def modify(self,
song,
sub_value_dict,
star_threshold=0.75,
iterstep_threshold=50):
value_dict = song.to_dict()
for key, value in sub_value_dict.items():
try:
provider = self._session.provider_for_key(key)
if value is None:
sub_value_dict[key] = None
else:
sub_value_dict[key] = provider.process(value)
except KeyError:
raise KeyError('key "{k}" is not in mask'.format(k=key))
value_dict.update(sub_value_dict)
new_song = Song(self._session,
value_dict,
max_neighbors=self._session.config['max_neighbors'],
max_distance=self._session.config['max_distance'])
new_song.uid = self.remove(song.uid)
self._song_list[song.uid] = new_song
# Clear all know distances:
new_song.distance_reset()
return self._insert_song_to_graph(new_song, star_threshold,
iterstep_threshold)
def modify(self, song, sub_value_dict, star_threshold=0.75, iterstep_threshold=50):
value_dict = song.to_dict()
for key, value in sub_value_dict.items():
try:
provider = self._session.provider_for_key(key)
if value is None:
sub_value_dict[key] = None
else:
sub_value_dict[key] = provider.process(value)
except KeyError:
raise KeyError('key "{k}" is not in mask'.format(k=key))
value_dict.update(sub_value_dict)
new_song = Song(
self._session, value_dict,
max_neighbors=self._session.config['max_neighbors'],
max_distance=self._session.config['max_distance']
)
new_song.uid = self.remove(song.uid)
self._song_list[song.uid] = new_song
# Clear all know distances:
new_song.distance_reset()
return self._insert_song_to_graph(
new_song, star_threshold, iterstep_threshold
)
def rebuild_stupid(self):
"""(Re)build the graph by calculating the combination of all songs.
This is a *very* expensive operation which takes quadratic time and
only should be ever used for a small amount of songs where accuracy
matters even more thant time.
"""
for song_a, song_b in combinations(self._song_list, 2):
distance = Song.distance_compute(song_a, song_b)
Song.distance_add(song_a, song_b, distance)
def rebuild_stupid(self):
"""(Re)build the graph by calculating the combination of all songs.
This is a *very* expensive operation which takes quadratic time and
only should be ever used for a small amount of songs where accuracy
matters even more thant time.
"""
for song_a, song_b in combinations(self._song_list, 2):
distance = Song.distance_compute(song_a, song_b)
Song.distance_add(song_a, song_b, distance)
def _insert_song_to_graph(self,
new_song,
star_threshold=0.75,
iterstep_threshold=50):
next_len = len(self._song_list)
if len(self) < iterstep_threshold:
iterstep = 1
else:
iterstep = round(max(1, math.log(max(next_len, 1))))
# Step 1: Find samples with similar songs (similar to the base step)
distances = deque()
for song in self._song_list[::iterstep]:
if song is not None:
distance = Song.distance_compute(song, new_song)
distances.append((song, distance))
new_song.distance_add(song, distance)
# Step 2: Short refinement step
for song, distance in distances:
if distance.distance > star_threshold:
for neighbor in song.neighbors():
distance = new_song.distance_compute(neighbor)
new_song.distance_add(neighbor, distance)
return new_song.uid
def test_recommendation_history(self):
history = RecommendationHistory()
session = Session('test',
{'artist': (None, None, 1), 'album': (None, None, 1)}
)
fst_song = Song(session, {
'artist': 'A',
'album': 'B'
})
self.assertEqual(history.allowed(fst_song), True)
history.feed(fst_song)
self.assertEqual(history.allowed(fst_song), False)
for expectation in [False, False, False, False, True]:
history.feed(Song(session, {
'artist': 'X',
'album': 'Y'
}))
self.assertEqual(history.allowed(fst_song), expectation)
def test_count_keys(self):
history = History(maxlen=19)
for _ in range(2000):
history.feed(Song(self._session, {choice('abcdef'): 1.0}))
counter = history.count_keys()
for char in 'abdef':
self.assertTrue(char in counter)
self.assertEqual(sum(counter.values()), 100)
self.assertEqual(len(list(history.groups())), 20)
for group in history.groups():
self.assertEqual(len(list(group)), 5)
def add(self, value_dict):
for key, value in value_dict.items():
try:
provider = self._session.provider_for_key(key)
if value is None:
value_dict[key] = None
else:
value_dict[key] = provider.process(value)
except KeyError:
raise KeyError('key "{k}" is not in mask'.format(k=key))
new_song = Song(self._session,
value_dict,
max_neighbors=self._session.config['max_neighbors'],
max_distance=self._session.config['max_distance'])
new_song.uid = self._current_uid()
if new_song.uid >= len(self._song_list):
self._song_list.append(new_song)
else:
self._song_list[new_song.uid] = new_song
return new_song.uid
def add(self, value_dict):
for key, value in value_dict.items():
try:
provider = self._session.provider_for_key(key)
if value is None:
value_dict[key] = None
else:
value_dict[key] = provider.process(value)
except KeyError:
raise KeyError('key "{k}" is not in mask'.format(k=key))
new_song = Song(
self._session, value_dict,
max_neighbors=self._session.config['max_neighbors'],
max_distance=self._session.config['max_distance']
)
new_song.uid = self._current_uid()
if new_song.uid >= len(self._song_list):
self._song_list.append(new_song)
else:
self._song_list[new_song.uid] = new_song
return new_song.uid
def test_relim(self):
history = ListenHistory()
songs = [Song(self._session, {'abcdef'[idx]: 1.0}) for idx in range(6)]
for idx, song in enumerate(songs):
song.uid = idx
N = 10000
for _ in range(N):
for i, ilem in enumerate(songs):
history.feed(ilem)
for j, jlem in enumerate(songs[i:]):
history.feed(jlem)
itemsets = history.frequent_itemsets()
print()
print('==================')
print('FREQUENT ITEMSETS:')
print('==================')
print()
for itemset, support in sorted(itemsets.items(), key=lambda x: x[1]):
print('{: 8d} ({:3.3f}%): {:>20s}'.format(
support, support / N * 10,
str([song.uid for song in itemset])
))
print()
print('==================')
print('ASSOCIATION RULES:')
print('==================')
print()
rules = history.find_rules(itemsets)
for left, right, support, rating in rules:
print('{:>15s} <-> {:<15s} [supp={:> 5d}, rating={:.5f}]'.format(
str([song.uid for song in left]),
str([song.uid for song in right]),
support, rating
))
def _insert_song_to_graph(self, new_song, star_threshold=0.75, iterstep_threshold=50):
next_len = len(self._song_list)
if len(self) < iterstep_threshold:
iterstep = 1
else:
iterstep = round(max(1, math.log(max(next_len, 1))))
# Step 1: Find samples with similar songs (similar to the base step)
distances = deque()
for song in self._song_list[::iterstep]:
if song is not None:
distance = Song.distance_compute(song, new_song)
distances.append((song, distance))
new_song.distance_add(song, distance)
# Step 2: Short refinement step
for song, distance in distances:
if distance.distance > star_threshold:
for neighbor in song.neighbors():
distance = new_song.distance_compute(neighbor)
new_song.distance_add(neighbor, distance)
return new_song.uid
