def main_evaluate():
I = my_i
records = ETLUtils.load_json_file(RECORDS_FILE)
# print('num_records', len(records))
test_file = RECORDS_FILE + '_test'
test_records = ETLUtils.load_json_file(test_file)
top_n_evaluator = TopNEvaluator(records, test_records, DATASET, 10, I)
top_n_evaluator.find_important_records()
# top_n_evaluator.initialize()
# records_to_predict_file = DATASET_FOLDER + 'generated/records_to_predict_' + DATASET + '.json'
top_n_evaluator.load_records_to_predict(RECORDS_TO_PREDICT_FILE)
predictions_file = GENERATED_FOLDER + 'predictions_' + DATASET + '.txt'
predictions = rmse_calculator.read_targets_from_txt(predictions_file)
# print('total predictions', len(predictions))
top_n_evaluator.evaluate(predictions)
# print('precision', top_n_evaluator.precision)
print('recall', top_n_evaluator.recall)
return top_n_evaluator.recall
def test_get_items_to_predict(self):
top_n_evaluator = TopNEvaluator(ratings, test_set)
top_n_evaluator.I = 4
top_n_evaluator.N = 2
top_n_evaluator.initialize()
items_to_predict = top_n_evaluator.get_records_to_predict()
predictions = [0] * len(test_set) * (top_n_evaluator.I + 1)
top_n_evaluator.evaluate(predictions)
print(items_to_predict)
for item in items_to_predict:
print(item)
def test_get_items_to_predict(self):
top_n_evaluator = TopNEvaluator(ratings, test_set)
top_n_evaluator.I = 4
top_n_evaluator.N = 2
top_n_evaluator.initialize()
items_to_predict = top_n_evaluator.get_records_to_predict()
predictions = [0] * len(test_set) * (top_n_evaluator.I + 1)
top_n_evaluator.evaluate(predictions)
print(items_to_predict)
for item in items_to_predict:
print(item)
class ContextTopNRunner(object):
def __init__(self):
self.records = None
self.original_records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.predictions = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = []
self.context_topics_map = None
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
def clear(self):
print('clear: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# self.records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = []
self.context_topics_map = None
if Constants.SOLVER == Constants.LIBFM:
os.remove(self.csv_train_file)
os.remove(self.csv_test_file)
os.remove(self.context_predictions_file)
os.remove(self.context_train_file)
os.remove(self.context_test_file)
os.remove(self.context_log_file)
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
gc.collect()
def create_tmp_file_names(self):
unique_id = uuid.uuid4().hex
prefix = Constants.GENERATED_FOLDER + unique_id + '_' + \
Constants.ITEM_TYPE
# prefix = constants.GENERATED_FOLDER + constants.ITEM_TYPE
print('unique id: %s' % unique_id)
self.csv_train_file = prefix + '_train.csv'
self.csv_test_file = prefix + '_test.csv'
self.context_predictions_file = prefix + '_predictions.txt'
self.context_train_file = self.csv_train_file + '.libfm'
self.context_test_file = self.csv_test_file + '.libfm'
self.context_log_file = prefix + '.log'
@staticmethod
def plant_seeds():
if Constants.RANDOM_SEED is not None:
print('random seed: %d' % Constants.RANDOM_SEED)
random.seed(Constants.RANDOM_SEED)
if Constants.NUMPY_RANDOM_SEED is not None:
print('numpy random seed: %d' % Constants.NUMPY_RANDOM_SEED)
numpy.random.seed(Constants.NUMPY_RANDOM_SEED)
def load(self):
print('load: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.original_records =\
ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
# ETLUtils.drop_fields(['tagged_words'], self.original_records)
print('num_records: %d' % len(self.original_records))
if not os.path.exists(Constants.USER_ITEM_MAP_FILE):
records = ETLUtils.load_json_file(Constants.RECORDS_FILE)
user_item_map = create_user_item_map(records)
with open(Constants.USER_ITEM_MAP_FILE, 'wb') as write_file:
pickle.dump(user_item_map, write_file, pickle.HIGHEST_PROTOCOL)
def shuffle(self):
print('shuffle: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
random.shuffle(self.original_records)
def get_records_to_predict_topn(self):
print('get_records_to_predict_topn: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
with open(Constants.USER_ITEM_MAP_FILE, 'rb') as read_file:
user_item_map = pickle.load(read_file)
self.top_n_evaluator = TopNEvaluator(self.records, self.test_records,
Constants.ITEM_TYPE, 10,
Constants.TOPN_NUM_ITEMS)
self.top_n_evaluator.initialize(user_item_map)
self.records_to_predict = self.top_n_evaluator.get_records_to_predict()
self.important_records = self.top_n_evaluator.important_records
self.test_records = None
gc.collect()
def get_records_to_predict_rmse(self):
print('get_records_to_predict_rmse: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.important_records = self.test_records
self.records_to_predict = self.test_records
self.test_records = None
gc.collect()
def get_records_to_predict(self):
if Constants.EVALUATION_METRIC == 'topn_recall':
self.get_records_to_predict_topn()
elif Constants.EVALUATION_METRIC == 'rmse':
self.get_records_to_predict_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def train_topic_model(self, cycle_index, fold_index):
if Constants.CACHE_TOPIC_MODEL:
print('loading topic model')
lda_based_context = topic_model_creator.load_topic_model(
cycle_index, fold_index)
else:
print('train topic model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
lda_based_context = LdaBasedContext(self.train_records)
lda_based_context.generate_review_corpus()
lda_based_context.build_topic_model()
lda_based_context.update_reviews_with_topics()
lda_based_context.get_context_rich_topics()
self.context_rich_topics = lda_based_context.context_rich_topics
print('Trained LDA Model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
return lda_based_context
def find_reviews_topics(self, lda_based_context):
print('find topics: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
lda_based_context.find_contextual_topics(self.train_records)
lda_based_context.find_contextual_topics(
self.important_records, Constants.TEXT_SAMPLING_PROPORTION)
self.context_topics_map = {}
for record in self.important_records:
topic_distribution = record[Constants.TOPICS_FIELD]
context_topics = {}
for i in self.context_rich_topics:
topic_id = 'topic' + str(i[0])
context_topics[topic_id] = topic_distribution[i[0]]
record[Constants.CONTEXT_TOPICS_FIELD] = context_topics
self.context_topics_map[record[Constants.REVIEW_ID_FIELD]] =\
context_topics
self.important_records = None
gc.collect()
def prepare_records_for_libfm(self):
print('prepare_records_for_libfm: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.headers = build_headers(self.context_rich_topics)
if Constants.USE_CONTEXT is True:
if Constants.REVIEW_TYPE == Constants.SPECIFIC or \
Constants.REVIEW_TYPE == Constants.GENERIC:
self.train_records = ETLUtils.filter_records(
self.train_records, Constants.PREDICTED_CLASS_FIELD,
[Constants.REVIEW_TYPE])
with open(self.csv_train_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.train_records:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
for topic in self.context_rich_topics:
context_topics = record[Constants.CONTEXT_TOPICS_FIELD]
# print('context_topics', context_topics)
row.append(context_topics['topic' + str(topic[0])])
writer.writerow(row)
self.train_records = None
gc.collect()
with open(self.csv_test_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.records_to_predict:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
for topic in self.context_rich_topics:
important_record = record[Constants.REVIEW_ID_FIELD]
context_topics =\
self.context_topics_map[important_record]
row.append(context_topics['topic' + str(topic[0])])
writer.writerow(row)
# self.records_to_predict = None
self.context_topics_map = None
self.context_rich_topics = None
gc.collect()
print('Exported CSV and JSON files: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
csv_files = [self.csv_train_file, self.csv_test_file]
print('num_cols', len(self.headers))
libfm_converter.csv_to_libfm(csv_files,
0, [1, 2], [],
',',
has_header=True,
suffix='.libfm')
print('Exported LibFM files: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
def predict_fastfm(self):
if Constants.USE_CONTEXT:
for record in self.records_to_predict:
important_record = record[Constants.REVIEW_ID_FIELD]
record[Constants.CONTEXT_TOPICS_FIELD] = \
self.context_topics_map[important_record]
all_records = self.train_records + self.records_to_predict
x_matrix, y_vector = fastfm_recommender.records_to_matrix(
all_records, self.context_rich_topics)
encoder = OneHotEncoder(categorical_features=[0, 1], sparse=True)
encoder.fit(x_matrix)
x_train = encoder.transform(x_matrix[:len(self.train_records)])
y_train = y_vector[:len(self.train_records)]
x_test = encoder.transform(x_matrix[len(self.train_records):])
if Constants.FASTFM_METHOD == 'mcmc':
# solver = mcmc.FMRegression(n_iter=num_iters, rank=num_factors)
solver = mcmc.FMRegression(rank=Constants.FM_NUM_FACTORS)
self.predictions = solver.fit_predict(x_train, y_train, x_test)
elif Constants.FASTFM_METHOD == 'als':
solver = als.FMRegression(rank=Constants.FM_NUM_FACTORS)
solver.fit(x_train, y_train)
self.predictions = solver.predict(x_test)
elif Constants.FASTFM_METHOD == 'sgd':
solver = sgd.FMRegression(rank=Constants.FM_NUM_FACTORS)
solver.fit(x_train, y_train)
self.predictions = solver.predict(x_test)
def predict_libfm(self):
print('predict: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
run_libfm(self.context_train_file, self.context_test_file,
self.context_predictions_file, self.context_log_file)
self.predictions = rmse_calculator.read_targets_from_txt(
self.context_predictions_file)
def predict(self):
if Constants.SOLVER == Constants.LIBFM:
self.prepare_records_for_libfm()
self.predict_libfm()
elif Constants.SOLVER == Constants.FASTFM:
self.predict_fastfm()
def evaluate_topn(self):
print('evaluate_topn: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.top_n_evaluator.evaluate(self.predictions)
recall = self.top_n_evaluator.recall
print('Recall: %f' % recall)
print('Specific recall: %f' % self.top_n_evaluator.specific_recall)
print('Generic recall: %f' % self.top_n_evaluator.generic_recall)
return recall
def evaluate_rmse(self):
print('evaluate_topn: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
true_values = [
record[Constants.RATING_FIELD]
for record in self.records_to_predict
]
rmse = rmse_calculator.calculate_rmse(true_values, self.predictions)
print('RMSE: %f' % rmse)
return rmse
def evaluate(self):
if Constants.EVALUATION_METRIC == 'topn_recall':
return self.evaluate_topn()
elif Constants.EVALUATION_METRIC == 'rmse':
return self.evaluate_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def perform_cross_validation(self):
print(Constants._properties)
self.plant_seeds()
total_metric = 0.0
# total_specific_recall = 0.0
# total_generic_recall = 0.0
total_cycle_time = 0.0
num_cycles = Constants.NUM_CYCLES
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
total_iterations = num_cycles * num_folds
split = 1 - (1 / float(num_folds))
self.load()
for i in range(num_cycles):
print('\n\nCycle: %d/%d' % ((i + 1), num_cycles))
if Constants.SHUFFLE_DATA:
self.shuffle()
self.records = copy.deepcopy(self.original_records)
for j in range(num_folds):
fold_start = time.time()
cv_start = float(j) / num_folds
print('\nFold: %d/%d' % ((j + 1), num_folds))
self.create_tmp_file_names()
self.train_records, self.test_records =\
ETLUtils.split_train_test_copy(
self.records, split=split, start=cv_start)
self.get_records_to_predict()
if Constants.USE_CONTEXT:
lda_based_context = self.train_topic_model(i, j)
self.find_reviews_topics(lda_based_context)
self.predict()
metric = self.evaluate()
# recall = self.top_n_evaluator.recall
# specific_recall = self.top_n_evaluator.specific_recall
# generic_recall = self.top_n_evaluator.generic_recall
total_metric += metric
# total_specific_recall += specific_recall
# total_generic_recall += generic_recall
fold_end = time.time()
fold_time = fold_end - fold_start
total_cycle_time += fold_time
self.clear()
print("Total fold %d time = %f seconds" % ((j + 1), fold_time))
metric_average = total_metric / total_iterations
# average_specific_recall = total_specific_recall / total_iterations
# average_generic_recall = total_generic_recall / total_iterations
average_cycle_time = total_cycle_time / total_iterations
print('average rmse: %f' % metric_average)
# print('average specific recall: %f' % average_specific_recall)
# print('average generic recall: %f' % average_generic_recall)
print('average cycle time: %f' % average_cycle_time)
print('End: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
#
results = copy.deepcopy(Constants._properties)
results[Constants.EVALUATION_METRIC] = metric_average
# results['specific_recall'] = average_specific_recall
# results['generic_recall'] = average_generic_recall
results['cycle_time'] = average_cycle_time
results['timestamp'] = time.strftime("%Y/%m/%d-%H:%M:%S")
write_results_to_csv(results)
write_results_to_json(results)
class ContextTopNRunner(object):
def __init__(self):
self.records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = None
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
self.no_context_predictions_file = None
self.no_context_train_file = None
self.no_context_test_file = None
self.no_context_log_file = None
def clear(self):
print('clear: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
self.records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = None
os.remove(self.csv_train_file)
os.remove(self.csv_test_file)
os.remove(self.context_predictions_file)
os.remove(self.context_train_file)
os.remove(self.context_test_file)
os.remove(self.context_log_file)
os.remove(self.no_context_predictions_file)
os.remove(self.no_context_train_file)
os.remove(self.no_context_test_file)
os.remove(self.no_context_log_file)
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
self.no_context_predictions_file = None
self.no_context_train_file = None
self.no_context_test_file = None
self.no_context_log_file = None
def create_tmp_file_names(self):
unique_id = uuid.uuid4().hex
prefix = GENERATED_FOLDER + unique_id + '_' + DATASET
self.csv_train_file = prefix + '_context_train.csv'
self.csv_test_file = prefix + '_context_test.csv'
self.context_predictions_file = prefix + '_context_predictions.txt'
self.context_train_file = self.csv_train_file + '.context.libfm'
self.context_test_file = self.csv_test_file + '.context.libfm'
self.context_log_file = prefix + '_context.log'
self.no_context_predictions_file =\
prefix + '_no_context_predictions.txt'
self.no_context_train_file = self.csv_train_file + '.no_context.libfm'
self.no_context_test_file = self.csv_test_file + '.no_context.libfm'
self.no_context_log_file = prefix + '_no_context.log'
def load(self):
print('load: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
self.records = ETLUtils.load_json_file(RECORDS_FILE)
print('num_records', len(self.records))
def shuffle(self):
print('shuffle: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
random.shuffle(self.records)
def split(self):
print('split: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
num_records = len(self.records)
num_split_records = int(float(SPLIT_PERCENTAGE)/100*num_records)
self.train_records = self.records[:num_split_records]
self.test_records = self.records[num_split_records:]
def export(self):
print('export: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
I = my_i
if REVIEW_TYPE:
self.records = ETLUtils.filter_records(
self.records, constants.PREDICTED_CLASS_FIELD, [REVIEW_TYPE])
self.test_records = ETLUtils.filter_records(
self.test_records, constants.PREDICTED_CLASS_FIELD,
[REVIEW_TYPE])
with open(USER_ITEM_MAP_FILE, 'rb') as read_file:
user_item_map = pickle.load(read_file)
self.top_n_evaluator = TopNEvaluator(
self.records, self.test_records, DATASET, 10, I)
self.top_n_evaluator.initialize(user_item_map)
self.records_to_predict = self.top_n_evaluator.get_records_to_predict()
# self.top_n_evaluator.export_records_to_predict(RECORDS_TO_PREDICT_FILE)
self.important_records = self.top_n_evaluator.important_records
def train_topic_model(self):
print('train topic model: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
# self.train_records = ETLUtils.load_json_file(TRAIN_RECORDS_FILE)
lda_based_context = LdaBasedContext(self.train_records)
lda_based_context.get_context_rich_topics()
self.context_rich_topics = lda_based_context.context_rich_topics
print('Trained LDA Model: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
# with open(TOPIC_MODEL_FILE, 'wb') as write_file:
# pickle.dump(lda_based_context, write_file, pickle.HIGHEST_PROTOCOL)
# with open(TOPIC_MODEL_FILE, 'rb') as read_file:
# lda_based_context = pickle.load(read_file)
self.context_rich_topics = lda_based_context.context_rich_topics
return lda_based_context
def find_reviews_topics(self, lda_based_context):
print('find topics: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
# self.records_to_predict =\
# ETLUtils.load_json_file(RECORDS_TO_PREDICT_FILE)
lda_based_context.find_contextual_topics(self.train_records)
lda_based_context.find_contextual_topics(self.records_to_predict)
# topics_map = {}
# lda_based_context.find_contextual_topics(self.important_records)
# for record in self.important_records:
# topics_map[record[constants.REVIEW_ID_FIELD]] =\
# record[constants.TOPICS_FIELD]
#
# for record in self.records_to_predict:
# topic_distribution = topics_map[record[constants.REVIEW_ID_FIELD]]
# for i in self.context_rich_topics:
# topic_id = 'topic' + str(i[0])
# record[topic_id] = topic_distribution[i[0]]
print('contextual test set size: %d' % len(self.records_to_predict))
self.headers = build_headers(self.context_rich_topics)
print('Exported contextual topics: %s' %
time.strftime("%Y/%d/%m-%H:%M:%S"))
return self.train_records, self.records_to_predict
def prepare(self):
print('prepare: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
contextual_train_set =\
ETLUtils.select_fields(self.headers, self.train_records)
contextual_test_set =\
ETLUtils.select_fields(self.headers, self.records_to_predict)
ETLUtils.save_csv_file(
self.csv_train_file, contextual_train_set, self.headers)
ETLUtils.save_csv_file(
self.csv_test_file, contextual_test_set, self.headers)
print('Exported CSV and JSON files: %s'
% time.strftime("%Y/%d/%m-%H:%M:%S"))
csv_files = [
self.csv_train_file,
self.csv_test_file
]
num_cols = len(self.headers)
context_cols = num_cols
print('num_cols', num_cols)
# print('context_cols', context_cols)
libfm_converter.csv_to_libfm(
csv_files, 0, [1, 2], range(3, context_cols), ',', has_header=True,
suffix='.no_context.libfm')
libfm_converter.csv_to_libfm(
csv_files, 0, [1, 2], [], ',', has_header=True,
suffix='.context.libfm')
print('Exported LibFM files: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
def predict(self):
print('predict: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
run_libfm(
self.no_context_train_file, self.no_context_test_file,
self.no_context_predictions_file, self.no_context_log_file)
run_libfm(
self.context_train_file, self.context_test_file,
self.context_predictions_file, self.context_log_file)
def evaluate(self):
print('evaluate: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
predictions = rmse_calculator.read_targets_from_txt(
self.no_context_predictions_file)
self.top_n_evaluator.evaluate(predictions)
no_context_recall = self.top_n_evaluator.recall
predictions = rmse_calculator.read_targets_from_txt(
self.context_predictions_file)
self.top_n_evaluator.evaluate(predictions)
context_recall = self.top_n_evaluator.recall
print('No context recall: %f' % no_context_recall)
print('Context recall: %f' % context_recall)
return context_recall, no_context_recall
def super_main_lda(self):
total_context_recall = 0.0
total_no_context_recall = 0.0
total_cycle_time = 0.0
num_iterations = 10
self.load()
self.split()
self.export()
for i in range(num_iterations):
cycle_start = time.time()
print('\nCycle: %d' % i)
lda_based_context = self.train_topic_model()
self.find_reviews_topics(lda_based_context)
self.prepare()
self.predict()
context_recall, no_context_recall = self.evaluate()
total_context_recall += context_recall
total_no_context_recall += no_context_recall
cycle_end = time.time()
cycle_time = cycle_end - cycle_start
total_cycle_time += cycle_time
print("Total cycle %d time = %f seconds" % (i, cycle_time))
average_context_recall = total_context_recall / num_iterations
average_no_context_recall = total_no_context_recall / num_iterations
average_cycle_time = total_cycle_time / num_iterations
improvement =\
(average_context_recall / average_no_context_recall - 1) * 100
print('average no context recall', average_no_context_recall)
print('average context recall', average_context_recall)
print('average improvement: %f2.3%%' % improvement)
print('average cycle time', average_cycle_time)
print('End: %s' % time.strftime("%Y/%d/%m-%H:%M:%S"))
class WordContextTopNRunner(object):
def __init__(self):
self.records = None
self.reviews = None
self.original_records = None
self.original_reviews = None
self.train_records = None
self.train_reviews = None
self.test_records = None
self.test_reviews = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.important_reviews = None
self.context_rich_topics = []
self.sense_groups = []
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
def clear(self):
print('clear: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# self.records = None
self.train_records = None
self.train_reviews = None
self.test_records = None
self.test_reviews = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.important_reviews = None
self.context_rich_topics = []
self.sense_groups = []
os.remove(self.csv_train_file)
os.remove(self.csv_test_file)
os.remove(self.context_predictions_file)
os.remove(self.context_train_file)
os.remove(self.context_test_file)
os.remove(self.context_log_file)
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
def create_tmp_file_names(self):
unique_id = uuid.uuid4().hex
prefix = Constants.GENERATED_FOLDER + unique_id + '_' + \
Constants.ITEM_TYPE
# prefix = constants.GENERATED_FOLDER + constants.ITEM_TYPE
print('unique id: %s' % unique_id)
self.csv_train_file = prefix + '_train.csv'
self.csv_test_file = prefix + '_test.csv'
self.context_predictions_file = prefix + '_predictions.txt'
self.context_train_file = self.csv_train_file + '.libfm'
self.context_test_file = self.csv_test_file + '.libfm'
self.context_log_file = prefix + '.log'
def load(self):
print('load: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.original_records = ETLUtils.load_json_file(Constants.RECORDS_FILE)
with open(Constants.REVIEWS_FILE, 'rb') as read_file:
self.original_reviews = pickle.load(read_file)
print('num_records: %d' % len(self.original_records))
for record, review in zip(self.original_records, self.original_reviews):
review.id = record[Constants.REVIEW_ID_FIELD]
review.rating = record[Constants.RATING_FIELD]
if not os.path.exists(Constants.USER_ITEM_MAP_FILE):
records = ETLUtils.load_json_file(Constants.RECORDS_FILE)
user_item_map = create_user_item_map(records)
with open(Constants.USER_ITEM_MAP_FILE, 'wb') as write_file:
pickle.dump(user_item_map, write_file, pickle.HIGHEST_PROTOCOL)
def shuffle(self):
print('shuffle: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# random.shuffle(self.records)
shuffled_records = []
shuffled_reviews = []
index_shuffle = range(len(self.original_records))
random.shuffle(index_shuffle)
for i in index_shuffle:
shuffled_records.append(self.original_records[i])
shuffled_reviews.append(self.original_reviews[i])
self.original_records = shuffled_records
self.original_reviews = shuffled_reviews
def export(self):
print('export: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
with open(Constants.USER_ITEM_MAP_FILE, 'rb') as read_file:
user_item_map = pickle.load(read_file)
self.top_n_evaluator = TopNEvaluator(
self.records, self.test_records, Constants.ITEM_TYPE, 10,
Constants.TOPN_NUM_ITEMS)
self.top_n_evaluator.initialize(user_item_map)
self.records_to_predict = self.top_n_evaluator.get_records_to_predict()
self.important_records = self.top_n_evaluator.important_records
self.important_reviews = [
review for review in self.test_reviews if review.rating == 5
]
def train_word_model(self):
print('train topic model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# lda_based_context = LdaBasedContext(self.train_records)
# lda_based_context.get_context_rich_topics()
# self.context_rich_topics = lda_based_context.context_rich_topics
word_based_context = WordBasedContext(self.train_reviews)
word_based_context.calculate_sense_group_ratios()
self.sense_groups = word_based_context.sense_groups
print('Trained LDA Model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
return word_based_context
def find_reviews_topics(self, word_based_context):
print('find topics: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# lda_based_context.find_contextual_topics(self.train_records)
for record, review in zip(self.train_records, self.train_reviews):
record[Constants.CONTEXT_WORDS_FIELD] =\
word_based_context.calculate_word_context(review)
for record, review in zip(self.important_records, self.important_reviews):
record[Constants.CONTEXT_WORDS_FIELD] =\
word_based_context.calculate_word_context(review)
topics_map = {}
for record in self.important_records:
topics_map[record[Constants.REVIEW_ID_FIELD]] =\
record[Constants.CONTEXT_WORDS_FIELD]
for record in self.records_to_predict:
word_distribution = topics_map[record[Constants.REVIEW_ID_FIELD]]
record[Constants.CONTEXT_WORDS_FIELD] = word_distribution
print('contextual test set size: %d' % len(self.records_to_predict))
print('Exported contextual topics: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
def prepare(self):
print('prepare: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.headers = build_headers(len(self.sense_groups))
if Constants.USE_CONTEXT is True:
for record in self.train_records:
record.update(record[Constants.CONTEXT_WORDS_FIELD])
for record in self.records_to_predict:
record.update(record[Constants.CONTEXT_WORDS_FIELD])
if Constants.FM_REVIEW_TYPE:
self.train_records = ETLUtils.filter_records(
self.train_records, Constants.PREDICTED_CLASS_FIELD,
[Constants.FM_REVIEW_TYPE])
# ETLUtils.drop_fields([Constants.TOPICS_FIELD], self.train_records)
ETLUtils.keep_fields(self.headers, self.train_records)
ETLUtils.keep_fields(self.headers, self.records_to_predict)
ETLUtils.save_csv_file(
self.csv_train_file, self.train_records, self.headers)
ETLUtils.save_csv_file(
self.csv_test_file, self.records_to_predict, self.headers)
print('Exported CSV and JSON files: %s'
% time.strftime("%Y/%m/%d-%H:%M:%S"))
csv_files = [
self.csv_train_file,
self.csv_test_file
]
print('num_cols', len(self.headers))
libfm_converter.csv_to_libfm(
csv_files, 0, [1, 2], [], ',', has_header=True,
suffix='.libfm')
print('Exported LibFM files: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
def predict(self):
print('predict: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
run_libfm(
self.context_train_file, self.context_test_file,
self.context_predictions_file, self.context_log_file)
def evaluate(self):
print('evaluate: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
predictions = rmse_calculator.read_targets_from_txt(
self.context_predictions_file)
self.top_n_evaluator.evaluate(predictions)
recall = self.top_n_evaluator.recall
print('Recall: %f' % recall)
print('Specific recall: %f' % self.top_n_evaluator.specific_recall)
print('Generic recall: %f' % self.top_n_evaluator.generic_recall)
return recall
def perform_cross_validation(self):
Constants.print_properties()
utilities.plant_seeds()
total_recall = 0.0
total_specific_recall = 0.0
total_generic_recall = 0.0
total_cycle_time = 0.0
num_cycles = Constants.NUM_CYCLES
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
total_iterations = num_cycles * num_folds
split = 1 - (1/float(num_folds))
self.load()
for i in range(num_cycles):
print('\n\nCycle: %d/%d' % ((i+1), num_cycles))
if Constants.SHUFFLE_DATA:
self.shuffle()
self.records = copy.deepcopy(self.original_records)
self.reviews = copy.deepcopy(self.original_reviews)
for j in range(num_folds):
fold_start = time.time()
cv_start = float(j) / num_folds
print('\nFold: %d/%d' % ((j+1), num_folds))
self.create_tmp_file_names()
self.train_records, self.test_records = \
ETLUtils.split_train_test_copy(
self.records, split=split, start=cv_start)
self.train_reviews, self.test_reviews = \
ETLUtils.split_train_test_copy(
self.reviews, split=split, start=cv_start)
self.export()
if Constants.USE_CONTEXT:
lda_based_context = self.train_word_model()
self.find_reviews_topics(lda_based_context)
self.prepare()
self.predict()
self.evaluate()
recall = self.top_n_evaluator.recall
specific_recall = self.top_n_evaluator.specific_recall
generic_recall = self.top_n_evaluator.generic_recall
total_recall += recall
total_specific_recall += specific_recall
total_generic_recall += generic_recall
fold_end = time.time()
fold_time = fold_end - fold_start
total_cycle_time += fold_time
self.clear()
print("Total fold %d time = %f seconds" % ((j+1), fold_time))
average_recall = total_recall / total_iterations
average_specific_recall = total_specific_recall / total_iterations
average_generic_recall = total_generic_recall / total_iterations
average_cycle_time = total_cycle_time / total_iterations
print('average recall: %f' % average_recall)
print('average specific recall: %f' % average_specific_recall)
print('average generic recall: %f' % average_generic_recall)
print('average cycle time: %f' % average_cycle_time)
print('End: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
results = Constants.get_properties_copy()
results['recall'] = average_recall
results['specific_recall'] = average_specific_recall
results['generic_recall'] = average_generic_recall
results['cycle_time'] = average_cycle_time
results['timestamp'] = time.strftime("%Y/%m/%d-%H:%M:%S")
if not os.path.exists(Constants.CSV_RESULTS_FILE):
with open(Constants.CSV_RESULTS_FILE, 'wb') as f:
w = csv.DictWriter(f, sorted(results.keys()))
w.writeheader()
w.writerow(results)
else:
with open(Constants.CSV_RESULTS_FILE, 'a') as f:
w = csv.DictWriter(f, sorted(results.keys()))
w.writerow(results)
class ContextTopNRunner(object):
def __init__(self):
self.records = None
self.original_records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.predictions = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = []
self.context_topics_map = None
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
def clear(self):
print('clear: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# self.records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = []
self.context_topics_map = None
if Constants.SOLVER == Constants.LIBFM:
os.remove(self.csv_train_file)
os.remove(self.csv_test_file)
os.remove(self.context_predictions_file)
os.remove(self.context_train_file)
os.remove(self.context_test_file)
os.remove(self.context_log_file)
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
gc.collect()
def create_tmp_file_names(self):
unique_id = uuid.uuid4().hex
prefix = Constants.GENERATED_FOLDER + unique_id + '_' + \
Constants.ITEM_TYPE
# prefix = constants.GENERATED_FOLDER + constants.ITEM_TYPE
print('unique id: %s' % unique_id)
self.csv_train_file = prefix + '_train.csv'
self.csv_test_file = prefix + '_test.csv'
self.context_predictions_file = prefix + '_predictions.txt'
self.context_train_file = self.csv_train_file + '.libfm'
self.context_test_file = self.csv_test_file + '.libfm'
self.context_log_file = prefix + '.log'
@staticmethod
def plant_seeds():
if Constants.RANDOM_SEED is not None:
print('random seed: %d' % Constants.RANDOM_SEED)
random.seed(Constants.RANDOM_SEED)
if Constants.NUMPY_RANDOM_SEED is not None:
print('numpy random seed: %d' % Constants.NUMPY_RANDOM_SEED)
numpy.random.seed(Constants.NUMPY_RANDOM_SEED)
def load(self):
print('load: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.original_records =\
ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
# ETLUtils.drop_fields(['tagged_words'], self.original_records)
print('num_records: %d' % len(self.original_records))
if not os.path.exists(Constants.USER_ITEM_MAP_FILE):
records = ETLUtils.load_json_file(Constants.RECORDS_FILE)
user_item_map = create_user_item_map(records)
with open(Constants.USER_ITEM_MAP_FILE, 'wb') as write_file:
pickle.dump(user_item_map, write_file, pickle.HIGHEST_PROTOCOL)
def shuffle(self):
print('shuffle: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
random.shuffle(self.original_records)
def get_records_to_predict_topn(self):
print(
'get_records_to_predict_topn: %s'
% time.strftime("%Y/%m/%d-%H:%M:%S")
)
with open(Constants.USER_ITEM_MAP_FILE, 'rb') as read_file:
user_item_map = pickle.load(read_file)
self.top_n_evaluator = TopNEvaluator(
self.records, self.test_records, Constants.ITEM_TYPE, 10,
Constants.TOPN_NUM_ITEMS)
self.top_n_evaluator.initialize(user_item_map)
self.records_to_predict = self.top_n_evaluator.get_records_to_predict()
self.important_records = self.top_n_evaluator.important_records
self.test_records = None
gc.collect()
def get_records_to_predict_rmse(self):
print(
'get_records_to_predict_rmse: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S")
)
self.important_records = self.test_records
self.records_to_predict = self.test_records
self.test_records = None
gc.collect()
def get_records_to_predict(self):
if Constants.EVALUATION_METRIC == 'topn_recall':
self.get_records_to_predict_topn()
elif Constants.EVALUATION_METRIC == 'rmse':
self.get_records_to_predict_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def train_topic_model(self, cycle_index, fold_index):
if Constants.CACHE_TOPIC_MODEL:
print('loading topic model')
lda_based_context = topic_model_creator.load_topic_model(
cycle_index, fold_index)
else:
print('train topic model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
lda_based_context = LdaBasedContext(self.train_records)
lda_based_context.generate_review_corpus()
lda_based_context.build_topic_model()
lda_based_context.update_reviews_with_topics()
lda_based_context.get_context_rich_topics()
self.context_rich_topics = lda_based_context.context_rich_topics
print('Trained LDA Model: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
return lda_based_context
def find_reviews_topics(self, lda_based_context):
print('find topics: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
lda_based_context.find_contextual_topics(self.train_records)
lda_based_context.find_contextual_topics(
self.important_records, Constants.TEXT_SAMPLING_PROPORTION)
self.context_topics_map = {}
for record in self.important_records:
topic_distribution = record[Constants.TOPICS_FIELD]
context_topics = {}
for i in self.context_rich_topics:
topic_id = 'topic' + str(i[0])
context_topics[topic_id] = topic_distribution[i[0]]
record[Constants.CONTEXT_TOPICS_FIELD] = context_topics
self.context_topics_map[record[Constants.REVIEW_ID_FIELD]] =\
context_topics
self.important_records = None
gc.collect()
def prepare_records_for_libfm(self):
print('prepare_records_for_libfm: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.headers = build_headers(self.context_rich_topics)
if Constants.REVIEW_TYPE == Constants.SPECIFIC or \
Constants.REVIEW_TYPE == Constants.GENERIC:
self.train_records = ETLUtils.filter_records(
self.train_records, Constants.PREDICTED_CLASS_FIELD,
[Constants.REVIEW_TYPE])
with open(self.csv_train_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.train_records:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
for topic in self.context_rich_topics:
context_topics = record[Constants.CONTEXT_TOPICS_FIELD]
# print('context_topics', context_topics)
row.append(context_topics['topic' + str(topic[0])])
writer.writerow(row)
self.train_records = None
gc.collect()
with open(self.csv_test_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.records_to_predict:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
for topic in self.context_rich_topics:
important_record = record[Constants.REVIEW_ID_FIELD]
context_topics =\
self.context_topics_map[important_record]
row.append(context_topics['topic' + str(topic[0])])
writer.writerow(row)
# self.records_to_predict = None
self.context_topics_map = None
self.context_rich_topics = None
gc.collect()
print('Exported CSV and JSON files: %s'
% time.strftime("%Y/%m/%d-%H:%M:%S"))
csv_files = [
self.csv_train_file,
self.csv_test_file
]
print('num_cols', len(self.headers))
libfm_converter.csv_to_libfm(
csv_files, 0, [1, 2], [], ',', has_header=True,
suffix='.libfm')
print('Exported LibFM files: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
def predict_fastfm(self):
if Constants.USE_CONTEXT:
for record in self.records_to_predict:
important_record = record[Constants.REVIEW_ID_FIELD]
record[Constants.CONTEXT_TOPICS_FIELD] = \
self.context_topics_map[important_record]
all_records = self.train_records + self.records_to_predict
x_matrix, y_vector = fastfm_recommender.records_to_matrix(
all_records, self.context_rich_topics)
encoder = OneHotEncoder(categorical_features=[0, 1], sparse=True)
encoder.fit(x_matrix)
x_train = encoder.transform(x_matrix[:len(self.train_records)])
y_train = y_vector[:len(self.train_records)]
x_test = encoder.transform(x_matrix[len(self.train_records):])
if Constants.FASTFM_METHOD == 'mcmc':
# solver = mcmc.FMRegression(n_iter=num_iters, rank=num_factors)
solver = mcmc.FMRegression(rank=Constants.FM_NUM_FACTORS)
self.predictions = solver.fit_predict(x_train, y_train, x_test)
elif Constants.FASTFM_METHOD == 'als':
solver = als.FMRegression(rank=Constants.FM_NUM_FACTORS)
solver.fit(x_train, y_train)
self.predictions = solver.predict(x_test)
elif Constants.FASTFM_METHOD == 'sgd':
solver = sgd.FMRegression(rank=Constants.FM_NUM_FACTORS)
solver.fit(x_train, y_train)
self.predictions = solver.predict(x_test)
def predict_libfm(self):
print('predict: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
run_libfm(
self.context_train_file, self.context_test_file,
self.context_predictions_file, self.context_log_file)
self.predictions = rmse_calculator.read_targets_from_txt(
self.context_predictions_file)
def predict(self):
if Constants.SOLVER == Constants.LIBFM:
self.prepare_records_for_libfm()
self.predict_libfm()
elif Constants.SOLVER == Constants.FASTFM:
self.predict_fastfm()
def evaluate_topn(self):
print('evaluate_topn: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.top_n_evaluator.evaluate(self.predictions)
recall = self.top_n_evaluator.recall
print('Recall: %f' % recall)
print('Specific recall: %f' % self.top_n_evaluator.specific_recall)
print('Generic recall: %f' % self.top_n_evaluator.generic_recall)
return recall, self.top_n_evaluator.specific_recall,\
self.top_n_evaluator.generic_recall
def evaluate_rmse(self):
print('evaluate_topn: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
true_values = [
record[Constants.RATING_FIELD] for record in self.records_to_predict
]
specific_true_values = []
specific_predictions = []
generic_true_values = []
generic_predictions = []
index = 0
for record, prediction in zip(
self.records_to_predict, self.predictions):
if record[Constants.PREDICTED_CLASS_FIELD] == 'specific':
specific_true_values.append(record[Constants.RATING_FIELD])
specific_predictions.append(prediction)
elif record[Constants.PREDICTED_CLASS_FIELD] == 'generic':
generic_true_values.append(record[Constants.RATING_FIELD])
generic_predictions.append(prediction)
index += 1
rmse = rmse_calculator.calculate_rmse(true_values, self.predictions)
specific_rmse = rmse_calculator.calculate_rmse(
specific_true_values, specific_predictions)
generic_rmse = rmse_calculator.calculate_rmse(
generic_true_values, generic_predictions)
print('RMSE: %f' % rmse)
print('Specific RMSE: %f' % specific_rmse)
print('Generic RMSE: %f' % generic_rmse)
return rmse, specific_rmse, generic_rmse
def evaluate(self):
if Constants.EVALUATION_METRIC == 'topn_recall':
return self.evaluate_topn()
elif Constants.EVALUATION_METRIC == 'rmse':
return self.evaluate_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def perform_cross_validation(self):
print(Constants._properties)
self.plant_seeds()
total_metric = 0.0
total_specific_metric = 0.0
total_generic_metric = 0.0
total_cycle_time = 0.0
num_cycles = Constants.NUM_CYCLES
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
total_iterations = num_cycles * num_folds
split = 1 - (1/float(num_folds))
self.load()
for i in range(num_cycles):
print('\n\nCycle: %d/%d' % ((i+1), num_cycles))
if Constants.SHUFFLE_DATA:
self.shuffle()
self.records = copy.deepcopy(self.original_records)
for j in range(num_folds):
fold_start = time.time()
cv_start = float(j) / num_folds
print('\nFold: %d/%d' % ((j+1), num_folds))
self.create_tmp_file_names()
self.train_records, self.test_records =\
ETLUtils.split_train_test_copy(
self.records, split=split, start=cv_start)
self.get_records_to_predict()
if Constants.USE_CONTEXT:
lda_based_context = self.train_topic_model(i, j)
self.find_reviews_topics(lda_based_context)
self.predict()
metric, specific_metric, generic_metric = self.evaluate()
total_metric += metric
total_specific_metric += specific_metric
total_generic_metric += generic_metric
fold_end = time.time()
fold_time = fold_end - fold_start
total_cycle_time += fold_time
self.clear()
print("Total fold %d time = %f seconds" % ((j+1), fold_time))
metric_name = Constants.EVALUATION_METRIC
metric_average = total_metric / total_iterations
average_specific_metric = total_specific_metric / total_iterations
average_generic_metric = total_generic_metric / total_iterations
average_cycle_time = total_cycle_time / total_iterations
print('average %s: %f' % (metric_name, metric_average))
print(
'average specific %s: %f' % (metric_name, average_specific_metric))
print('average generic %s: %f' % (metric_name, average_generic_metric))
print('average cycle time: %f' % average_cycle_time)
print('End: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
#
results = copy.deepcopy(Constants._properties)
results[Constants.EVALUATION_METRIC] = metric_average
results['specific_' + metric_name] = average_specific_metric
results['generic_' + metric_name] = average_generic_metric
results['cycle_time'] = average_cycle_time
results['timestamp'] = time.strftime("%Y/%m/%d-%H:%M:%S")
write_results_to_csv(results)
write_results_to_json(results)
class ContextTopNRunner(object):
def __init__(self):
self.records = None
self.original_records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.predictions = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = None
self.context_topics_map = None
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
self.libfm_model_file = None
self.num_variables_in_model = None
def clear(self):
print('clear: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# self.records = None
self.train_records = None
self.test_records = None
self.records_to_predict = None
self.top_n_evaluator = None
self.headers = None
self.important_records = None
self.context_rich_topics = None
self.context_topics_map = None
if Constants.SOLVER == Constants.LIBFM:
os.remove(self.csv_train_file)
os.remove(self.csv_test_file)
os.remove(self.context_predictions_file)
os.remove(self.context_train_file)
os.remove(self.context_test_file)
os.remove(self.context_log_file)
os.remove(self.libfm_model_file)
self.csv_train_file = None
self.csv_test_file = None
self.context_predictions_file = None
self.context_train_file = None
self.context_test_file = None
self.context_log_file = None
self.libfm_model_file = None
gc.collect()
def create_tmp_file_names(self, cycle_index, fold_index):
unique_id = uuid.uuid4().hex
prefix = Constants.GENERATED_FOLDER + unique_id + '_' + \
Constants.ITEM_TYPE
print('unique id: %s' % unique_id)
self.csv_train_file = prefix + '_train.csv'
self.csv_test_file = prefix + '_test.csv'
self.context_predictions_file = prefix + '_predictions.txt'
self.context_train_file = self.csv_train_file + '.libfm'
self.context_test_file = self.csv_test_file + '.libfm'
self.context_log_file = prefix + '.log'
self.libfm_model_file = prefix + '_trained_model.libfm'
# self.csv_train_file = Constants.generate_file_name(
# 'libfm_train', 'csv', Constants.GENERATED_FOLDER, cycle_index, fold_index, Constants.USE_CONTEXT)
# self.csv_test_file = Constants.generate_file_name(
# 'libfm_test', 'csv', Constants.GENERATED_FOLDER, cycle_index, fold_index, Constants.USE_CONTEXT)
# self.context_predictions_file = Constants.generate_file_name(
# 'libfm_predictions', 'txt', Constants.GENERATED_FOLDER, cycle_index, fold_index, Constants.USE_CONTEXT)
# self.context_train_file = self.csv_train_file + '.libfm'
# self.context_test_file = self.csv_test_file + '.libfm'
# self.context_log_file = Constants.generate_file_name(
# 'libfm_log', 'log', Constants.GENERATED_FOLDER, cycle_index, fold_index, Constants.USE_CONTEXT)
# self.libfm_model_file = Constants.generate_file_name(
# 'libfm_model', 'csv', Constants.GENERATED_FOLDER, cycle_index, fold_index, Constants.USE_CONTEXT)
def load(self):
print('load: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
self.original_records =\
ETLUtils.load_json_file(
Constants.RECSYS_CONTEXTUAL_PROCESSED_RECORDS_FILE)
else:
self.original_records =\
ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
print('num_records: %d' % len(self.original_records))
user_ids = extractor.get_groupby_list(self.original_records,
Constants.USER_ID_FIELD)
item_ids = extractor.get_groupby_list(self.original_records,
Constants.ITEM_ID_FIELD)
print('total users', len(user_ids))
print('total items', len(item_ids))
def shuffle(self, records):
print('shuffle: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
random.shuffle(records)
def get_records_to_predict_topn(self):
print('get_records_to_predict_topn: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.top_n_evaluator = TopNEvaluator(self.records, self.test_records,
Constants.ITEM_TYPE, 10,
Constants.TOPN_NUM_ITEMS)
self.top_n_evaluator.initialize()
self.important_records = self.top_n_evaluator.important_records
if Constants.TEST_CONTEXT_REVIEWS_ONLY:
self.important_records = ETLUtils.filter_records(
self.important_records, Constants.HAS_CONTEXT_FIELD, [True])
self.records_to_predict =\
self.top_n_evaluator.get_records_to_predict()
if Constants.MAX_SAMPLE_TEST_SET is not None:
print('important_records %d' % len(self.important_records))
if len(self.important_records) > Constants.MAX_SAMPLE_TEST_SET:
self.important_records = random.sample(
self.important_records, Constants.MAX_SAMPLE_TEST_SET)
else:
message = 'WARNING max_sample_test_set is greater than the ' \
'number of important records'
print(message)
self.top_n_evaluator.important_records = self.important_records
self.records_to_predict = self.top_n_evaluator.get_records_to_predict()
self.test_records = None
gc.collect()
def get_records_to_predict_rmse(self):
print('get_records_to_predict_rmse: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.important_records = self.test_records
if Constants.TEST_CONTEXT_REVIEWS_ONLY:
self.important_records = ETLUtils.filter_records(
self.important_records, Constants.HAS_CONTEXT_FIELD, [True])
self.records_to_predict = self.important_records
self.test_records = None
gc.collect()
def get_records_to_predict(self, use_random_seeds):
if use_random_seeds:
utilities.plant_seeds()
if Constants.EVALUATION_METRIC == 'topn_recall':
self.get_records_to_predict_topn()
elif Constants.EVALUATION_METRIC in ['rmse', 'mae']:
self.get_records_to_predict_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def train_topic_model(self, cycle_index, fold_index):
context_extractor = topic_model_creator.create_topic_model(
self.train_records, cycle_index, fold_index)
self.context_rich_topics = context_extractor.context_rich_topics
topics_file_path = Constants.generate_file_name(
'context_topics', 'json', Constants.CACHE_FOLDER, cycle_index,
fold_index, True)
ETLUtils.save_json_file(topics_file_path,
[dict(self.context_rich_topics)])
print('Trained Context Extractor: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
return context_extractor
def load_context_reviews(self, cycle_index, fold_index):
train_records_file_path = Constants.generate_file_name(
'context_train_records', 'json', Constants.CACHE_FOLDER,
cycle_index, fold_index, True)
important_records_file_path = Constants.generate_file_name(
'context_important_records', 'json', Constants.CACHE_FOLDER,
cycle_index, fold_index, True)
self.train_records = ETLUtils.load_json_file(train_records_file_path)
self.important_records = \
ETLUtils.load_json_file(important_records_file_path)
self.load_cache_context_topics(cycle_index, fold_index)
self.context_topics_map = {}
for record in self.important_records:
self.context_topics_map[record[Constants.REVIEW_ID_FIELD]] = \
record[Constants.CONTEXT_TOPICS_FIELD]
# self.train_records = self.filter_context_words(self.train_records)
# self.print_context_topics(self.important_records)
self.important_records = None
gc.collect()
def load_cache_context_topics(self, cycle_index, fold_index):
print('load cache context topics: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
topics_file_path = Constants.generate_file_name(
'context_topics', 'json', Constants.CACHE_FOLDER, cycle_index,
fold_index, True)
self.context_rich_topics = sorted(
ETLUtils.load_json_file(topics_file_path)[0].items(),
key=operator.itemgetter(1),
reverse=True)
self.context_topics_map = {}
for record in self.important_records:
self.context_topics_map[record[Constants.REVIEW_ID_FIELD]] = \
record[Constants.CONTEXT_TOPICS_FIELD]
def find_reviews_topics(self, context_extractor, cycle_index, fold_index):
print('find topics: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
train_records_file_path = Constants.generate_file_name(
'context_train_records', 'json', Constants.CACHE_FOLDER,
cycle_index, fold_index, Constants.USE_CONTEXT)
if os.path.exists(train_records_file_path):
self.train_records = \
ETLUtils.load_json_file(train_records_file_path)
else:
context_extractor.find_contextual_topics(self.train_records)
ETLUtils.save_json_file(train_records_file_path,
self.train_records)
context_extractor.find_contextual_topics(
self.important_records, Constants.TEXT_SAMPLING_PROPORTION)
self.context_topics_map = {}
for record in self.important_records:
self.context_topics_map[record[Constants.REVIEW_ID_FIELD]] = \
record[Constants.CONTEXT_TOPICS_FIELD]
self.important_records = None
gc.collect()
@staticmethod
def print_context_topics(records):
dictionary = corpora.Dictionary.load(Constants.DICTIONARY_FILE)
all_context_topics = set()
for record in records:
words = []
corpus = record[Constants.CORPUS_FIELD]
for element in corpus:
word_id = element[0]
word = dictionary[word_id]
words.append(word + ' (' + str(word_id) + ')')
context_topics = record[Constants.CONTEXT_TOPICS_FIELD]
used_context_topics =\
dict((k, v) for k, v in context_topics.items() if v > 0.0)
all_context_topics |= set(used_context_topics.keys())
print('words: %s' % ', '.join(words))
print('text: %s' % record[Constants.TEXT_FIELD])
print('bow', record[Constants.BOW_FIELD])
# print('pos tags', record[Constants.POS_TAGS_FIELD])
print(record[Constants.TOPICS_FIELD])
# # print(record[Constants.CONTEXT_TOPICS_FIELD])
print(used_context_topics)
# print('')
# print('important records: %d' % len(records))
# print('context records: %d' % len(context_records))
# print('no context records: %d' % len(no_context_records))
# print('all used context words', all_context_words)
print('all used context topics', all_context_topics)
# print('all used context words count: %d' % len(all_context_words))
print('all used context topics: %d' % len(all_context_topics))
def prepare_records_for_libfm(self):
print('prepare_records_for_libfm: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
self.headers = build_headers(self.context_rich_topics)
if Constants.FM_REVIEW_TYPE == Constants.SPECIFIC or \
Constants.FM_REVIEW_TYPE == Constants.GENERIC:
self.train_records = ETLUtils.filter_records(
self.train_records, Constants.PREDICTED_CLASS_FIELD,
[Constants.FM_REVIEW_TYPE])
with open(self.csv_train_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.train_records:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
context_topics = record[Constants.CONTEXT_TOPICS_FIELD]
for topic in self.context_rich_topics:
# print('context_topics', context_topics)
row.append(context_topics['topic' + str(topic[0])])
if Constants.USE_NO_CONTEXT_TOPICS_SUM:
row.append(context_topics['nocontexttopics'])
writer.writerow(row)
self.train_records = None
gc.collect()
with open(self.csv_test_file, 'w') as out_file:
writer = csv.writer(out_file)
# Write header
writer.writerow(self.headers)
for record in self.records_to_predict:
row = []
for header in basic_headers:
row.append(record[header])
if Constants.USE_CONTEXT is True:
important_record = record[Constants.REVIEW_ID_FIELD]
context_topics = \
self.context_topics_map[important_record]
for topic in self.context_rich_topics:
row.append(context_topics['topic' + str(topic[0])])
if Constants.USE_NO_CONTEXT_TOPICS_SUM:
row.append(context_topics['nocontexttopics'])
writer.writerow(row)
# self.records_to_predict = None
self.context_topics_map = None
self.context_rich_topics = None
gc.collect()
print('Exported CSV and JSON files: %s' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
csv_files = [self.csv_train_file, self.csv_test_file]
print('num_cols', len(self.headers))
self.num_variables_in_model = libfm_converter.csv_to_libfm(
csv_files, 0, [1, 2], [], ',', has_header=True, suffix='.libfm')
print('Exported LibFM files: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
# def predict_fastfm(self):
#
# if Constants.USE_CONTEXT:
# for record in self.records_to_predict:
# important_record = record[Constants.REVIEW_ID_FIELD]
# record[Constants.CONTEXT_TOPICS_FIELD] = \
# self.context_topics_map[important_record]
#
# all_records = self.train_records + self.records_to_predict
# x_matrix, y_vector = fastfm_recommender.records_to_matrix(
# all_records, self.context_rich_topics)
#
# encoder = OneHotEncoder(categorical_features=[0, 1], sparse=True)
# encoder.fit(x_matrix)
#
# x_train = encoder.transform(x_matrix[:len(self.train_records)])
# y_train = y_vector[:len(self.train_records)]
# x_test = encoder.transform(x_matrix[len(self.train_records):])
#
# if Constants.FASTFM_METHOD == 'mcmc':
# # solver = mcmc.FMRegression(n_iter=num_iters, rank=num_factors)
# solver = mcmc.FMRegression(rank=Constants.FM_NUM_FACTORS)
# self.predictions = solver.fit_predict(x_train, y_train, x_test)
# elif Constants.FASTFM_METHOD == 'als':
# solver = als.FMRegression(rank=Constants.FM_NUM_FACTORS)
# solver.fit(x_train, y_train)
# self.predictions = solver.predict(x_test)
# elif Constants.FASTFM_METHOD == 'sgd':
# solver = sgd.FMRegression(rank=Constants.FM_NUM_FACTORS)
# solver.fit(x_train, y_train)
# self.predictions = solver.predict(x_test)
def predict_libfm(self):
print('predict: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
run_libfm(self.context_train_file, self.context_test_file,
self.context_predictions_file, self.context_log_file,
self.libfm_model_file)
self.predictions = rmse_calculator.read_targets_from_txt(
self.context_predictions_file)
def predict(self):
if Constants.SOLVER == Constants.LIBFM:
self.prepare_records_for_libfm()
self.predict_libfm()
# elif Constants.SOLVER == Constants.FASTFM:
# self.predict_fastfm()
def evaluate_topn(self):
print('evaluate_topn: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
self.top_n_evaluator.evaluate(self.predictions)
recall = self.top_n_evaluator.recall
print('Recall: %f' % recall)
print('Specific recall: %f' % self.top_n_evaluator.specific_recall)
print('Generic recall: %f' % self.top_n_evaluator.generic_recall)
results = {
Constants.TOPN_RECALL:
self.top_n_evaluator.recall,
Constants.SPECIFIC + '_' + Constants.TOPN_RECALL:
self.top_n_evaluator.specific_recall,
Constants.GENERIC + '_' + Constants.TOPN_RECALL:
self.top_n_evaluator.generic_recall,
Constants.HAS_CONTEXT + '_' + Constants.TOPN_RECALL:
self.top_n_evaluator.has_context_recall,
Constants.HAS_NO_CONTEXT + '_' + Constants.TOPN_RECALL:
self.top_n_evaluator.has_no_context_recall,
}
return results
def evaluate_rmse(self):
print('evaluate_rmse: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
true_values = [
record[Constants.RATING_FIELD]
for record in self.records_to_predict
]
specific_true_values = []
specific_predictions = []
generic_true_values = []
generic_predictions = []
has_context_true_values = []
has_context_predictions = []
has_no_context_true_values = []
has_no_context_predictions = []
index = 0
for record, prediction in zip(self.records_to_predict,
self.predictions):
if record[Constants.PREDICTED_CLASS_FIELD] == 'specific':
specific_true_values.append(record[Constants.RATING_FIELD])
specific_predictions.append(prediction)
elif record[Constants.PREDICTED_CLASS_FIELD] == 'generic':
generic_true_values.append(record[Constants.RATING_FIELD])
generic_predictions.append(prediction)
if record[Constants.HAS_CONTEXT_FIELD]:
has_context_true_values.append(record[Constants.RATING_FIELD])
has_context_predictions.append(prediction)
else:
has_no_context_true_values.append(
record[Constants.RATING_FIELD])
has_no_context_predictions.append(prediction)
index += 1
metric = Constants.EVALUATION_METRIC
overall_result = None
specific_result = None
generic_result = None
has_context_result = None
has_no_context_result = None
if metric == 'rmse':
overall_result = \
rmse_calculator.calculate_rmse(true_values, self.predictions)
specific_result = rmse_calculator.calculate_rmse(
specific_true_values, specific_predictions)
generic_result = rmse_calculator.calculate_rmse(
generic_true_values, generic_predictions)
has_context_result = rmse_calculator.calculate_rmse(
has_context_true_values, has_context_predictions)
has_no_context_result = rmse_calculator.calculate_rmse(
has_no_context_true_values, has_no_context_predictions)
elif metric == 'mae':
overall_result = \
rmse_calculator.calculate_mae(true_values, self.predictions)
specific_result = rmse_calculator.calculate_mae(
specific_true_values, specific_predictions)
generic_result = rmse_calculator.calculate_mae(
generic_true_values, generic_predictions)
has_context_result = rmse_calculator.calculate_mae(
has_context_true_values, has_context_predictions)
has_no_context_result = rmse_calculator.calculate_mae(
has_no_context_true_values, has_no_context_predictions)
print(metric + ': %f' % overall_result)
print('Specific ' + metric + ': %f' % specific_result)
print('Generic ' + metric + ': %f' % generic_result)
print('Has context ' + metric + ': %f' % has_context_result)
print('Has no ' + metric + ': %f' % has_no_context_result)
results = {
metric: overall_result,
Constants.SPECIFIC + '_' + metric: specific_result,
Constants.GENERIC + '_' + metric: generic_result,
Constants.HAS_CONTEXT + '_' + metric: has_context_result,
Constants.HAS_NO_CONTEXT + '_' + metric: has_no_context_result,
}
return results
def evaluate(self):
if Constants.EVALUATION_METRIC == 'topn_recall':
return self.evaluate_topn()
elif Constants.EVALUATION_METRIC in ['rmse', 'mae']:
return self.evaluate_rmse()
else:
raise ValueError('Unrecognized evaluation metric')
def perform_cross_validation(self, records):
Constants.print_properties()
# self.plant_seeds()
metrics_list = []
total_cycle_time = 0.0
num_cycles = Constants.NUM_CYCLES
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
total_iterations = num_cycles * num_folds
split = 1 - (1 / float(num_folds))
metric_name = Constants.EVALUATION_METRIC
# self.load()
for i in range(num_cycles):
print('\n\nCycle: %d/%d' % ((i + 1), num_cycles))
self.records = copy.deepcopy(records)
if Constants.SHUFFLE_DATA:
self.shuffle(self.records)
for j in range(num_folds):
fold_start = time.time()
cv_start = float(j) / num_folds
print('\nFold: %d/%d' % ((j + 1), num_folds))
self.create_tmp_file_names(i, j)
self.train_records, self.test_records = \
ETLUtils.split_train_test_copy(
self.records, split=split, start=cv_start)
# subsample_size = int(len(self.train_records)*0.5)
# self.train_records = self.train_records[:subsample_size]
self.get_records_to_predict(True)
if Constants.USE_CONTEXT:
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
self.load_cache_context_topics(None, None)
else:
context_extractor = self.train_topic_model(i, j)
self.find_reviews_topics(context_extractor, i, j)
else:
self.context_rich_topics = []
self.predict()
metrics = self.evaluate()
metrics_list.append(metrics)
print('Accumulated %s: %f' %
(metric_name,
numpy.mean([k[metric_name] for k in metrics_list])))
fold_end = time.time()
fold_time = fold_end - fold_start
total_cycle_time += fold_time
self.clear()
print("Total fold %d time = %f seconds" % ((j + 1), fold_time))
results = self.summarize_results(metrics_list)
average_cycle_time = total_cycle_time / total_iterations
results['cycle_time'] = average_cycle_time
print('average cycle time: %f' % average_cycle_time)
write_results_to_csv(results)
write_results_to_json(results)
return results
@staticmethod
def summarize_results(metrics_list):
metric_name = Constants.EVALUATION_METRIC
specific_metric_name = Constants.SPECIFIC + '_' + metric_name
generic_metric_name = Constants.GENERIC + '_' + metric_name
has_context_metric_name = Constants.HAS_CONTEXT + '_' + metric_name
has_no_context_metric_name = Constants.HAS_NO_CONTEXT + '_' + metric_name
metric_average = \
numpy.mean(numpy.mean([k[metric_name] for k in metrics_list]))
metric_stdev = numpy.std([k[metric_name] for k in metrics_list])
average_specific_metric = numpy.mean(
[k[specific_metric_name] for k in metrics_list])
average_generic_metric = numpy.mean(
[k[generic_metric_name] for k in metrics_list])
average_has_context_metric = numpy.mean(
[k[has_context_metric_name] for k in metrics_list])
average_has_no_context_metric = numpy.mean(
[k[has_no_context_metric_name] for k in metrics_list])
print('average %s:\t\t\t%f' % (metric_name, metric_average))
print('average specific %s:\t%f' %
(metric_name, average_specific_metric))
print('average generic %s:\t%f' %
(metric_name, average_generic_metric))
print('average has context %s:\t%f' %
(metric_name, average_has_context_metric))
print('average has no context %s:\t%f' %
(metric_name, average_has_no_context_metric))
print('standard deviation %s:\t%f (%f%%)' %
(metric_name, metric_stdev,
(metric_stdev / metric_average * 100)))
print('End: %s' % time.strftime("%Y/%m/%d-%H:%M:%S"))
#
results = Constants.get_properties_copy()
results[metric_name] = metric_average
results[specific_metric_name] = average_specific_metric
results[generic_metric_name] = average_generic_metric
results[has_context_metric_name] = average_has_context_metric
results[has_no_context_metric_name] = average_has_no_context_metric
results[metric_name + '_stdev'] = metric_stdev
results['timestamp'] = time.strftime("%Y/%m/%d-%H:%M:%S")
write_results_to_csv(results)
write_results_to_json(results)
return results
def run_single_fold(self, parameters):
fold = parameters['fold']
Constants.update_properties(parameters)
Constants.print_properties()
utilities.plant_seeds()
self.load()
records = self.original_records
# self.plant_seeds()
total_cycle_time = 0.0
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
split = 1 - (1 / float(num_folds))
self.records = copy.deepcopy(records)
if Constants.SHUFFLE_DATA:
self.shuffle(self.records)
fold_start = time.time()
cv_start = float(fold) / num_folds
print('\nFold: %d/%d' % ((fold + 1), num_folds))
self.create_tmp_file_names(0, fold)
self.train_records, self.test_records = \
ETLUtils.split_train_test_copy(
self.records, split=split, start=cv_start)
# subsample_size = int(len(self.train_records)*0.5)
# self.train_records = self.train_records[:subsample_size]
self.get_records_to_predict(True)
if Constants.USE_CONTEXT:
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
self.load_cache_context_topics(None, None)
else:
context_extractor = self.train_topic_model(0, fold)
self.find_reviews_topics(context_extractor, 0, fold)
else:
self.context_rich_topics = []
self.predict()
metrics = self.evaluate()
fold_end = time.time()
fold_time = fold_end - fold_start
total_cycle_time += fold_time
self.clear()
print("Total fold %d time = %f seconds" % ((fold + 1), fold_time))
return metrics
def run(self):
utilities.plant_seeds()
self.load()
records = self.original_records
if Constants.CROSS_VALIDATION_STRATEGY == 'nested_validate':
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
cycle = Constants.NESTED_CROSS_VALIDATION_CYCLE
split = 1 - (1 / float(num_folds))
cv_start = float(cycle) / num_folds
print('cv_start', cv_start)
records, _ = ETLUtils.split_train_test(self.original_records,
split, cv_start)
return self.perform_cross_validation(records)
elif Constants.CROSS_VALIDATION_STRATEGY == 'nested_test':
return self.perform_cross_validation(records)
else:
raise ValueError('Unknown cross-validation strategy')