time = time.time()
if split is None:
split = 1 - (1 / float(num_folds))
# split = 0.984
total_recall = 0.
total_coverage = 0.
num_cycles = 0.0
for i in xrange(0, num_folds):
print('Fold', i)
print('started training', time.strftime("%Y/%d/%m-%H:%M:%S"))
start = float(i) / num_folds
cluster_labels = None
train_records, test_records = ETLUtils.split_train_test(records,
split=split,
start=start)
if cache_reviews:
train_reviews, test_reviews = ETLUtils.split_train_test(
cache_reviews, split=split, start=start)
if reviews_type is not None:
cluster_labels = reviews_clusterer.cluster_reviews(
test_reviews)
recommender.reviews = train_reviews
recommender.load(train_records)
print('finished training', time.strftime("%Y/%d/%m-%H:%M:%S"))
if cluster_labels is not None:
separated_records = reviews_clusterer.split_list_by_labels(
test_records, cluster_labels)
if reviews_type == 'specific':
test_records = separated_records[0]
if reviews_type == 'generic':
test_records = separated_records[1]
positive_reviews = \
[review for review in test_records if review['overall_rating'] >= min_score]
if len(positive_reviews) == 0:
continue
num_hits = 0.0
num_predictions = 0.0
for review in positive_reviews:
user_id = review['user_id']
item_id = review['offering_id']
if not recommender.has_context:
hit = calculate_is_a_hit(test_records, recommender, user_id,
item_id, n)
else:
text_review = review['text']
hit = calculate_is_a_hit(test_records, recommender, user_id,
item_id, n, text_review)
if hit is None:
continue
if hit:
num_hits += 1
num_predictions += 1
# print('num predictions: %d/%d %s' % (num_predictions, len(positive_reviews), time.strftime("%Y/%d/%m-%H:%M:%S")))
if num_predictions == 0:
continue
recommender.clear()
recall = num_hits / num_predictions
coverage = num_predictions / len(positive_reviews)
print('recall', recall, time.strftime("%Y/%d/%m-%H:%M:%S"))
print('coverage', coverage, time.strftime("%Y/%d/%m-%H:%M:%S"))
total_recall += recall
total_coverage += coverage
num_cycles += 1
final_recall = total_recall / num_cycles
final_coverage = total_coverage / num_cycles
execution_time = time.time() - start_time
print('Final Top N Precision: %f' % final_recall)
print('Final Coverage: %f' % final_coverage)
print("--- %s seconds ---" % execution_time)
result = {
'Top N': final_recall,
'Coverage': final_coverage,
'Execution time': execution_time
}
return result
def perform_cross_validation(
records, recommender, num_folds, cache_reviews=None, reviews_type=None):
start_time = time.time()
split = 1 - (1/float(num_folds))
total_mean_absolute_error = 0.
total_mean_square_error = 0.
total_coverage = 0.
num_cycles = 0
for i in range(0, num_folds):
print('Num cycles: %d' % i)
start = float(i) / num_folds
cluster_labels = None
train_records, test_records = ETLUtils.split_train_test(
records, split=split, start=start)
if cache_reviews:
train_reviews, test_reviews = ETLUtils.split_train_test(
cache_reviews, split=split, start=start)
if reviews_type is not None:
cluster_labels = reviews_clusterer.cluster_reviews(test_reviews)
recommender.reviews = train_reviews
recommender.load(train_records)
if cluster_labels is not None:
separated_records = reviews_clusterer.split_list_by_labels(
test_records, cluster_labels)
if reviews_type == 'specific':
test_records = separated_records[0]
if reviews_type == 'generic':
test_records = separated_records[1]
_, errors, num_unknown_ratings = predict_rating_list(recommender, test_records)
recommender.clear()
mean_absolute_error = MeanAbsoluteError.compute_list(errors)
root_mean_square_error = RootMeanSquareError.compute_list(errors)
num_samples = len(test_records)
coverage = float((num_samples - num_unknown_ratings) / num_samples)
# print('Total length:', len(test))
# print('Unknown ratings:', num_unknown_ratings)
# print('Coverage:', coverage)
if mean_absolute_error is not None:
total_mean_absolute_error += mean_absolute_error
total_mean_square_error += root_mean_square_error
total_coverage += coverage
num_cycles += 1
else:
print('Mean absolute error is None!!!')
final_mean_absolute_error = total_mean_absolute_error / num_cycles
final_root_squared_error = total_mean_square_error / num_cycles
final_coverage = total_coverage / num_cycles
execution_time = time.time() - start_time
print('Final mean absolute error: %f' % final_mean_absolute_error)
print('Final root mean square error: %f' % final_root_squared_error)
print('Final coverage: %f' % final_coverage)
print("--- %s seconds ---" % execution_time)
result = {
'MAE': final_mean_absolute_error,
'RMSE': final_root_squared_error,
'Coverage': final_coverage,
'Execution time': execution_time
}
return result
def calculate_recall_in_top_n(
records, recommender, n, num_folds, split=None, min_score=5.0,
cache_reviews=None, reviews_type=None):
start_time = time.time()
if split is None:
split = 1 - (1/float(num_folds))
# split = 0.984
total_recall = 0.
total_coverage = 0.
num_cycles = 0.0
for i in xrange(0, num_folds):
print('Fold', i)
print('started training', time.strftime("%Y/%d/%m-%H:%M:%S"))
start = float(i) / num_folds
cluster_labels = None
train_records, test_records = ETLUtils.split_train_test(
records, split=split, shuffle_data=False, start=start)
if cache_reviews:
train_reviews, test_reviews = ETLUtils.split_train_test(
cache_reviews, split=split, shuffle_data=False, start=start)
if reviews_type is not None:
cluster_labels = reviews_clusterer.cluster_reviews(test_reviews)
recommender.reviews = train_reviews
recommender.load(train_records)
print('finished training', time.strftime("%Y/%d/%m-%H:%M:%S"))
if cluster_labels is not None:
separated_records = reviews_clusterer.split_list_by_labels(
test_records, cluster_labels)
if reviews_type == 'specific':
test_records = separated_records[0]
if reviews_type == 'generic':
test_records = separated_records[1]
positive_reviews = \
[review for review in test_records if review['overall_rating'] >= min_score]
if len(positive_reviews) == 0:
continue
num_hits = 0.0
num_predictions = 0.0
for review in positive_reviews:
user_id = review['user_id']
item_id = review['offering_id']
if not recommender.has_context:
hit = calculate_is_a_hit(
test_records, recommender, user_id, item_id, n)
else:
text_review = review['text']
hit = calculate_is_a_hit(
test_records, recommender, user_id, item_id, n, text_review)
if hit is None:
continue
if hit:
num_hits += 1
num_predictions += 1
# print('num predictions: %d/%d %s' % (num_predictions, len(positive_reviews), time.strftime("%Y/%d/%m-%H:%M:%S")))
if num_predictions == 0:
continue
recommender.clear()
recall = num_hits / num_predictions
coverage = num_predictions / len(positive_reviews)
print('recall', recall, time.strftime("%Y/%d/%m-%H:%M:%S"))
print('coverage', coverage, time.strftime("%Y/%d/%m-%H:%M:%S"))
total_recall += recall
total_coverage += coverage
num_cycles += 1
final_recall = total_recall / num_cycles
final_coverage = total_coverage / num_cycles
execution_time = time.time() - start_time
print('Final Top N Precision: %f' % final_recall)
print('Final Coverage: %f' % final_coverage)
print("--- %s seconds ---" % execution_time)
result = {
'Top N': final_recall,
'Coverage': final_coverage,
'Execution time': execution_time
}
return result
