def main(): db_helper = ApeicDBHelper() users = db_helper.get_users() accs = [] mfu_accs = [] for user in users: sessions = db_helper.get_sessions(user) if len(sessions) < 20: continue else: print colored(user, attrs=['blink']), training_sessions, testing_sessions = split(sessions, 0.84) counter = Counter(map(lambda x: x['application'], list(itertools.chain(*training_sessions)))) # counter = Counter(map(lambda x: x['application'], map(lambda x: x[0], training_sessions))) # installed_apps = train(training_sessions) installed_apps = train_1(training_sessions) # installed_apps = train_2(training_sessions) acc = test(testing_sessions, installed_apps, counter) accs.append(acc) predictor = MFUPredictor() predictor.train(list(itertools.chain(*training_sessions))) # testing_sessions = filter(lambda x: len(x) > 1, testing_sessions) mfu_acc, mrr = predictor.test(list(itertools.chain(*testing_sessions)), 4) print mfu_acc mfu_accs.append(mfu_acc) # installed_apps = train_by_ic([list(itertools.chain(*training_sessions))]) # test([list(itertools.chain(*testing_sessions))], installed_apps) # break print sum(accs)/len(accs), sum(mfu_accs)/len(mfu_accs)
def main():
db_helper = ApeicDBHelper()
users = db_helper.get_users()
accs = []
mfu_accs = []
for user in users:
sessions = db_helper.get_sessions(user)
if len(sessions) < 20:
continue
else:
print colored(user, attrs=['blink']),
training_sessions, testing_sessions = split(sessions, 0.84)
counter = Counter(
map(lambda x: x['application'],
list(itertools.chain(*training_sessions))))
# counter = Counter(map(lambda x: x['application'], map(lambda x: x[0], training_sessions)))
# installed_apps = train(training_sessions)
installed_apps = train_1(training_sessions)
# installed_apps = train_2(training_sessions)
acc = test(testing_sessions, installed_apps, counter)
accs.append(acc)
predictor = MFUPredictor()
predictor.train(list(itertools.chain(*training_sessions)))
# testing_sessions = filter(lambda x: len(x) > 1, testing_sessions)
mfu_acc, mrr = predictor.test(list(itertools.chain(*testing_sessions)),
4)
print mfu_acc
mfu_accs.append(mfu_acc)
# installed_apps = train_by_ic([list(itertools.chain(*training_sessions))])
# test([list(itertools.chain(*testing_sessions))], installed_apps)
# break
print sum(accs) / len(accs), sum(mfu_accs) / len(mfu_accs)
def main():
hits = 0.0
misses = 0.0
db_helper = ApeicDBHelper()
users = db_helper.get_users()
accuracies = []
for user in users:
if user == '11d1ef9f845ec10e':
continue
print colored(user, attrs=['blink'])
sessions = db_helper.get_sessions(user)
last = sessions[0][-1]['application']
test = [sessions[0]]
for s in sessions[1:]:
if s[0]['application'] == last:
if s[1:]:
test.append(s[1:])
else:
test.append(s)
last = s[-1]['application']
training_sessions, testing_sessions = split(test, 0.8)
training_sessions, testing_sessions = split(sessions, 0.8)
logs = aggregate_sessions(training_sessions)
extractor = FeatureExtractor()
X, y = extractor.generate_training_instances(logs)
nb = MultinomialNB()
predictor = nb.fit(X, y)
last_used_app = ''
for session in testing_sessions:
for log in session:
# if log['application'] in ['com.android.settings', \
# 'com.android.packageinstaller', 'com.htc.android.worldclock', 'com.android.systemui']:
# continue
instance = extractor.transform(last_used_app, log)
ranking = sorted(zip(predictor.classes_, predictor.predict_proba(instance)[0]), \
key=operator.itemgetter(1), reverse=True)
candidates = map(lambda x: x[0], ranking[:4])
if log['application'] in candidates:
hits += 1.0
else:
misses += 1.0
last_used_app = log['application']
acc = hits/(hits + misses)
accuracies.append(acc)
print acc, hits, misses
# break
print sum(accuracies)/len(accuracies)
def get_app_usage_distrs(self):
app_usage_distrs = {}
db = ApeicDBHelper()
for user in db.get_users():
sessions = db.get_sessions(user)
logs = list(itertools.chain(*sessions))
logs = map(lambda x: self._format(x), logs)
user_traces = sorted(logs, key=itemgetter(-1))
for app, traces in groupby(user_traces, lambda x: x[-1]):
cluster = list(traces)
count = app_usage_distrs.keys().count(app)
app_usage_distrs['%s_%d' % (app, count + 1)] = \
tuple(map(lambda i: self._get_feature_distr(i, cluster), range(len(self.feature_names))))
return app_usage_distrs
def get_app_usage_distrs(self):
app_usage_distrs = {}
db = ApeicDBHelper()
for user in db.get_users():
sessions = db.get_sessions(user)
logs = list(itertools.chain(*sessions))
logs = map(lambda x: self._format(x), logs)
user_traces = sorted(logs, key=itemgetter(-1))
for app, traces in groupby(user_traces, lambda x: x[-1]):
cluster = list(traces)
count = app_usage_distrs.keys().count(app)
app_usage_distrs['%s_%d' % (app, count + 1)] = \
tuple(map(lambda i: self._get_feature_distr(i, cluster), range(len(self.feature_names))))
return app_usage_distrs
def compute_repeatability(user): db_helper = ApeicDBHelper() sessions = db_helper.get_sessions(user) results = defaultdict(list) for session in sessions: apps = [x['application'] for x in session] counts = Counter(apps) for app in counts: results[app].append(counts[app]) for app in results: results[app] = sum(results[app])/float(len(results[app])) return sorted(results.iteritems(), key=operator.itemgetter(1), reverse=True)
def get_segment_len_distrs(self):
distr = defaultdict(list)
db = ApeicDBHelper()
for user in db.get_users():
sessions = db.get_sessions(user)
for session in sessions:
int_context = map(lambda x: x['application'], session)
session_len = len(int_context)
initiator = max(set(int_context), key=int_context.count)
indices = [i for i, x in enumerate(int_context) if x == initiator] + [session_len]
if session_len <= 2:
distr[session_len].append((session_len, ))
else:
segment_lens = ()
for i in xrange(indices[0]):
segment_lens += (1, )
for i in xrange(1, len(indices)):
segment_lens += (indices[i] - indices[i-1], )
distr[session_len].append(segment_lens)
distr = dict(map(lambda x: (x, Counter(distr[x]).items()), distr))
return distr
def get_segment_len_distrs(self):
distr = defaultdict(list)
db = ApeicDBHelper()
for user in db.get_users():
sessions = db.get_sessions(user)
for session in sessions:
int_context = map(lambda x: x['application'], session)
session_len = len(int_context)
initiator = max(set(int_context), key=int_context.count)
indices = [
i for i, x in enumerate(int_context) if x == initiator
] + [session_len]
if session_len <= 2:
distr[session_len].append((session_len, ))
else:
segment_lens = ()
for i in xrange(indices[0]):
segment_lens += (1, )
for i in xrange(1, len(indices)):
segment_lens += (indices[i] - indices[i - 1], )
distr[session_len].append(segment_lens)
distr = dict(map(lambda x: (x, Counter(distr[x]).items()), distr))
return distr
def test(k=4, ignore_initiator=True):
db_helper = ApeicDBHelper()
users = filter(lambda x: x != '11d1ef9f845ec10e', db_helper.get_users())
total_hits = 0.0
total_misses = 0.0
m = 0.0
initial_misses = 0.0
nb_total_hits = 0.0
nb_total_misses = 0.0
nb_m = 0.0
lu_total_hits = 0.0
lu_total_misses = 0.0
lu_m = 0.0
mru_total_hits = 0.0
mru_total_misses = 0.0
mru_m = 0.0
mfu_total_hits = 0.0
mfu_total_misses = 0.0
mfu_m = 0.0
for user in users:
# print colored(user, attrs=['blink'])
sessions = db_helper.get_sessions(user)
training_sessions, testing_sessions = split(sessions, 0.8)
past_launches = map(lambda x: x['application'], chain(*training_sessions))
mru_candidates = past_launches[-k:]
l1 = past_launches[-1]
l2 = past_launches[-2]
used_apps = []
for session in training_sessions:
used_apps += map(lambda x: x['application'], session)
counter = Counter(used_apps)
mfu_candidates = map(lambda x: x[0], counter.most_common(k))
predictor = ApeicPredictor()
predictor.train(training_sessions)
lu_predictor = LUPredictor(0.25)
lu_predictor.train(list(chain(*training_sessions)))
hits = 0.0
misses = 0.0
# initial_misses = 0.0
unseen_misses = 0.0
starter = ''
terminator = ''
last_app = ''
last_date = ''
# ss = 1
for session in testing_sessions:
"""
if last_date != session[0]['datetime'].day and last_date != '':
if hits + misses > 0:
print ss, hits/(hits + misses)
else:
pass
print 'zero'
# hits = 0.0
# misses = 0.0
ss += 1
# if ss == 7:
# sys.exit()
"""
# print '==='
# if len(session) > 0:
# print '\n'.join(map(lambda x: x['application'], session))
# print
for i in xrange(len(session)):
if ignore_initiator and i == 0:
continue
apeic_candidates, nb_candidates = predictor.predict(session[:i+1], last_app, terminator, k)
assert len(apeic_candidates) <= k and len(nb_candidates) <= k
lu_candidates = lu_predictor.predict(l2, l1, k)
if session[i]['application'] in apeic_candidates:
total_hits += 1.0
hits += 1.0
m += 1.0/(apeic_candidates.index(session[i]['application']) + 1)
else:
total_misses += 1.0
# print i, session[i]['application']
# print '\t', apeic_candidates
misses += 1.0
if i == 0:
initial_misses += 1.0
if session[i]['application'] not in counter.keys() and i == 0:
# print session[i]['application']
unseen_misses += 1.0
if session[i]['application'] in nb_candidates:
nb_total_hits += 1.0
nb_m += 1.0/(nb_candidates.index(session[i]['application']) + 1)
else:
nb_total_misses += 1.0
if session[i]['application'] in lu_candidates:
lu_total_hits += 1.0
lu_m += 1.0/(lu_candidates.index(session[i]['application']) + 1)
else:
lu_total_misses += 1.0
if session[i]['application'] in mru_candidates:
mru_total_hits += 1.0
mru_m += 1.0/(mru_candidates.index(session[i]['application']) + 1)
else:
mru_total_misses += 1.0
if session[i]['application'] in mfu_candidates:
mfu_total_hits += 1.0
mfu_m += 1.0/(mfu_candidates.index(session[i]['application']) + 1)
else:
mfu_total_misses += 1.0
mru_candidates = mru_candidates[1:] + [session[i]['application']]
l2 = l1
l1 = session[i]['application']
last_app = session[i]['application']
starter = session[0]['application']
terminator = session[-1]['application']
last_date = session[0]['datetime'].day
# predictor.update(session)
if hits + misses == 0:
continue
acc = (hits)/(hits + misses)
# print unseen_misses
print acc, hits, misses, initial_misses, unseen_misses
break
print k
print colored('APEIC', 'cyan'), \
(total_hits)/(total_hits + total_misses), m/(total_hits + total_misses), initial_misses, total_hits, total_misses
print colored('NB ', 'cyan'), \
(nb_total_hits)/(nb_total_hits + nb_total_misses), nb_m/(nb_total_hits + nb_total_misses)
print colored('LU ', 'cyan'), \
(lu_total_hits)/(lu_total_hits + lu_total_misses), lu_m/(lu_total_hits + lu_total_misses)
print colored('MRU ', 'cyan'), \
(mru_total_hits)/(mru_total_hits + mru_total_misses), mru_m/(mru_total_hits + mru_total_misses)
print colored('MFU ', 'cyan'), \
(mfu_total_hits)/(mfu_total_hits + mfu_total_misses), mfu_m/(mfu_total_hits + mfu_total_misses)
def main():
hits = 0.0
misses = 0.0
db_helper = ApeicDBHelper()
users = db_helper.get_users()
# users = ['475f258ecc566658']
accuracies = []
trigger_errors = []
for user in users:
if user == '11d1ef9f845ec10e':
continue
sessions = db_helper.get_sessions(user)
training_sessions, testing_sessions = split(sessions, 0.8)
logs = aggregate_sessions(training_sessions)
extractor = FeatureExtractor()
X, y = extractor.generate_training_instances(logs)
nb = MultinomialNB()
nb_predictor = nb.fit(X, y)
sessions = db_helper.get_sessions(user)
print colored(user, attrs=['blink'])
predictor = ApeicPredictor()
for session in training_sessions:
predictor.update(session)
lengths = map(lambda x: len(x), training_sessions)
avg = sum(lengths) / float(len(lengths))
# print avg, len(db_helper.get_used_apps(user))
# hits = 0.0
# misses = 0.0
new = 0.0
trigger_misses = 0.0
error = defaultdict(int)
last_session = testing_sessions[0]
last = ''
for session in testing_sessions:
# print '\n'.join(map(lambda x: x['application'], session))
for i in xrange(len(session)):
# print session[i]['activity'], session[i]['activity_conf']
# if session[i]['application'] in ['com.android.settings', \
# 'com.android.packageinstaller', 'com.htc.android.worldclock', 'com.android.systemui']:
# continue
instance = extractor.transform(last, session[i])
ei = dict(zip(nb_predictor.classes_, nb_predictor.predict_proba(instance)[0]), \
key=operator.itemgetter(1), reverse=True)
candidates = predictor.predict(session[:i],
last_session[-1]['application'],
ei, 5)
if session[i]['application'] in candidates:
hits += 1.0
else:
misses += 1.0
if session[i]['application'] not in ei:
new += 1.0
if i == 0:
# print '\t', session[i]['application']
trigger_misses += 1.0
error[session[i]['application']] += 1
# print
last_session = session
last = session[0]['application']
predictor.update(session)
# logs = aggregate_sessions(training_sessions + [session])
# extractor = FeatureExtractor()
# X, y = extractor.generate_training_instances(logs)
# nb = MultinomialNB()
# nb_predictor = nb.fit(X, y)
acc = (hits) / (hits + misses)
accuracies.append(acc)
trigger_errors.append(trigger_misses / misses if misses > 0 else 0)
print acc, trigger_misses / misses if misses > 0 else 0, trigger_misses, new, misses, (
hits + misses)
# print error
# break
print sum(accuracies) / len(accuracies), sum(trigger_errors) / len(
trigger_errors)
def get_session_len_distr(verbose=True): db_helper = ApeicDBHelper() sessions = chain(*map(lambda x: db_helper.get_sessions(x), db_helper.get_users())) session_lengths = map(lambda x: len(x), sessions) app_numbers = map(lambda x: len(set(map(lambda y: y['application']))), x), sessions))
def test(k=4, ignore_initiator=True):
db_helper = ApeicDBHelper()
users = filter(lambda x: x != '11d1ef9f845ec10e', db_helper.get_users())
total_hits = 0.0
total_misses = 0.0
m = 0.0
initial_misses = 0.0
nb_total_hits = 0.0
nb_total_misses = 0.0
nb_m = 0.0
lu_total_hits = 0.0
lu_total_misses = 0.0
lu_m = 0.0
mru_total_hits = 0.0
mru_total_misses = 0.0
mru_m = 0.0
mfu_total_hits = 0.0
mfu_total_misses = 0.0
mfu_m = 0.0
for user in users:
# print colored(user, attrs=['blink'])
sessions = db_helper.get_sessions(user)
training_sessions, testing_sessions = split(sessions, 0.8)
past_launches = map(lambda x: x['application'],
chain(*training_sessions))
mru_candidates = past_launches[-k:]
l1 = past_launches[-1]
l2 = past_launches[-2]
used_apps = []
for session in training_sessions:
used_apps += map(lambda x: x['application'], session)
counter = Counter(used_apps)
mfu_candidates = map(lambda x: x[0], counter.most_common(k))
predictor = ApeicPredictor()
predictor.train(training_sessions)
lu_predictor = LUPredictor(0.25)
lu_predictor.train(list(chain(*training_sessions)))
hits = 0.0
misses = 0.0
# initial_misses = 0.0
unseen_misses = 0.0
starter = ''
terminator = ''
last_app = ''
last_date = ''
# ss = 1
for session in testing_sessions:
"""
if last_date != session[0]['datetime'].day and last_date != '':
if hits + misses > 0:
print ss, hits/(hits + misses)
else:
pass
print 'zero'
# hits = 0.0
# misses = 0.0
ss += 1
# if ss == 7:
# sys.exit()
"""
# print '==='
# if len(session) > 0:
# print '\n'.join(map(lambda x: x['application'], session))
# print
for i in xrange(len(session)):
if ignore_initiator and i == 0:
continue
apeic_candidates, nb_candidates = predictor.predict(
session[:i + 1], last_app, terminator, k)
assert len(apeic_candidates) <= k and len(nb_candidates) <= k
lu_candidates = lu_predictor.predict(l2, l1, k)
if session[i]['application'] in apeic_candidates:
total_hits += 1.0
hits += 1.0
m += 1.0 / (
apeic_candidates.index(session[i]['application']) + 1)
else:
total_misses += 1.0
# print i, session[i]['application']
# print '\t', apeic_candidates
misses += 1.0
if i == 0:
initial_misses += 1.0
if session[i]['application'] not in counter.keys(
) and i == 0:
# print session[i]['application']
unseen_misses += 1.0
if session[i]['application'] in nb_candidates:
nb_total_hits += 1.0
nb_m += 1.0 / (
nb_candidates.index(session[i]['application']) + 1)
else:
nb_total_misses += 1.0
if session[i]['application'] in lu_candidates:
lu_total_hits += 1.0
lu_m += 1.0 / (
lu_candidates.index(session[i]['application']) + 1)
else:
lu_total_misses += 1.0
if session[i]['application'] in mru_candidates:
mru_total_hits += 1.0
mru_m += 1.0 / (
mru_candidates.index(session[i]['application']) + 1)
else:
mru_total_misses += 1.0
if session[i]['application'] in mfu_candidates:
mfu_total_hits += 1.0
mfu_m += 1.0 / (
mfu_candidates.index(session[i]['application']) + 1)
else:
mfu_total_misses += 1.0
mru_candidates = mru_candidates[1:] + [
session[i]['application']
]
l2 = l1
l1 = session[i]['application']
last_app = session[i]['application']
starter = session[0]['application']
terminator = session[-1]['application']
last_date = session[0]['datetime'].day
# predictor.update(session)
if hits + misses == 0:
continue
acc = (hits) / (hits + misses)
# print unseen_misses
print acc, hits, misses, initial_misses, unseen_misses
break
print k
print colored('APEIC', 'cyan'), \
(total_hits)/(total_hits + total_misses), m/(total_hits + total_misses), initial_misses, total_hits, total_misses
print colored('NB ', 'cyan'), \
(nb_total_hits)/(nb_total_hits + nb_total_misses), nb_m/(nb_total_hits + nb_total_misses)
print colored('LU ', 'cyan'), \
(lu_total_hits)/(lu_total_hits + lu_total_misses), lu_m/(lu_total_hits + lu_total_misses)
print colored('MRU ', 'cyan'), \
(mru_total_hits)/(mru_total_hits + mru_total_misses), mru_m/(mru_total_hits + mru_total_misses)
print colored('MFU ', 'cyan'), \
(mfu_total_hits)/(mfu_total_hits + mfu_total_misses), mfu_m/(mfu_total_hits + mfu_total_misses)
def main():
hits = 0.0
misses = 0.0
db_helper = ApeicDBHelper()
users = db_helper.get_users()
# users = ['475f258ecc566658']
accuracies = []
trigger_errors = []
for user in users:
if user == '11d1ef9f845ec10e':
continue
sessions = db_helper.get_sessions(user)
training_sessions, testing_sessions = split(sessions, 0.8)
logs = aggregate_sessions(training_sessions)
extractor = FeatureExtractor()
X, y = extractor.generate_training_instances(logs)
nb = MultinomialNB()
nb_predictor = nb.fit(X, y)
sessions = db_helper.get_sessions(user)
print colored(user, attrs=['blink'])
predictor = ApeicPredictor()
for session in training_sessions:
predictor.update(session)
lengths = map(lambda x: len(x), training_sessions)
avg = sum(lengths)/float(len(lengths))
# print avg, len(db_helper.get_used_apps(user))
# hits = 0.0
# misses = 0.0
new = 0.0
trigger_misses = 0.0
error = defaultdict(int)
last_session = testing_sessions[0]
last = ''
for session in testing_sessions:
# print '\n'.join(map(lambda x: x['application'], session))
for i in xrange(len(session)):
# print session[i]['activity'], session[i]['activity_conf']
# if session[i]['application'] in ['com.android.settings', \
# 'com.android.packageinstaller', 'com.htc.android.worldclock', 'com.android.systemui']:
# continue
instance = extractor.transform(last, session[i])
ei = dict(zip(nb_predictor.classes_, nb_predictor.predict_proba(instance)[0]), \
key=operator.itemgetter(1), reverse=True)
candidates = predictor.predict(session[:i], last_session[-1]['application'], ei, 5)
if session[i]['application'] in candidates:
hits += 1.0
else:
misses += 1.0
if session[i]['application'] not in ei:
new += 1.0
if i == 0:
# print '\t', session[i]['application']
trigger_misses += 1.0
error[session[i]['application']] += 1
# print
last_session = session
last = session[0]['application']
predictor.update(session)
# logs = aggregate_sessions(training_sessions + [session])
# extractor = FeatureExtractor()
# X, y = extractor.generate_training_instances(logs)
# nb = MultinomialNB()
# nb_predictor = nb.fit(X, y)
acc = (hits)/(hits + misses)
accuracies.append(acc)
trigger_errors.append(trigger_misses/misses if misses > 0 else 0)
print acc, trigger_misses/misses if misses > 0 else 0, trigger_misses, new, misses, (hits + misses)
# print error
# break
print sum(accuracies)/len(accuracies), sum(trigger_errors)/len(trigger_errors)