def calculate_topic_stability(records):
Constants.update_properties({
Constants.NUMPY_RANDOM_SEED_FIELD: Constants.NUMPY_RANDOM_SEED + 10,
Constants.RANDOM_SEED_FIELD: Constants.RANDOM_SEED + 10
})
utilities.plant_seeds()
Constants.print_properties()
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
num_records = len(records)
records = records[:num_records / 2]
print('num_reviews', len(records))
all_term_rankings = []
context_extractor =\
topic_model_creator.create_topic_model(records, None, None)
terms_matrix = get_topic_model_terms(
context_extractor, Constants.TOPIC_MODEL_STABILITY_NUM_TERMS)
all_term_rankings.append(terms_matrix)
sample_ratio = 0.8
print('Total iterations: %d' % Constants.TOPIC_MODEL_STABILITY_ITERATIONS)
for _ in range(Constants.TOPIC_MODEL_STABILITY_ITERATIONS - 1):
sampled_records = sample_list(records, sample_ratio)
context_extractor = \
topic_model_creator.train_context_extractor(sampled_records)
terms_matrix = get_topic_model_terms(
context_extractor, Constants.TOPIC_MODEL_STABILITY_NUM_TERMS)
all_term_rankings.append(terms_matrix)
return calculate_stability(all_term_rankings)
def full_cycle(self):
Constants.print_properties()
print('%s: full cycle' % time.strftime("%Y/%m/%d-%H:%M:%S"))
utilities.plant_seeds()
if self.use_cache and \
os.path.exists(Constants.PROCESSED_RECORDS_FILE):
print('Records have already been processed')
self.records = \
ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
else:
self.load_records()
if 'yelp' in Constants.ITEM_TYPE:
self.transform_yelp_records()
elif 'fourcity' in Constants.ITEM_TYPE:
self.transform_fourcity_records()
self.add_integer_ids()
self.clean_reviews()
self.remove_duplicate_reviews()
self.tag_reviews_language()
self.remove_foreign_reviews()
self.lemmatize_records()
self.remove_users_with_low_reviews()
self.remove_items_with_low_reviews()
self.count_frequencies()
self.shuffle_records()
print('total_records: %d' % len(self.records))
self.classify_reviews()
self.build_bag_of_words()
self.tag_contextual_reviews()
# self.load_full_records()
self.build_dictionary()
self.build_corpus()
self.label_review_targets()
self.export_records()
self.count_specific_generic_ratio()
# self.export_to_triplet()
rda = ReviewsDatasetAnalyzer(self.records)
print('density: %f' % rda.calculate_density_approx())
print('sparsity: %f' % rda.calculate_sparsity_approx())
print('total_records: %d' % len(self.records))
user_ids = \
extractor.get_groupby_list(self.records, Constants.USER_ID_FIELD)
item_ids = \
extractor.get_groupby_list(self.records, Constants.ITEM_ID_FIELD)
print('total users', len(user_ids))
print('total items', len(item_ids))
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
self.separate_recsys_topic_model_records()
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 full_cycle(self):
Constants.print_properties()
print('%s: full cycle' % time.strftime("%Y/%m/%d-%H:%M:%S"))
utilities.plant_seeds()
if self.use_cache and \
os.path.exists(Constants.PROCESSED_RECORDS_FILE):
print('Records have already been processed')
self.records = \
ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
else:
self.preprocess()
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
self.separate_recsys_topic_model_records()
def create_topic_model(num_topics):
print('%s: evaluating topic model' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
Constants.update_properties({
Constants.NUMPY_RANDOM_SEED_FIELD: Constants.NUMPY_RANDOM_SEED + 10,
Constants.RANDOM_SEED_FIELD: Constants.RANDOM_SEED + 10,
Constants.TOPIC_MODEL_NUM_TOPICS_FIELD: num_topics
})
utilities.plant_seeds()
Constants.print_properties()
file_path = Constants.ENSEMBLED_RESULTS_FOLDER + \
"factors_final_k%02d.pkl" % Constants.TOPIC_MODEL_NUM_TOPICS
if os.path.exists(file_path):
print('Ensemble topic model already exists')
return
# topic_ensemble_caller.run_local_parse_directory()
topic_ensemble_caller.run_generate_kfold()
topic_ensemble_caller.run_combine_nmf()
def create_single_topic_model(cycle_index, fold_index, check_exists=True):
Constants.print_properties()
print('%s: Start' % time.strftime("%Y/%m/%d-%H:%M:%S"))
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
msg = 'This function shouldn\'t be used when the ' \
'separate_topic_model_recsys_reviews property is set to True'
raise ValueError(msg)
records = ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
if Constants.CROSS_VALIDATION_STRATEGY == 'nested_test':
pass
elif 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(records, split, cv_start)
else:
raise ValueError('Unknown cross-validation strategy')
utilities.plant_seeds()
num_folds = Constants.CROSS_VALIDATION_NUM_FOLDS
split = 1 - (1/float(num_folds))
for i in range(cycle_index+1):
if Constants.SHUFFLE_DATA:
random.shuffle(records)
cv_start = float(fold_index) / num_folds
train_records, test_records = \
ETLUtils.split_train_test(records, split=split, start=cv_start)
return create_topic_model(
train_records, cycle_index, fold_index, check_exists)
def evaluate_topic_model(metric):
print('%s: evaluating topic model' % time.strftime("%Y/%m/%d-%H:%M:%S"))
Constants.update_properties({
Constants.NUMPY_RANDOM_SEED_FIELD:
Constants.NUMPY_RANDOM_SEED + 10,
Constants.RANDOM_SEED_FIELD:
Constants.RANDOM_SEED + 10
})
utilities.plant_seeds()
Constants.print_properties()
records = ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
num_records = len(records)
records = records[:num_records / 2]
print('num_reviews', len(records))
all_term_rankings = None
topic_model_type = Constants.TOPIC_MODEL_TYPE
if topic_model_type in ['lda', 'nmf']:
all_term_rankings = create_all_term_rankings(records, metric)
elif topic_model_type == 'ensemble':
all_term_rankings = create_all_term_rankings_from_ensemble()
else:
raise ValueError('Unrecognized topic modeling algorithm: \'%s\'' %
topic_model_type)
print('Total iterations: %d' % len(all_term_rankings))
if metric == TERM_STABILITY_REFERENCE:
return eval_term_stability_reference(all_term_rankings)
if metric == TERM_STABILITY_PAIRWISE:
return eval_term_stability_pairwise(all_term_rankings)
elif metric == TERM_DIFFERENCE:
return eval_term_difference(all_term_rankings)
else:
raise ValueError('Unknown evaluation metric: \'%s\'' % metric)
def evaluate_topic_model(metric):
print('%s: evaluating topic model' %
time.strftime("%Y/%m/%d-%H:%M:%S"))
Constants.update_properties({
Constants.NUMPY_RANDOM_SEED_FIELD: Constants.NUMPY_RANDOM_SEED + 10,
Constants.RANDOM_SEED_FIELD: Constants.RANDOM_SEED + 10
})
utilities.plant_seeds()
Constants.print_properties()
records = ETLUtils.load_json_file(Constants.PROCESSED_RECORDS_FILE)
if Constants.SEPARATE_TOPIC_MODEL_RECSYS_REVIEWS:
num_records = len(records)
records = records[:num_records / 2]
print('num_reviews', len(records))
all_term_rankings = None
topic_model_type = Constants.TOPIC_MODEL_TYPE
if topic_model_type in ['lda', 'nmf']:
all_term_rankings = create_all_term_rankings(records, metric)
elif topic_model_type == 'ensemble':
all_term_rankings = create_all_term_rankings_from_ensemble()
else:
raise ValueError(
'Unrecognized topic modeling algorithm: \'%s\'' % topic_model_type)
print('Total iterations: %d' % len(all_term_rankings))
if metric == TERM_STABILITY_REFERENCE:
return eval_term_stability_reference(all_term_rankings)
if metric == TERM_STABILITY_PAIRWISE:
return eval_term_stability_pairwise(all_term_rankings)
elif metric == TERM_DIFFERENCE:
return eval_term_difference(all_term_rankings)
else:
raise ValueError('Unknown evaluation metric: \'%s\'' % metric)
def create_topic_model(num_topics):
print('%s: evaluating topic model' % time.strftime("%Y/%m/%d-%H:%M:%S"))
Constants.update_properties({
Constants.NUMPY_RANDOM_SEED_FIELD:
Constants.NUMPY_RANDOM_SEED + 10,
Constants.RANDOM_SEED_FIELD:
Constants.RANDOM_SEED + 10,
Constants.TOPIC_MODEL_NUM_TOPICS_FIELD:
num_topics
})
utilities.plant_seeds()
Constants.print_properties()
file_path = Constants.ENSEMBLED_RESULTS_FOLDER + \
"factors_final_k%02d.pkl" % Constants.TOPIC_MODEL_NUM_TOPICS
if os.path.exists(file_path):
print('Ensemble topic model already exists')
return
# topic_ensemble_caller.run_local_parse_directory()
topic_ensemble_caller.run_generate_kfold()
topic_ensemble_caller.run_combine_nmf()
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)
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
