def reciprocal_rank_at_recall(run, label_store):
scores_by_topic = dict()
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = set(get_all_subtopics(label_store, topic_id))
seen_subtopics = set()
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
# check off seen subtopics
for subtopic, conf in get_best_subtopics(result['subtopics']):
seen_subtopics.add(subtopic)
if len(seen_subtopics) == len(subtopic_ids):
break
scores_by_topic[topic_id] = 1 / (idx + 1)
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
run['scores']['reciprocal_rank_at_recall'] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}
def precision_at_recall(run, label_store):
scores_by_topic = dict()
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = set(get_all_subtopics(label_store, topic_id))
seen_subtopics = set()
relevant_docs = 0
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
result_subtopics = \
{subtopic for subtopic, conf in get_best_subtopics(result['subtopics'])}
if result['on_topic']:
relevant_docs += 1
seen_subtopics.update(result_subtopics)
if len(seen_subtopics) == len(subtopic_ids):
break
## precision is number of documents relevant at stopping point
p = relevant_docs / (idx + 1)
scores_by_topic[topic_id] = p
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
run['scores']['precision_at_recall'] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}
def reciprocal_rank_at_recall(run, label_store):
scores_by_topic = dict()
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = set(get_all_subtopics(label_store, topic_id))
seen_subtopics = set()
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
# check off seen subtopics
for subtopic, conf in get_best_subtopics(result['subtopics']):
seen_subtopics.add(subtopic)
if len(seen_subtopics) == len(subtopic_ids):
break
scores_by_topic[topic_id] = 1/(idx + 1)
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
run['scores']['reciprocal_rank_at_recall'] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}
def precision_at_recall(run, label_store):
scores_by_topic = dict()
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = set(get_all_subtopics(label_store, topic_id))
seen_subtopics = set()
relevant_docs = 0
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
result_subtopics = \
{subtopic for subtopic, conf in get_best_subtopics(result['subtopics'])}
if result['on_topic']:
relevant_docs += 1
seen_subtopics.update(result_subtopics)
if len(seen_subtopics) == len(subtopic_ids):
break
## precision is number of documents relevant at stopping point
p = relevant_docs/(idx + 1)
scores_by_topic[topic_id] = p
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
run['scores']['precision_at_recall'] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}
def average_err(run,
label_store,
mean_type='arithmetic',
relevance_metric='graded'):
'''
mean_type can be `arithmetic' or `harmonic'
'''
scores_by_topic = dict()
relevance_func = relevance_metrics[relevance_metric]
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = list(set(get_all_subtopics(label_store, topic_id)))
# for each subtopic, compute a running stopping_p
# and score
p_continue = defaultdict(lambda: 1)
score = defaultdict(float)
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
for subtopic, conf in get_best_subtopics(result['subtopics']):
rel = relevance_func(conf)
p_stop_here = p_continue[subtopic] * rel
score[subtopic] += p_stop_here / (idx + 1)
## update stopping probabilities
p_continue[subtopic] *= (1 - rel)
## precision is number of documents relevant at stopping point
if mean_type == 'arithmetic':
scores_by_topic[topic_id] = mean(score.values())
elif mean_type == 'harmonic':
scores_by_topic[topic_id] = harmonic_mean(score.values())
else:
sys.exit('Error: invalid mean type specified.')
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
scorer_name = 'average_err_%s' % mean_type
run['scores'][scorer_name] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}
def average_err(run, label_store, mean_type='arithmetic', relevance_metric='graded'):
'''
mean_type can be `arithmetic' or `harmonic'
'''
scores_by_topic = dict()
relevance_func = relevance_metrics[relevance_metric]
## score for each topic
for topic_id, results in run['results'].items():
## get all subtopics for the topic
subtopic_ids = list(set(get_all_subtopics(label_store, topic_id)))
# for each subtopic, compute a running stopping_p
# and score
p_continue = defaultdict(lambda: 1)
score = defaultdict(float)
for idx, result in enumerate(results):
assert idx == result['rank'] - 1
for subtopic, conf in get_best_subtopics(result['subtopics']):
rel = relevance_func(conf)
p_stop_here = p_continue[subtopic]*rel
score[subtopic] += p_stop_here/(idx+1)
## update stopping probabilities
p_continue[subtopic] *= (1-rel)
## precision is number of documents relevant at stopping point
if mean_type == 'arithmetic':
scores_by_topic[topic_id] = mean(score.values())
elif mean_type == 'harmonic':
scores_by_topic[topic_id] = harmonic_mean(score.values())
else:
sys.exit('Error: invalid mean type specified.')
## macro average over all the topics
macro_avg = mean(scores_by_topic.values())
scorer_name = 'average_err_%s' % mean_type
run['scores'][scorer_name] = \
{'scores_by_topic': scores_by_topic, 'macro_average': macro_avg}