def update_records(self, update_messages, human):
# if we're using naive tracing, we actually don't care which records we update
if not update_messages:
return self
grouped_update_messages = self.group_by_received_at(update_messages)
for received_at, update_messages in grouped_update_messages.items():
for update_message in update_messages:
old_message_dec = Message(update_message.uid,
update_message.risk,
update_message.contact_day,
update_message.unobs_id,
update_message.has_app)
old_message_enc = encode_message(old_message_dec)
old_cluster = None
for cluster, messages in self.clusters_by_day[
update_message.contact_day].items():
for message in messages:
if message == old_message_enc:
old_cluster = cluster
#print(old_cluster)
break
if old_cluster:
break
updated_message = Message(old_message_dec.uid,
update_message.new_risk,
old_message_dec.day,
old_message_dec.unobs_id,
old_message_dec.has_app)
new_cluster = hash_to_cluster(updated_message)
self.update_record(old_cluster, new_cluster, old_message_dec,
updated_message)
return self
def update_records(self, update_messages):
""" Updates old encounter messages with a new risk"""
if not update_messages:
return self
# TODO: implement a 24-hour cycle of message updates and batch the update messages by this.
grouped_update_messages = self.group_by_received_at(update_messages)
# for each batch of update messages
for received_at, update_messages in grouped_update_messages.items():
# for each update message in the batch
for update_message in update_messages:
old_message_dec = Message(update_message.uid, update_message.risk, update_message.day, update_message.unobs_id)
old_message_enc = encode_message(old_message_dec)
old_cluster = None
# TODO: don't use the secret info when finding the message to update. Also, optimize this loop.
# Find the matching message to update
for cluster, messages in self.clusters_by_day[update_message.day].items():
for message in messages:
if message == old_message_enc:
old_cluster = cluster
break
if old_cluster:
break
# Create the a new encounter message with the update risk and replace the old encounter message
updated_message = Message(old_message_dec.uid, update_message.new_risk, old_message_dec.day, old_message_dec.unobs_id)
new_cluster = hash_to_cluster(updated_message)
self.update_record(old_cluster, new_cluster, old_message_dec, updated_message)
return self
def add_messages(self, messages, current_day, rng=None):
""" This function clusters new messages by scoring them against old messages in a sort of naive nearest neighbors approach"""
for message in messages:
m_dec = decode_message(message)
# otherwise score against previous messages
best_cluster, best_message, best_score = self.score_matches(
m_dec, current_day, rng=rng)
if best_score >= 0:
cluster_id = best_cluster
else:
cluster_id = hash_to_cluster(m_dec)
self.all_messages.append(message)
self.clusters[cluster_id].append(message)
self.add_to_clusters_by_day(cluster_id, m_dec.day, message)
def score_matches(self, m_new, current_day, rng=None):
""" This function checks a new risk message against all previous messages, and assigns to the closest one in a brute force manner"""
best_cluster = hash_to_cluster(m_new)
best_message = None
best_score = -1
for i in range(current_day - 3, current_day + 1, -1):
for cluster_id, messages in self.clusters_by_day[i].items():
for m_enc in messages:
obs_uid, risk, day, unobs_uid, has_app = decode_message(
m_enc)
if m_new.uid == obs_uid and m_new.day == day:
best_cluster = cluster_id
best_message = m_enc
best_score = 3
break
elif compare_uids(
m_new.uid, obs_uid,
1) and m_new.day - 1 == day and m_new.risk == risk:
best_cluster = cluster_id
best_message = m_enc
best_score = 2
elif compare_uids(
m_new.uid, obs_uid,
2) and m_new.day - 2 == day and best_score < 1:
best_cluster = cluster_id
best_message = m_enc
best_score = 1
elif compare_uids(
m_new.uid, obs_uid,
3) and m_new.day - 3 == day and best_score < 0:
best_cluster = cluster_id
best_message = m_enc
best_score = 0
else:
best_cluster = cluster_id
best_message = m_enc
best_score = -1
if best_score == 3:
break
if best_score == 3:
break
# print(f"best_cluster: {best_cluster}, m_new: {m_new}, best_score: {best_score}")
# print(self.clusters)
if best_message:
best_message = decode_message(best_message)
return best_cluster, best_message, best_score
def score_matches(self, m_new, current_day):
""" This function checks a new risk message against previous clusterings, and assigns to the closest one in a brute force manner"""
cluster_days = hash_to_cluster_day(m_new)
best_cluster = hash_to_cluster(m_new)
if self.clusters_by_day[current_day].get(best_cluster, None):
return best_cluster
found = False
for day, cluster_ids in cluster_days.items():
for cluster_id in cluster_ids:
if self.clusters_by_day[current_day - day].get(cluster_id, None):
best_cluster = cluster_id
found = True
break
if found:
break
return best_cluster
def update_records(self, update_messages, human):
# if we're using naive tracing, we actually don't care which records we update
#if not config.CLUSTER_MESSAGES and config.CLUSTER_TYPE == "heuristic":
# for update_message in update_messages:
# self.clusters_by_day
if not update_messages:
return self
grouped_update_messages = self.group_by_received_at(update_messages)
for received_at, update_messages in grouped_update_messages.items():
# num days x num clusters
cluster_cards = np.zeros((max(self.clusters_by_day.keys()) + 1,
max(self.clusters.keys()) + 1))
update_cards = np.zeros((max(self.clusters_by_day.keys()) + 1, 1))
# figure out the cardinality of each day's message set
for day, clusters in self.clusters_by_day.items():
for cluster_id, messages in clusters.items():
cluster_cards[day][cluster_id] = len(messages)
for update_message in update_messages:
update_cards[update_message.day] += 1
# find the nearest cardinality cluster
perfect_signatures = np.where(
(cluster_cards == update_cards).all(axis=0))[0]
if not any(perfect_signatures):
# calculate the wasserstein distance between every signature
scores = []
for cluster_idx in range(cluster_cards.shape[1]):
scores.append(
dist(cluster_cards[:, cluster_idx],
update_cards.reshape(-1)))
best_cluster = int(np.argmin(scores))
# for each day
for day in range(len(update_cards)):
cur_cardinality = int(cluster_cards[day, best_cluster])
target_cardinality = int(update_cards[day])
# if (and while) the cardinality is not what it should be, as determined by the update_messages
while cur_cardinality - target_cardinality != 0:
# print(f"day: {day}, cur_cardinality: {cur_cardinality}, target_cardinality: {target_cardinality}")
# if we need to remove messages from this cluster on this day,
if cur_cardinality > target_cardinality:
best_score = -1
best_message = None
new_cluster_id = None
# then for each message in that day/cluster,
for message in self.clusters_by_day[day][
best_cluster]:
for cluster_id, messages in self.clusters_by_day[
day].items():
if cluster_id == best_cluster:
continue
# and for each alternative cluster on that day
for candidate_cluster_message in messages:
# check if it's a good cluster to move this message to
score = self.score_two_messages(
decode_message(
candidate_cluster_message),
message)
if (score > best_score
or not best_message):
best_message = message
new_cluster_id = cluster_id
# if there are no other clusters on that day make a new cluster
if not best_message:
best_message = message
message = decode_message(message)
new_cluster_id = hash_to_cluster(message)
best_message = decode_message(best_message)
# for the message which best fits another cluster, move it there
self.update_record(best_cluster, new_cluster_id,
best_message, best_message)
cur_cardinality -= 1
# print(f"removing from cluster {best_cluster} to cluster {new_cluster_id} on day {day}")
#otherwise we need to add messages to this cluster/day
else:
# so look for messages which closely match our update messages, and add them
for update_message in update_messages:
if update_message.day == day:
break
best_score = -2
best_message = None
old_cluster_id = None
for cluster_id, messages in self.clusters_by_day[
day].items():
for message in messages:
score = self.score_two_messages(
update_message, message)
if (score > best_score
and cluster_id != best_cluster):
best_message = message
old_cluster_id = cluster_id
best_message = decode_message(best_message)
updated_message = Message(best_message.uid,
update_message.new_risk,
best_message.day,
best_message.unobs_id)
# print(f"adding from cluster {old_cluster_id} to cluster {best_cluster} on day {day}")
self.update_record(old_cluster_id, best_cluster,
best_message, updated_message)
cur_cardinality += 1
else:
best_cluster = self.score_clusters(update_messages,
perfect_signatures)
for update_message in update_messages:
best_score = -1
best_message = self.clusters_by_day[
update_message.day][best_cluster][0]
for risk_message in self.clusters_by_day[
update_message.day][best_cluster]:
score = self.score_two_messages(update_message,
risk_message)
if score > best_score:
best_message = risk_message
best_message = decode_message(best_message)
updated_message = Message(best_message.uid,
update_message.new_risk,
best_message.day,
best_message.unobs_id)
self.update_record(best_cluster, best_cluster, best_message,
updated_message)
return self