def evaluate_fit(trained_model, test_sessions, test_queries):
"""
Evaluate the model's fit to the observed data - i.e. whether C==0 or C==1 for every
item/session.
There are two measures:
- Log likelihood goes from negative infinity (bad) to 0 (good)
- It measures the likelihood of observing the clicks in all the test sessions if the model is correct
- Perplexity goes from 1 (good) to 2 (bad).
- It's a measure of how surprised we are about all clicks and non-clicks in all of the test sessions if the model is correct.
- When comparing models you can use perplexity gain (pB - pA) / (pB - 1)
- It can be computed at individual ranks, or averaged across all ranks. Perplexity is normally higher for higher ranks.
"""
print("-------------------------------")
print("Testing on %d search sessions (%d unique queries)." % (len(test_sessions), len(test_queries)))
print("-------------------------------")
loglikelihood = LogLikelihood()
perplexity = Perplexity()
start = time.time()
ll_value = loglikelihood.evaluate(trained_model, test_sessions)
end = time.time()
print("\tlog-likelihood: %f; time: %i secs" % (ll_value, end - start))
start = time.time()
perp_value = perplexity.evaluate(trained_model, test_sessions)[0]
end = time.time()
print("\tperplexity: %f; time: %i secs" % (perp_value, end - start))
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("-------------------------------")
print("Testing on %d search sessions (%d unique queries)." %
(len(test_sessions), len(test_queries)))
print("-------------------------------")
loglikelihood = LogLikelihood()
perplexity = Perplexity()
start = time.time()
ll_value = loglikelihood.evaluate(click_model, test_sessions)
end = time.time()
print("\tlog-likelihood: %f; time: %i secs" % (ll_value, end - start))
start = time.time()
perp_value = perplexity.evaluate(click_model, test_sessions)[0]
end = time.time()
print("\tperplexity: %f; time: %i secs" % (perp_value, end - start))
end = time.time()
logger.info('Finish training. Time consumed: {} seconds'.format(end -
start))
# Log likelihood
logger.info('Computing log likelihood')
loglikelihood = LogLikelihood()
start = time.time()
ll_value = loglikelihood.evaluate(click_model, test_dataset)
end = time.time()
logger.info('Log likelihood: {}. Time consumed: {} seconds'.format(
ll_value, end - start))
# Perplexity
logger.info('Computing perplexity')
perplexity = Perplexity()
start = time.time()
perp_value = perplexity.evaluate(click_model, test_dataset)[0]
end = time.time()
logger.info('Perplexity: {}. Time consumed: {} seconds'.format(
perp_value, end - start))
# NDCG
logger.info('Computing NDCG@k')
RelevanceEstimation = RankingPerformance(args)
start = time.time()
ndcg_version1 = {}
ndcg_version2 = {}
ks = [1, 3, 5, 10]
for k in ks:
ndcg_version1[k], ndcg_version2[k] = RelevanceEstimation.evaluate(
sys.exit(1)
click_model = globals()[sys.argv[1]]()
search_sessions_path = sys.argv[2]
search_sessions_num = int(sys.argv[3])
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)
print "-------------------------------"
print "Training on %d search sessions (%d unique queries)." % (len(train_sessions), len(train_queries))
print "-------------------------------"
click_model.train(train_sessions)
print "\tTrained %s click model:\n%r" % (click_model.__class__.__name__, click_model)
print "-------------------------------"
print "Testing on %d search sessions (%d unique queries)." % (len(test_sessions), len(test_queries))
print "-------------------------------"
loglikelihood = LogLikelihood()
print "\tlog-likelihood: %f" % loglikelihood.evaluate(click_model, test_sessions)
perplexity = Perplexity()
print "\tperplexity: %f" % perplexity.evaluate(click_model, test_sessions)[0]
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)
print "-------------------------------"
print "Training on %d search sessions (%d unique queries)." % (
len(train_sessions), len(train_queries))
print "-------------------------------"
click_model.train(train_sessions)
print "\tTrained %s click model:\n%r" % (click_model.__class__.__name__,
click_model)
print "-------------------------------"
print "Testing on %d search sessions (%d unique queries)." % (
len(test_sessions), len(test_queries))
print "-------------------------------"
loglikelihood = LogLikelihood()
print "\tlog-likelihood: %f" % loglikelihood.evaluate(
click_model, test_sessions)
perplexity = Perplexity()
print "\tperplexity: %f" % perplexity.evaluate(click_model,
test_sessions)[0]
sys.exit(1)
click_model = globals()[sys.argv[1]]()
search_sessions_path = sys.argv[2]
search_sessions_num = int(sys.argv[3])
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)
print "-------------------------------"
print "Training on %d search sessions (%d unique queries)." % (len(train_sessions), len(train_queries))
print "-------------------------------"
click_model.train(search_sessions)
#print "\tTrained %s click model:\n%r" % (click_model.__class__.__name__, click_model)
print "-------------------------------"
print "Testing on %d search sessions (%d unique queries)." % (len(test_sessions), len(test_queries))
print "-------------------------------"
loglikelihood = LogLikelihood()
print "\tlog-likelihood: %f" % loglikelihood.evaluate(click_model, test_queries)
perplexity = Perplexity()
print "\tperplexity: %f" % perplexity.evaluate(click_model, test_queries)[0]