def get_train_test(search_sessions_path, search_sessions_num):
"""
Reads the given number of search sessions from the given path
and splits them into train and test sets with the ratio 3/1.
Queries that are not present in the train set are removed from the test set.
Returns the train/test sessions and queries.
:param search_sessions_path: The path to the file with search sessions
in the format of Yandex Relevance Prediction Challenge
(http://imat-relpred.yandex.ru/en).
:param search_sessions_num: The number of search sessions to consider.
:returns: Returns train sessions, train queries (distinct), test sessions, test queries (distinct).
"""
search_sessions = YandexRelPredChallengeParser().parse(
search_sessions_path, search_sessions_num)
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = Utils.get_unique_queries(train_sessions)
test_sessions = Utils.filter_sessions(search_sessions[train_test_split:],
train_queries)
test_queries = Utils.get_unique_queries(test_sessions)
return train_sessions, train_queries, test_sessions, test_queries
def fit_click_model(search_sessions, cm):
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = pc_utils.get_unique_queries(train_sessions)
test_sessions = pc_utils.filter_sessions(
search_sessions[train_test_split:], train_queries)
test_queries = pc_utils.get_unique_queries(test_sessions)
module = importlib.import_module(modules[cm])
click_model = getattr(module, cm)
click_model.train(train_sessions)
with open(
Path(config.DATASET_OUTPUT_FOLDER + 'click_models/' + cm +
'.pkl')) as out:
pickle.dump(click_model, out)
ranker = QueryDocumentRanker(PyClickModelAdapter(model))
df = expand_content_ids(ranker.rank(query).to_frame()).sort_values(0)
for idx, row in df.iterrows():
a = sdbn_click_model.params[sdbn_click_model.param_names.attr].get(query, idx)
s = sdbn_click_model.params[sdbn_click_model.param_names.sat].get(query, idx)
n = sdbn_click_model.params[sdbn_click_model.param_names.attr]._container[query][idx]._denominator
print(f'a={a} s={s}, n={n}: {idx} ({row["title"]})')
if __name__ == "__main__":
logging.basicConfig(filename='estimate_with_pyclick.log',level=logging.INFO)
training, test = load_from_csv()
train_sessions = map_to_pyclick_format(training)
test_sessions = map_to_pyclick_format(test)
train_queries = Utils.get_unique_queries(train_sessions)
# PyClick normally filters out any test sessions that aren't in the training set.
# I shouldn't need to do this, because my train/test split shouldn't let this happen.
assert len(test_sessions) == len(Utils.filter_sessions(test_sessions, train_queries))
test_queries = Utils.get_unique_queries(test_sessions)
print('SDBN')
train_model(sdbn_click_model, train_sessions, train_queries)
evaluate_fit(sdbn_click_model, test_sessions, test_queries)
with open('sdbn_model.json', 'w') as f:
f.write(sdbn_click_model.to_json())
from evaluate_model import PyClickModelAdapter, QueryDocumentRanker
result = SearchResult(search.final_click_url, 1)
session.web_results.append(result)
sessions.append(session)
return sessions
if __name__ == "__main__":
click_model = SDBN()
search_sessions = get_sessions()
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = Utils.get_unique_queries(train_sessions)
test_sessions = Utils.filter_sessions(search_sessions[train_test_split:],
train_queries)
test_queries = Utils.get_unique_queries(test_sessions)
print("===============================")
print("Training on %d search sessions (%d unique queries)." %
(len(train_sessions), len(train_queries)))
print("===============================")
start = time.time()
click_model.train(train_sessions)
end = time.time()
print("\tTrained %s click model in %i secs:\n%r" %
(click_model.__class__.__name__, end - start, click_model))
print "\tdataset - the path to the dataset from Yandex Relevance Prediction Challenge"
print "\tsessions_max - the maximum number of one-query search sessions to consider"
print ""
sys.exit(1)
click_model = TCM()
search_sessions_path = sys.argv[1]
search_sessions_num = int(sys.argv[2])
# search_sessions = YandexRelPredChallengeParser().parse(search_sessions_path, search_sessions_num)
search_sessions = YandexPersonalizedChallengeParser().parse(
search_sessions_path, search_sessions_num)
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = Utils.get_unique_queries(train_sessions)
train_tasks = SearchTask.get_search_tasks(train_sessions)
# test_sessions = Utils.filter_sessions(search_sessions[train_test_split:], train_queries)
# test_queries = Utils.get_unique_queries(test_sessions)
# test_tasks = SearchTask.get_search_tasks(test_sessions)
print "-------------------------------"
print "Training on %d search tasks (%d search sessions, %d unique queries)." % \
(len(train_tasks), len(train_sessions), len(train_queries))
print "-------------------------------"
# print "data ", train_tasks[:5]
click_model.train(train_tasks)
print "\tTrained %s click model:\n%r" % (click_model.__class__.__name__,
click_model)
with open('./data/search_sessions.pkl', 'wb') as fout:
pickle.dump(search_sessions, fout, protocol=-1)
else:
with open('./data/search_sessions.pkl') as fin:
search_sessions = pickle.load(fin)
# Because we are interested mostly in getting parameter estimates
# for `search_sessions` queries, they are also used as 'hold-out set'.
holdout_search_sessions = list(search_sessions)
n_search_sessions = 1000000
# Reads `n_search_session` number of sessions from the Yandex dataset ...
additional_search_sessions = Utils.filter_sessions(YandexRelPredChallengeParser().parse(search_sessions_path, sessions_max=n_search_sessions),
Utils.get_unique_queries(search_sessions),
operation='remove')
# ... and merges them with `search_sessions` while taking care of not
# including any session twice.
search_sessions.extend(additional_search_sessions)
queries = Utils.get_unique_queries(search_sessions)
print "---------------------------------------------------------"
print "Training on %d search sessions (%d unique queries)." % (len(search_sessions), len(queries))
print "---------------------------------------------------------"
for model_name, model_filename in models_to_train:
# Makes sure the output file for model parameters
# does not exist (overwrite protection).
import numpy as np
from operator import add
import pickle
import scipy.stats as st
import matplotlib.pyplot as mp
click_model = UBM()
search_sessions_path = "YandexRelPredChallenge.txt"
search_sessions_num = 10000
search_sessions = YandexRelPredChallengeParser().parse(search_sessions_path, search_sessions_num)
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = Utils.get_unique_queries(train_sessions)
test_sessions = Utils.filter_sessions(search_sessions[train_test_split:], train_queries)
test_queries = Utils.get_unique_queries(test_sessions)
click_model.train(train_sessions)
measures = {'dcg':cm.dcg, 'rbp':cm.rbp, 'err':cm.err}
ep_pairs = cm.pairs()
for m in measures:
print "making groups"
groups = cm.make_groups(measures[m], ep_pairs)
print 'ubm with %s' % m
if len(sys.argv) < 3:
print "USAGE: %s <dataset> <sessions_max>" % sys.argv[0]
print "\tdataset - the path to the dataset from Yandex Relevance Prediction Challenge"
print "\tsessions_max - the maximum number of one-query search sessions to consider"
print ""
sys.exit(1)
click_model = TCM()
search_sessions_path = sys.argv[1]
search_sessions_num = int(sys.argv[2])
search_sessions = YandexRelPredChallengeParser().parse(search_sessions_path, search_sessions_num)
train_test_split = int(len(search_sessions) * 0.75)
train_sessions = search_sessions[:train_test_split]
train_queries = Utils.get_unique_queries(train_sessions)
train_tasks = SearchTask.get_search_tasks(train_sessions)
# test_sessions = Utils.filter_sessions(search_sessions[train_test_split:], train_queries)
# test_queries = Utils.get_unique_queries(test_sessions)
# test_tasks = SearchTask.get_search_tasks(test_sessions)
print "-------------------------------"
print "Training on %d search tasks (%d search sessions, %d unique queries)." % \
(len(train_tasks), len(train_sessions), len(train_queries))
print "-------------------------------"
click_model.train(train_tasks)
print "\tTrained %s click model:\n%r" % (click_model.__class__.__name__, click_model)
# print "-------------------------------"