def fit(self,
topK=50,
shrink=100,
similarity='cosine',
feature_weighting="none",
**similarity_args):
# Similaripy returns also self similarity, which will be set to 0 afterwards
topK += 1
self.topK = topK
self.shrink = shrink
if feature_weighting not in self.FEATURE_WEIGHTING_VALUES:
raise ValueError(
"Value for 'feature_weighting' not recognized. Acceptable values are {}, provided was '{}'"
.format(self.FEATURE_WEIGHTING_VALUES, feature_weighting))
if feature_weighting == "BM25":
self.URM_train = self.URM_train.astype(np.float32)
self.URM_train = okapi_BM_25(self.URM_train.T).T
self.URM_train = check_matrix(self.URM_train, 'csr')
elif feature_weighting == "TF-IDF":
self.URM_train = self.URM_train.astype(np.float32)
self.URM_train = TF_IDF(self.URM_train.T).T
self.URM_train = check_matrix(self.URM_train, 'csr')
if similarity == "cosine":
self.W_sparse = sim.cosine(self.URM_train,
k=topK,
shrink=shrink,
**similarity_args)
elif similarity == "jaccard":
self.W_sparse = sim.jaccard(self.URM_train,
k=topK,
shrink=shrink,
**similarity_args)
elif similarity == "dice":
self.W_sparse = sim.dice(self.URM_train,
k=topK,
shrink=shrink,
**similarity_args)
elif similarity == "jaccard":
self.W_sparse = sim.tversky(self.URM_train,
k=topK,
shrink=shrink,
**similarity_args)
elif similarity == "splus":
self.W_sparse = sim.s_plus(self.URM_train,
k=topK,
shrink=shrink,
**similarity_args)
else:
raise ValueError(
"Unknown value '{}' for similarity".format(similarity))
self.W_sparse.setdiag(0)
self.W_sparse = self.W_sparse.transpose().tocsr()
def fit(self,
topK=50,
shrink=100,
similarity='cosine',
feature_weighting="none"):
self.topK = topK
self.shrink = shrink
if feature_weighting == "bm25":
self.ICM_train = similaripy.normalization.bm25(self.ICM_train)
elif feature_weighting == "bm25plus":
self.ICM_train = similaripy.normalization.bm25plus(self.ICM_train)
elif feature_weighting == "tfidf":
self.ICM_train = similaripy.normalization.tfidf(self.ICM_train)
if similarity == "cosine":
similarity_matrix = similaripy.cosine(self.ICM_train,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
if similarity == "s_plus":
similarity_matrix = similaripy.s_plus(self.ICM_train,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
if similarity == "dice":
similarity_matrix = similaripy.dice(self.ICM_train,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
if similarity == "rp3beta":
similarity_matrix = similaripy.rp3beta(self.ICM_train,
alpha=0.3,
beta=0.61,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
if similarity == "p3alpha":
similarity_matrix = similaripy.p3alpha(self.ICM_train,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
if similarity == "jaccard":
similarity_matrix = similaripy.jaccard(self.ICM_train,
k=self.topK,
shrink=self.shrink,
binary=False,
verbose=False)
self.W_sparse = similarity_matrix.transpose().tocsr()
self.W_sparse = check_matrix(self.W_sparse, format='csr')
def fit(self, topK=50, shrink=100, similarity='cosine', pre_normalization="none", post_normalization = "none", **similarity_args):
self.topK = topK
self.shrink = shrink
#ucm = sps.load_npz("FULL_UCM.npz")
interactions = self.URM_train
if pre_normalization == "bm25plus":
interactions = similaripy.normalization.bm25plus(self.URM_train, axis=1, k1=1.2, b=0.75, delta=0.85, tf_mode="raw", idf_mode="bm25", inplace=False)
if pre_normalization == "tfidf":
interactions = similaripy.normalization.tfidf(self.URM_train, axis=1)
#interactions = sps.hstack((interactions, ucm))
if similarity == "cosine":
similarity_matrix = similaripy.cosine(interactions, k=self.topK, shrink=self.shrink, binary=False, verbose=False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if similarity == "s_plus":
similarity_matrix = similaripy.s_plus(interactions, k=self.topK, shrink=self.shrink, binary = False, verbose = False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if similarity == "dice":
similarity_matrix = similaripy.dice(interactions, k=self.topK, shrink=self.shrink, binary=False, verbose=False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if similarity == "rp3beta":
similarity_matrix = similaripy.rp3beta(interactions, alpha=0.3, beta=0.61, k=self.topK, shrink=self.shrink, binary=False, verbose=False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if similarity == "asym":
similarity_matrix = similaripy.asymmetric_cosine(interactions, k=self.topK, shrink=self.shrink, alpha=0.5, binary=False, verbose=False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if similarity == "jaccard":
similarity_matrix = similaripy.jaccard(interactions, k=self.topK, shrink=self.shrink, binary=False, verbose=False)
similarity_matrix = similarity_matrix.transpose().tocsr()
if post_normalization == "bm25plus_once":
self.URM_train = similaripy.normalization.bm25plus(self.URM_train, axis=1, k1=1.2, b=0.75, delta=0.8, tf_mode='raw', idf_mode='bm25', inplace=False)
if post_normalization == "bm25plus_twice":
self.URM_train = similaripy.normalization.bm25plus(self.URM_train, axis=1, k1=1.2, b=0.75, delta=0.8, tf_mode='raw', idf_mode='bm25', inplace=False)
self.URM_train = similaripy.normalization.bm25plus(self.URM_train, axis=1, k1=1.2, b=0.75, delta=0.8, tf_mode='raw', idf_mode='bm25', inplace=False)
if post_normalization == "tfidf":
self.URM_train = similaripy.normalization.tfidf(self.URM_train, axis=1)
if post_normalization == "bm25":
self.URM_train = similaripy.normalization.bm25(self.URM_train, axis=1)
self.W_sparse = similarity_matrix
self.W_sparse = check_matrix(self.W_sparse, format='csr')
choices=['thresh', 'knn'],
type=str,
required=True)
args = parser.parse_args()
# first load the data
df_train = pd.read_csv(f'../dataset/{args.split}/train.csv', escapechar="\\")
df_test = pd.read_csv(f'../dataset/{args.split}/test.csv', escapechar="\\")
# ALWAYS sort the data by record_id
df_train = df_train.sort_values(by=['record_id']).reset_index(drop=True)
df_test = df_test.sort_values(by=['record_id']).reset_index(drop=True)
df_train.address = df_train.address.fillna('').astype(str)
df_test.address = df_test.address.fillna('').astype(str)
corpus = list(df_train.address) + list(df_test.address)
vectorizer = CountVectorizer(preprocessor=remove_spaces,
analyzer=remove_spaces)
X = vectorizer.fit_transform(corpus)
X_train = X[:df_train.shape[0], :]
X_test = X[df_train.shape[0]:, :]
if args.mode == 'thresh':
cosmatrixxx = sim.jaccard(X_test, X_train.T, k=2000)
cosmatrixxx.data[cosmatrixxx.data <= 0.3] = 0
else:
cosmatrixxx = sim.jaccard(X_test, X_train.T, k=300)
if not os.path.isdir(f"../dataset/{args.split}/similarities"):
os.makedirs(f"../dataset/{args.split}/similarities")
save_npz(
f'../dataset/{args.split}/similarities/jaccard_uncleaned_address_300k_{args.split}_2ngrams.npz',
cosmatrixxx.tocsr())
def fit(self,
topK=50,
shrink=100,
similarity='cosine',
normalization="none",
feature_weighting="none",
**similarity_args):
self.topK = topK
self.shrink = shrink
reader = DataReader()
icm = reader.load_icm()
if normalization == "bm25":
self.URM_train = similaripy.normalization.bm25(self.URM_train,
axis=1)
if normalization == "tfidf":
self.URM_train = similaripy.normalization.tfidf(self.URM_train,
axis=1)
if normalization == "bm25plus":
self.URM_train = similaripy.normalization.bm25plus(self.URM_train,
axis=1)
if feature_weighting == "bm25":
icm = similaripy.normalization.bm25(icm, axis=1)
if feature_weighting == "tfidf":
icm = similaripy.normalization.tfidf(icm, axis=1)
if feature_weighting == "bm25plus":
icm = similaripy.normalization.bm25plus(icm, axis=1)
matrix = sps.hstack((self.URM_train.transpose().tocsr(), icm))
if similarity == "cosine":
similarity_matrix = similaripy.cosine(matrix,
k=self.topK,
shrink=self.shrink,
binary=False,
threshold=0)
if similarity == "dice":
similarity_matrix = similaripy.dice(matrix,
k=self.topK,
shrink=self.shrink,
binary=False,
threshold=0)
if similarity == "jaccard":
similarity_matrix = similaripy.jaccard(matrix,
k=self.topK,
shrink=self.shrink,
binary=False,
threshold=0)
if similarity == "asym":
similarity_matrix = similaripy.asymmetric_cosine(
matrix,
k=self.topK,
shrink=self.shrink,
binary=False,
threshold=0)
if similarity == "rp3beta":
similarity_matrix = similaripy.rp3beta(matrix,
k=self.topK,
shrink=self.shrink,
binary=False,
threshold=0,
alpha=0.3,
beta=0.61)
self.W_sparse = similarity_matrix
self.W_sparse = check_matrix(self.W_sparse, format='csr')
def check_similarity(m, k, rtol=0.0001, full=False):
# cython
dot = sim.dot_product(m, k=k)
cosine = sim.cosine(m, k=k)
asy_cosine = sim.asymmetric_cosine(m, alpha=0.2, k=k)
jaccard = sim.jaccard(m, k=k)
dice = sim.dice(m, k=k)
tversky = sim.tversky(m, alpha=0.8, beta=0.4, k=k)
p3alpha = sim.p3alpha(m, alpha=0.8, k=k)
rp3beta = sim.rp3beta(m, alpha=0.8, beta=0.4, k=k)
# python
dot2 = py_dot(m, k)
cosine2 = py_cosine(m, k).tocsr()
asy_cosine2 = py_asy_cosine(m, 0.2, k=k)
jaccard2 = py_jaccard(m, k)
dice2 = py_dice(m, k)
tversky2 = py_tversky(m, alpha=0.8, beta=0.4, k=k)
p3alpha2 = py_p3alpha(m, alpha=0.8, k=k)
rp3beta2 = py_rp3beta(m, alpha=0.8, beta=0.4, k=k)
# test
np.testing.assert_allclose(check_sum(dot),
check_sum(dot2),
rtol=rtol,
err_msg='dot error')
np.testing.assert_allclose(check_sum(cosine),
check_sum(cosine2),
rtol=rtol,
err_msg='cosine error')
np.testing.assert_allclose(check_sum(asy_cosine),
check_sum(asy_cosine2),
rtol=rtol,
err_msg='asy_cosine error')
np.testing.assert_allclose(check_sum(jaccard),
check_sum(jaccard2),
rtol=rtol,
err_msg='jaccard error')
np.testing.assert_allclose(check_sum(dice),
check_sum(dice2),
rtol=rtol,
err_msg='dice error')
np.testing.assert_allclose(check_sum(tversky),
check_sum(tversky2),
rtol=rtol,
err_msg='tversky error')
np.testing.assert_allclose(check_sum(p3alpha),
check_sum(p3alpha2),
rtol=rtol,
err_msg='p3alpha error')
np.testing.assert_allclose(check_sum(rp3beta),
check_sum(rp3beta2),
rtol=rtol,
err_msg='rp3beta error')
# test full rows
if full:
np.testing.assert_(check_full(dot, dot2, rtol) == 0, msg='dot error')
np.testing.assert_(check_full(cosine, cosine2, rtol) == 0,
msg='cosine error')
np.testing.assert_(check_full(asy_cosine, asy_cosine2, rtol) == 0,
msg='asy_cosine error')
np.testing.assert_(check_full(jaccard, jaccard2, rtol) == 0,
msg='jaccard error')
np.testing.assert_(check_full(dice, dice2, rtol) == 0,
msg='dice error')
np.testing.assert_(check_full(tversky, tversky2, rtol) == 0,
msg='tversky error')
np.testing.assert_(check_full(p3alpha, p3alpha2, rtol) == 0,
msg='p3alpha error')
np.testing.assert_(check_full(rp3beta, rp3beta2, rtol) == 0,
msg='rp3beta error')
return
def fit(self,
matrix,
k,
distance,
shrink=0,
threshold=0,
implicit=True,
alpha=None,
beta=None,
l=None,
c=None,
verbose=False):
"""
Initialize the model and compute the similarity matrix S with a distance metric.
Access the similarity matrix using: self._sim_matrix
Parameters
----------
matrix : csr_matrix
A sparse matrix. For example, it can be the URM of shape (number_users, number_items).
k : int
K nearest neighbour to consider.
distance : str
One of the supported distance metrics, check collaborative_filtering_base constants.
shrink : float, optional
Shrink term used in the normalization
threshold: float, optional
All the values under this value are cutted from the final result
implicit: bool, optional
If true, treat the URM as implicit, otherwise consider explicit ratings (real values) in the URM
alpha: float, optional, included in [0,1]
beta: float, optional, included in [0,1]
l: float, optional, balance coefficient used in s_plus distance, included in [0,1]
c: float, optional, cosine coefficient, included in [0,1]
"""
alpha = -1 if alpha is None else alpha
beta = -1 if beta is None else beta
l = -1 if l is None else l
c = -1 if c is None else c
if distance == self.SIM_ASYMCOSINE and not (0 <= alpha <= 1):
log.error(
'Invalid parameter alpha in asymmetric cosine similarity!')
return
if distance == self.SIM_TVERSKY and not (0 <= alpha <= 1
and 0 <= beta <= 1):
log.error('Invalid parameter alpha/beta in tversky similarity!')
return
if distance == self.SIM_P3ALPHA and alpha is None:
log.error('Invalid parameter alpha in p3alpha similarity')
return
if distance == self.SIM_RP3BETA and alpha is None and beta is None:
log.error('Invalid parameter alpha/beta in rp3beta similarity')
return
if distance == self.SIM_SPLUS and not (0 <= l <= 1 and 0 <= c <= 1
and 0 <= alpha <= 1
and 0 <= beta <= 1):
log.error('Invalid parameter alpha/beta/l/c in s_plus similarity')
return
# compute and stores the similarity matrix using one of the distance metric: S = R•R'
if distance == self.SIM_COSINE:
self._sim_matrix = sim.cosine(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
elif distance == self.SIM_ASYMCOSINE:
self._sim_matrix = sim.asymmetric_cosine(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha)
elif distance == self.SIM_JACCARD:
self._sim_matrix = sim.jaccard(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
elif distance == self.SIM_DICE:
self._sim_matrix = sim.dice(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
elif distance == self.SIM_TVERSKY:
self._sim_matrix = sim.tversky(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha,
beta=beta)
elif distance == self.SIM_P3ALPHA:
self._sim_matrix = sim.p3alpha(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha)
elif distance == self.SIM_RP3BETA:
self._sim_matrix = sim.rp3beta(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha,
beta=beta)
elif distance == self.SIM_SPLUS:
self._sim_matrix = sim.s_plus(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
l=l,
t1=alpha,
t2=beta,
c=c)
else:
log.error('Invalid distance metric: {}'.format(distance))
#self.SIM_DOTPRODUCT: sim.dot_product(matrix, k=k, shrink=shrink, threshold=threshold, binary=implicit)
return self._sim_matrix
#setup parser
parser = argparse.ArgumentParser()
parser.add_argument("-s",
"--split",
help="The dataset split to use",
choices=['original', 'validation'],
type=str,
required=True)
args = parser.parse_args()
# first load the data
df_train = pd.read_csv(f"../dataset/{args.split}/train.csv", escapechar="\\")
df_test = pd.read_csv(f"../dataset/{args.split}/test.csv", escapechar="\\")
# ALWAYS sort the data by record_id
df_train = df_train.sort_values(by=['record_id']).reset_index(drop=True)
df_test = df_test.sort_values(by=['record_id']).reset_index(drop=True)
df_train.address = df_train.address.astype(str)
df_test.address = df_test.address.astype(str)
# mi serve una colonna con tutti i telefoni su cui fare tfidf
all_adds = list(df_train.address) + list(df_test.address)
# daje con tfidf
vectorizer = TfidfVectorizer(min_df=1, analyzer=ngrams)
tf_idf_matrix = vectorizer.fit_transform(all_adds)
# split
tf_idf_train = tf_idf_matrix[:df_train.
shape[0], :] # 691440 è la lunghezza del train
tf_idf_test = tf_idf_matrix[df_train.shape[0]:, :]
jac = sim.jaccard(tf_idf_test, tf_idf_train.T, k=300)
save_npz(f'jaccard_tfidf_address_{args.split}.npz', jac.tocsr())
def similarity(matrix,
k=100,
sim_type='cosine',
shrink=0,
threshold=0,
implicit=True,
alpha=None,
beta=None,
l=None,
c=None):
# similarity type
SIM_COSINE = 'cosine'
SIM_TVERSKY = 'tversky'
SIM_P3ALPHA = 'p3alpha'
SIM_ASYMCOSINE = 'asymcosine'
SIM_RP3BETA = 'rp3beta'
SIM_SPLUS = 'splus'
SIM_JACCARD = 'jaccard'
SIM_DICE = 'dice'
matrix = matrix.T
if sim_type == SIM_COSINE:
return sim.cosine(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
elif sim_type == SIM_ASYMCOSINE:
return sim.asymmetric_cosine(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha)
elif sim_type == SIM_JACCARD:
return sim.jaccard(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
elif sim_type == SIM_TVERSKY:
return sim.tversky(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha,
beta=beta)
elif sim_type == SIM_P3ALPHA:
return sim.p3alpha(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha)
elif sim_type == SIM_RP3BETA:
return sim.rp3beta(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit,
alpha=alpha,
beta=beta)
elif sim_type == SIM_SPLUS:
return sim.s_plus(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit) #, l=l, t1=alpha, t2=beta, c=c)
elif sim_type == SIM_DICE:
return sim.dice(matrix,
k=k,
shrink=shrink,
threshold=threshold,
binary=implicit)
else:
print('Error wrong distance metric')
def fit(self):
self.alpha = -1 if self.alpha is None else self.alpha
self.beta = -1 if self.beta is None else self.beta
self.l = -1 if self.l is None else self.l
self.c = -1 if self.c is None else self.c
if self.distance == self.SIM_ASYMCOSINE and not (0 <= self.alpha <= 1):
log.error(
'Invalid parameter alpha in asymmetric cosine Similarity_MFD!')
return
if self.distance == self.SIM_TVERSKY and not (0 <= self.alpha <= 1
and 0 <= self.beta <= 1):
log.error(
'Invalid parameter alpha/beta in tversky Similarity_MFD!')
return
if self.distance == self.SIM_P3ALPHA and self.alpha is None:
log.error('Invalid parameter alpha in p3alpha Similarity_MFD')
return
if self.distance == self.SIM_RP3BETA and self.alpha is None and self.beta is None:
log.error('Invalid parameter alpha/beta in rp3beta Similarity_MFD')
return
if self.distance == self.SIM_SPLUS and not (
0 <= self.l <= 1 and 0 <= self.c <= 1 and 0 <= self.alpha <= 1
and 0 <= self.beta <= 1):
log.error(
'Invalid parameter alpha/beta/l/c in s_plus Similarity_MFD')
return
# compute and stores the Similarity_MFD matrix using one of the distance metric: S = R•R'
if self.distance == self.SIM_COSINE:
self._sim_matrix = sim.cosine(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit)
elif self.distance == self.SIM_ASYMCOSINE:
self._sim_matrix = sim.asymmetric_cosine(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit,
alpha=self.alpha)
elif self.distance == self.SIM_JACCARD:
self._sim_matrix = sim.jaccard(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit)
elif self.distance == self.SIM_DICE:
self._sim_matrix = sim.dice(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit)
elif self.distance == self.SIM_TVERSKY:
self._sim_matrix = sim.tversky(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit,
alpha=self.alpha,
beta=self.beta)
elif self.distance == self.SIM_P3ALPHA:
self._sim_matrix = sim.p3alpha(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit,
alpha=self.alpha)
elif self.distance == self.SIM_RP3BETA:
self._sim_matrix = sim.rp3beta(self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit,
alpha=self.alpha,
beta=self.beta)
elif self.distance == self.SIM_SPLUS:
self._sim_matrix = prep.normalize(sim.s_plus(
self.matrix,
k=self.k,
shrink=self.shrink,
threshold=self.threshold,
binary=self.implicit,
l=self.l,
t1=self.alpha,
t2=self.beta,
c=self.c),
norm='l2',
axis=0)
else:
log.error('Invalid distance metric: {}'.format(self.distance))
return self._sim_matrix
string = re.sub(r'[,-./]',r'', string)
ngrams = zip(*[string[i:] for i in range(n)])
return [''.join(ngram) for ngram in ngrams]
#setup parser
parser = argparse.ArgumentParser()
parser.add_argument("-s","--split",
help="The dataset split to use",
choices=['original','validation'],
type=str,
required=True)
args = parser.parse_args()
# first load the data
df_train = pd.read_csv(f"../dataset/{args.split}/train.csv", escapechar="\\")
df_test = pd.read_csv(f"../dataset/{args.split}/test.csv", escapechar="\\")
# ALWAYS sort the data by record_id
df_train = df_train.sort_values(by=['record_id']).reset_index(drop=True)
df_test = df_test.sort_values(by=['record_id']).reset_index(drop=True)
df_train.name = df_train.name.astype(str)
df_test.name = df_test.name.astype(str)
# mi serve una colonna con tutti i nomi su cui fare tfidf
all_names = list(df_train.name) + list(df_test.name)
# daje con tfidf
vectorizer = TfidfVectorizer(min_df=1, analyzer=ngrams)
tf_idf_matrix = vectorizer.fit_transform(all_names)
# split
tf_idf_train = tf_idf_matrix[:df_train.shape[0],:] # 691440 è la lunghezza del train
tf_idf_test = tf_idf_matrix[df_train.shape[0]:,:]
cos_tfidf = sim.jaccard(tf_idf_test, tf_idf_train.T, k=300)
save_npz(f'jaccard_tfidf_name_{args.split}_noclean.npz', cos_tfidf.tocsr())
