In [1]:
%matplotlib inline
import os
import sys
import gzip
import numpy as np
import pickle as pkl
from scipy.sparse import lil_matrix, issparse, hstack, vstack
from collections import Counter
import gensim
import matplotlib.pyplot as plt
import seaborn as sns
from tqdm import tqdm
In [2]:
np_settings0 = np.seterr(all='raise')
RAND_SEED = 0
n_feature_artist = 30
plt.style.use('seaborn')
In [3]:
datasets = ['aotm2011', '30music']
ffeature = 'data/msd/song2feature.pkl.gz'
fgenre = 'data/msd/song2genre.pkl.gz'
fsong2artist = 'data/msd/song2artist.pkl.gz'
audio_feature_indices = [20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 185, 186, 187, 198, 199, 200, 201]
test_user_ratio = 0.3
In [4]:
dix = 1
dataset_name = datasets[dix]
data_dir = 'data/%s' % dataset_name
print(dataset_name)
Load playlists.
In [5]:
fplaylist = os.path.join(data_dir, '%s-playlist.pkl.gz' % dataset_name)
_all_playlists = pkl.load(gzip.open(fplaylist, 'rb'))
In [6]:
# _all_playlists[0]
In [7]:
all_playlists = []
if type(_all_playlists[0][1]) == tuple:
for pl, u in _all_playlists:
user = '%s_%s' % (u[0], u[1]) # user string
all_playlists.append((pl, user))
else:
all_playlists = _all_playlists
In [8]:
# user_playlists = dict()
# for pl, u in all_playlists:
# try:
# user_playlists[u].append(pl)
# except KeyError:
# user_playlists[u] = [pl]
In [9]:
# all_playlists = []
# for u in user_playlists:
# if len(user_playlists[u]) > 4:
# all_playlists += [(pl, u) for pl in user_playlists[u]]
In [10]:
all_users = sorted(set({user for _, user in all_playlists}))
In [11]:
print('#user : {:,}'.format(len(all_users)))
print('#playlist: {:,}'.format(len(all_playlists)))
In [12]:
pl_lengths = [len(pl) for pl, _ in all_playlists]
plt.hist(pl_lengths, bins=100)
print('Average playlist length: %.1f' % np.mean(pl_lengths))
check duplicated songs in the same playlist.
In [13]:
print('{:,} | {:,}'.format(np.sum(pl_lengths), np.sum([len(set(pl)) for pl, _ in all_playlists])))
Load song_id --> feature array mapping: map a song to the audio features of one of its corresponding tracks in MSD.
In [14]:
_song2feature = pkl.load(gzip.open(ffeature, 'rb'))
In [15]:
song2feature = dict()
for sid in sorted(_song2feature):
song2feature[sid] = _song2feature[sid][audio_feature_indices]
Song genres from MSD Allmusic Genre Dataset (Top MAGD) and tagtraum.
In [16]:
song2genre = pkl.load(gzip.open(fgenre, 'rb'))
In [17]:
_all_songs = sorted([(sid, int(song2feature[sid][-1])) for sid in {s for pl, _ in all_playlists for s in pl}],
key=lambda x: (x[1], x[0]))
print('{:,}'.format(len(_all_songs)))
Randomise the order of song with the same age.
In [18]:
song_age_dict = dict()
for sid, age in _all_songs:
age = int(age)
try:
song_age_dict[age].append(sid)
except KeyError:
song_age_dict[age] = [sid]
In [19]:
all_songs = []
np.random.seed(RAND_SEED)
for age in sorted(song_age_dict.keys()):
all_songs += [(sid, age) for sid in np.random.permutation(song_age_dict[age])]
In [20]:
pkl.dump(all_songs, gzip.open(os.path.join(data_dir, 'setting4/all_songs.pkl.gz'), 'wb'))
Check if all songs have genre info.
In [21]:
print('#songs missing genre: {:,}'.format(len(all_songs) - np.sum([sid in song2genre for (sid, _) in all_songs])))
Song popularity.
In [22]:
song2index = {sid: ix for ix, (sid, _) in enumerate(all_songs)}
song_pl_mat = lil_matrix((len(all_songs), len(all_playlists)), dtype=np.int8)
for j in range(len(all_playlists)):
pl = all_playlists[j][0]
ind = [song2index[sid] for sid in pl]
song_pl_mat[ind, j] = 1
In [23]:
song_pop = song_pl_mat.tocsc().sum(axis=1)
In [24]:
max_pop = np.max(song_pop)
max_pop
Out[24]:
In [25]:
song2pop = {sid: song_pop[song2index[sid], 0] for (sid, _) in all_songs}
In [26]:
pkl.dump(song2pop, gzip.open(os.path.join(data_dir, 'setting4/song2pop.pkl.gz'), 'wb'))
Songs as rows, playlists as columns.
In [27]:
def gen_dataset(playlists, song2feature, song2genre, song2artist, artist2vec,
train_song_set, dev_song_set=[], test_song_set=[], song2pop_train=None):
"""
Create labelled dataset: rows are songs, columns are users.
Input:
- playlists: a set of playlists
- train_song_set: a list of songIDs in training set
- dev_song_set: a list of songIDs in dev set
- test_song_set: a list of songIDs in test set
- song2feature: dictionary that maps songIDs to features from MSD
- song2genre: dictionary that maps songIDs to genre
- song2pop_train: a dictionary that maps songIDs to its popularity
Output:
- (Feature, Label) pair (X, Y)
X: #songs by #features
Y: #songs by #users
"""
song_set = train_song_set + dev_song_set + test_song_set
N = len(song_set)
K = len(playlists)
genre_set = sorted({v for v in song2genre.values()})
genre2index = {genre: ix for ix, genre in enumerate(genre_set)}
def onehot_genre(songID):
"""
One-hot encoding of genres.
Data imputation:
- mean imputation (default)
- one extra entry for songs without genre info
- sampling from the distribution of genre popularity
"""
num = len(genre_set) # + 1
vec = np.zeros(num, dtype=np.float)
if songID in song2genre:
genre_ix = genre2index[song2genre[songID]]
vec[genre_ix] = 1
else:
vec[:] = np.nan
#vec[-1] = 1
return vec
def song_artist_feature(songID):
"""
Return the artist feature for a given song
"""
if songID in song2artist:
aid = song2artist[songID]
return artist2vec[aid]
else:
return artist2vec['$UNK$']
X = np.array([np.concatenate([song2feature[sid], song_artist_feature(sid), onehot_genre(sid)], axis=-1) \
for sid in song_set])
Y = lil_matrix((N, K), dtype=np.bool)
song2index = {sid: ix for ix, sid in enumerate(song_set)}
for k in range(K):
pl = playlists[k]
indices = [song2index[sid] for sid in pl if sid in song2index]
Y[indices, k] = True
# genre imputation
genre_ix_start = -len(genre_set)
genre_nan = np.isnan(X[:, genre_ix_start:])
genre_mean = np.nansum(X[:, genre_ix_start:], axis=0) / (X.shape[0] - np.sum(genre_nan, axis=0))
#print(np.nansum(X[:, genre_ix_start:], axis=0))
#print(genre_set)
#print(genre_mean)
for j in range(len(genre_set)):
X[genre_nan[:,j], j+genre_ix_start] = genre_mean[j]
# normalise the sum of all genres per song to 1
# X[:, -len(genre_set):] /= X[:, -len(genre_set):].sum(axis=1).reshape(-1, 1)
# NOTE: this is not necessary, as the imputed values are guaranteed to be normalised (sum to 1)
# due to the above method to compute mean genres.
# the log of song popularity
if song2pop_train is not None:
# for sid in song_set:
# assert sid in song2pop_train # trust the input
logsongpop = np.log2([song2pop_train[sid]+1 for sid in song_set]) # deal with 0 popularity
X = np.hstack([X, logsongpop.reshape(-1, 1)])
#return X, Y
Y = Y.tocsr()
train_ix = [song2index[sid] for sid in train_song_set]
X_train = X[train_ix, :]
Y_train = Y[train_ix, :]
dev_ix = [song2index[sid] for sid in dev_song_set]
X_dev = X[dev_ix, :]
Y_dev = Y[dev_ix, :]
test_ix = [song2index[sid] for sid in test_song_set]
X_test = X[test_ix, :]
Y_test = Y[test_ix, :]
if len(dev_song_set) > 0:
if len(test_song_set) > 0:
return X_train, Y_train.tocsc(), X_dev, Y_dev.tocsc(), X_test, Y_test.tocsc()
else:
return X_train, Y_train.tocsc(), X_dev, Y_dev.tocsc()
else:
if len(test_song_set) > 0:
return X_train, Y_train.tocsc(), X_test, Y_test.tocsc()
else:
return X_train, Y_train.tocsc()
Split playlists such that
In [28]:
user_playlists = dict()
for j in range(len(all_playlists)):
u = all_playlists[j][1]
try:
user_playlists[u].append(j)
except KeyError:
user_playlists[u] = [j]
In [29]:
# sanity check
npl_all = np.sum([len(user_playlists[u]) for u in user_playlists])
print('{:30s} {:,}'.format('#users:', len(user_playlists)))
print('{:30s} {:,}'.format('#playlists:', npl_all))
print('{:30s} {:.2f}'.format('Average #playlists per user:', npl_all / len(user_playlists)))
Make sure every song in test set should also be in training set.
In [30]:
user_songcnt = dict()
for u in all_users:
songcnt_u = np.zeros(len(all_songs), dtype=np.int32)
for pl, _ in [all_playlists[j] for j in user_playlists[u]]:
for sid in pl:
songcnt_u[song2index[sid]] += 1
user_songcnt[u] = songcnt_u
In [31]:
all_songcnt = np.zeros(len(all_songs), dtype=np.int32)
for u in all_users:
all_songcnt += user_songcnt[u]
In [32]:
candidate_users = set()
other_users = set()
for u in all_users:
_songcnt = all_songcnt - user_songcnt[u]
if np.all(_songcnt > 0):
candidate_users.add(u)
else:
other_users.add(u)
In [33]:
print(len(candidate_users), len(other_users), len(all_users))
In [34]:
npl_candidate = sorted([len(user_playlists[u]) for u in candidate_users])
In [35]:
print('%d, %.1f, %d, %d' % (min(npl_candidate), np.mean(npl_candidate), max(npl_candidate), np.sum(npl_candidate)))
In [36]:
train_users = set(all_users)
test_users = set()
train_songcnt = all_songcnt.copy()
np.random.seed(RAND_SEED)
for u in np.random.permutation(sorted(candidate_users)):
_songcnt = train_songcnt - user_songcnt[u]
if np.all(_songcnt > 0):
train_users = train_users - {u}
test_users.add(u)
train_songcnt[:] = _songcnt
npl_test = np.sum([len(user_playlists[u]) for u in test_users])
if len(test_users) >= int(test_user_ratio * len(all_users)):
break
In [37]:
train_playlists = [all_playlists[j] for u in sorted(train_users) for j in user_playlists[u]]
test_playlists = [all_playlists[j] for u in sorted(test_users) for j in user_playlists[u]]
Sanity check that every song in test set should also be in training set.
In [38]:
print('#Songs in train set: %d, #Songs total: %d' % \
(len(set([sid for pl, _ in train_playlists for sid in pl])), len(all_songs)))
In [39]:
print('{:30s} {:,}'.format('#playlists (train):', len(train_playlists)))
print('{:30s} {:,}'.format('#playlists (test) :', len(test_playlists)))
print('{:30s} {:,} out of {:,}'.format('#users in test set:', len(test_users), len(all_users)))
In [40]:
assert 0 == len(train_users & test_users)
print('#users (train): {:,}'.format(len(train_users)))
In [41]:
xmax = np.max([len(pl) for (pl, _) in all_playlists]) + 1
In [42]:
ax = plt.subplot(111)
ax.hist([len(pl) for (pl, _) in train_playlists], bins=100)
ax.set_yscale('log')
ax.set_xlim(0, xmax)
ax.set_title('Histogram of playlist length in TRAINING set')
pass
In [43]:
ax = plt.subplot(111)
ax.hist([len(pl) for (pl, _) in test_playlists], bins=100)
ax.set_yscale('log')
ax.set_xlim(0, xmax)
ax.set_title('Histogram of playlist length in TEST set')
pass
In [44]:
song2pop_train = song2pop.copy()
for pl, _ in test_playlists:
for sid in pl:
song2pop_train[sid] -= 1
In [45]:
pkl.dump(song2pop_train, gzip.open(os.path.join(data_dir, 'setting4/song2pop_train.pkl.gz'), 'wb'))
In [46]:
song2artist = pkl.load(gzip.open(fsong2artist, 'rb'))
In [47]:
artist_playlist = []
for pl, _ in train_playlists:
pl_artists = [song2artist[sid] if sid in song2artist else '$UNK$' for sid in pl]
artist_playlist.append(pl_artists)
In [48]:
fartist2vec_bin = os.path.join(data_dir, 'setting4/artist2vec.bin')
if os.path.exists(fartist2vec_bin):
artist2vec = gensim.models.KeyedVectors.load_word2vec_format(fartist2vec_bin, binary=True)
else:
artist2vec_model = gensim.models.Word2Vec(sentences=artist_playlist, size=n_feature_artist, seed=RAND_SEED,
window=10, iter=10, min_count=1)
artist2vec_model.wv.save_word2vec_format(fartist2vec_bin, binary=True)
artist2vec = artist2vec_model.wv
In [5]:
pkl_dir = os.path.join(data_dir, 'coldstart/setting4')
fpl = os.path.join(pkl_dir, 'playlists_train_test_s4.pkl.gz')
fx = os.path.join(pkl_dir, 'X.pkl.gz')
fytrain = os.path.join(pkl_dir, 'Y_train.pkl.gz')
fytest = os.path.join(pkl_dir, 'Y_test.pkl.gz')
fclique_train = os.path.join(pkl_dir, 'cliques_train.pkl.gz')
fclique_all = os.path.join(pkl_dir, 'cliques_all.pkl.gz')
In [50]:
X, Y = gen_dataset(playlists = [t[0] for t in train_playlists + test_playlists],
song2feature = song2feature, song2genre = song2genre,
song2artist = song2artist, artist2vec = artist2vec,
train_song_set = [t[0] for t in all_songs], song2pop_train=song2pop_train)
In [51]:
split_ix = len(train_playlists)
In [52]:
Y_train = Y[:, :split_ix].tocsc()
Y_test = Y[:, split_ix:].tocsc()
In [53]:
assert Y_train.shape[0] == Y_test.shape[0] == X.shape[0] == len(all_songs)
assert Y_train.shape[1] + Y_test.shape[1] == Y.shape[1] == len(all_playlists)
In [54]:
pkl.dump({'train_playlists': train_playlists, 'test_playlists': test_playlists}, gzip.open(fpl, 'wb'))
Feature normalisation.
In [55]:
X_mean = np.mean(X, axis=0).reshape((1, -1))
X_std = np.std(X, axis=0).reshape((1, -1)) + 10 ** (-6)
X -= X_mean
X /= X_std
In [56]:
print(np.mean(np.mean(X, axis=0)))
print(np.mean( np.std(X, axis=0)) - 1)
In [57]:
print('Train :', Y_train.shape)
print('Test :', Y_test.shape)
print('All: %s, %s' % (X.shape, Y.shape))
In [58]:
pkl.dump(X, gzip.open(fx, 'wb'))
pkl.dump(Y_train, gzip.open(fytrain, 'wb'))
pkl.dump(Y_test, gzip.open(fytest, 'wb'))
Playlists of the same user form a clique.
Cliques in train set.
In [59]:
pl_users = [u for (_, u) in train_playlists]
cliques_train = []
for u in sorted(set(pl_users)):
clique = np.where(u == np.array(pl_users, dtype=np.object))[0]
cliques_train.append(clique)
In [60]:
pkl.dump(cliques_train, gzip.open(fclique_train, 'wb'))
In [61]:
clqsize = [len(clq) for clq in cliques_train]
print(np.min(clqsize), np.max(clqsize), len(clqsize), np.sum(clqsize))
In [62]:
assert np.all(np.arange(Y_train.shape[1]) == np.asarray(sorted([k for clq in cliques_train for k in clq])))
In [ ]:
In [7]:
pldata = pkl.load(gzip.open(fpl, 'rb'))
train_playlists = pldata['train_playlists']
test_playlists = pldata['test_playlists']
In [8]:
pl_users = [u for (_, u) in train_playlists + test_playlists]
train_users = [u for (_, u) in train_playlists]
test_users = [u for (_, u) in test_playlists]
pl_users = train_users + test_users
user_set = sorted(set(train_users)) + sorted(set(test_users))
clique_all = []
for u in user_set:
clique = np.where(u == np.array(pl_users, dtype=np.object))[0]
#if len(clique) > 1:
clique_all.append(clique)
In [9]:
clqsize = [len(clq) for clq in clique_all]
print(np.min(clqsize), np.max(clqsize), len(clqsize), np.sum(clqsize))
In [10]:
Y_train = pkl.load(gzip.open(fytrain, 'rb'))
Y_test = pkl.load(gzip.open(fytest, 'rb'))
N = Y_train.shape[1] + Y_test.shape[1]
assert np.all(np.arange(N) == np.asarray(sorted([k for clq in clique_all for k in clq])))
In [11]:
pkl.dump(clique_all, gzip.open(fclique_all, 'wb'))