def __init__(self,
d,
gamma,
C,
batch_size=10,
positive_thresh=0.00001,
max_iters=1000,
validation_iters=100,
max_trials=50,
sample_item_rate=1.0,
model_type='WARP'):
self.d = d
self.gamma = gamma
self.C = C
self.batch_size = batch_size
self.positive_thresh = positive_thresh
self.max_iters = max_iters
self.validation_iters = validation_iters
self.max_trials = max_trials
self.sample_item_rate = sample_item_rate
self.model_type = model_type
self.model = None
if self.model_type == 'WARP':
self.model = WARP(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters, self.batch_size,
self.positive_thresh, self.max_trials)
elif self.model_type == 'WARPLimitedItem':
self.model = WARPLimitedItem(self.d, self.gamma, self.C,
self.max_iters, self.validation_iters,
self.batch_size, self.positive_thresh,
self.max_trials,
self.sample_item_rate)
else:
print "Invalid model type:[%s]" % self.model_type
sys.exit(2)
def __init__(self,d,gamma,C,batch_size=10,positive_thresh=0.00001,max_iters=1000,
validation_iters = 100, max_trials = 50, sample_item_rate = 1.0,
model_type = 'WARP'):
self.d = d
self.gamma = gamma
self.C = C
self.batch_size = batch_size
self.positive_thresh = positive_thresh
self.max_iters = max_iters
self.validation_iters = validation_iters
self.max_trials = max_trials
self.sample_item_rate = sample_item_rate
self.model_type = model_type
self.model = None
if self.model_type == 'WARP':
self.model = WARP(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters,self.batch_size, self.positive_thresh,
self.max_trials)
elif self.model_type == 'WARPLimitedItem':
self.model = WARPLimitedItem(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters,self.batch_size, self.positive_thresh,
self.max_trials, self.sample_item_rate)
else:
print "Invalid model type:[%s]" % self.model_type
sys.exit(2)
class WARPMFRecommender(MatrixFactorizationRecommender):
"""
Learn matrix factorization optimizing the WARP loss.
Parameters
==========
d : int
Dimensionality of factors.
gamma : float
Learning rate.
C : float
Regularization constant.
batch_size : int
Mini batch size for SGD updates.
positive_thresh: float
Consider an item to be "positive" i.e. liked if its rating is at least this.
max_iters : int
Number of attempts allowed to find a violating negative example during updates.
In practice it means that we optimize for ranks 1 to max_iters-1.
"""
def __init__(self,
d,
gamma,
C,
batch_size=10,
positive_thresh=0.00001,
max_iters=1000,
validation_iters=100,
max_trials=50,
sample_item_rate=1.0,
model_type='WARP'):
self.d = d
self.gamma = gamma
self.C = C
self.batch_size = batch_size
self.positive_thresh = positive_thresh
self.max_iters = max_iters
self.validation_iters = validation_iters
self.max_trials = max_trials
self.sample_item_rate = sample_item_rate
self.model_type = model_type
self.model = None
if self.model_type == 'WARP':
self.model = WARP(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters, self.batch_size,
self.positive_thresh, self.max_trials)
elif self.model_type == 'WARPLimitedItem':
self.model = WARPLimitedItem(self.d, self.gamma, self.C,
self.max_iters, self.validation_iters,
self.batch_size, self.positive_thresh,
self.max_trials,
self.sample_item_rate)
else:
print "Invalid model type:[%s]" % self.model_type
sys.exit(2)
def recommend(self, users, topK):
return self.model.recommend(users, topK)
def fit(self, train, item_features=None):
"""
Learn factors from training set.
Parameters
==========
train : scipy.sparse.csr_matrix
User-item matrix.
item_features : array_like, shape = [num_items, num_features]
Features for each item in the dataset, ignored here.
"""
max_iters, validation_iters, validation = self.create_validation_set(
train)
self.description = 'WARPMF({0})'.format(self.model)
print "Start model fit with param:[%s]" % self.description
self.model.fit(train, validation)
self.U = self.model.U_
self.V = self.model.V_
@staticmethod
def create_validation_set(train):
"""
Hide and return half of the known items for a sample of users,
and estimate the number of sgd iterations to run.
Parameters
==========
train : scipy.sparse.csr_matrix
User-item matrix.
Returns
=======
max_iters : int
Total number of sgd iterations to run.
validation_iters : int
Check progress after this many iterations.
validation : dict
Validation set.
"""
# use 1% of users for validation, with a floor
num_users = train.shape[0]
num_validation_users = max(num_users / 100, 10)
# ensure reasonable expected number of updates per validation user
validation_iters = 100 * num_users / num_validation_users
# and reasonable number of validation cycles
max_iters = 30 * validation_iters
print num_validation_users, 'validation users'
print validation_iters, 'validation iters'
print max_iters, 'max_iters'
validation = dict()
for u in xrange(num_validation_users):
positive = np.where(train[u].data > 0)[0]
hidden = random.sample(positive, positive.shape[0] / 2)
if hidden:
train[u].data[hidden] = 0
#index_value_list = []
#for index in train[u].indices[hidden]:
# index_value_list.append((index, 0))
validation[u] = train[u].indices[hidden]
#validation[u] = index_value_list
return max_iters, validation_iters, validation
class WARPMFRecommender(MatrixFactorizationRecommender):
"""
Learn matrix factorization optimizing the WARP loss.
Parameters
==========
d : int
Dimensionality of factors.
gamma : float
Learning rate.
C : float
Regularization constant.
batch_size : int
Mini batch size for SGD updates.
positive_thresh: float
Consider an item to be "positive" i.e. liked if its rating is at least this.
max_iters : int
Number of attempts allowed to find a violating negative example during updates.
In practice it means that we optimize for ranks 1 to max_iters-1.
"""
def __init__(self,d,gamma,C,batch_size=10,positive_thresh=0.00001,max_iters=1000,
validation_iters = 100, max_trials = 50, sample_item_rate = 1.0,
model_type = 'WARP'):
self.d = d
self.gamma = gamma
self.C = C
self.batch_size = batch_size
self.positive_thresh = positive_thresh
self.max_iters = max_iters
self.validation_iters = validation_iters
self.max_trials = max_trials
self.sample_item_rate = sample_item_rate
self.model_type = model_type
self.model = None
if self.model_type == 'WARP':
self.model = WARP(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters,self.batch_size, self.positive_thresh,
self.max_trials)
elif self.model_type == 'WARPLimitedItem':
self.model = WARPLimitedItem(self.d, self.gamma, self.C, self.max_iters,
self.validation_iters,self.batch_size, self.positive_thresh,
self.max_trials, self.sample_item_rate)
else:
print "Invalid model type:[%s]" % self.model_type
sys.exit(2)
def recommend(self, users, topK):
return self.model.recommend(users, topK)
def fit(self,train,item_features=None):
"""
Learn factors from training set.
Parameters
==========
train : scipy.sparse.csr_matrix
User-item matrix.
item_features : array_like, shape = [num_items, num_features]
Features for each item in the dataset, ignored here.
"""
max_iters,validation_iters,validation = self.create_validation_set(train)
self.description = 'WARPMF({0})'.format(self.model)
print "Start model fit with param:[%s]" % self.description
self.model.fit(train, validation)
self.U = self.model.U_
self.V = self.model.V_
@staticmethod
def create_validation_set(train):
"""
Hide and return half of the known items for a sample of users,
and estimate the number of sgd iterations to run.
Parameters
==========
train : scipy.sparse.csr_matrix
User-item matrix.
Returns
=======
max_iters : int
Total number of sgd iterations to run.
validation_iters : int
Check progress after this many iterations.
validation : dict
Validation set.
"""
# use 1% of users for validation, with a floor
num_users = train.shape[0]
num_validation_users = max(num_users/100,10)
# ensure reasonable expected number of updates per validation user
validation_iters = 100*num_users/num_validation_users
# and reasonable number of validation cycles
max_iters = 30*validation_iters
print num_validation_users,'validation users'
print validation_iters,'validation iters'
print max_iters,'max_iters'
validation = dict()
for u in xrange(num_validation_users):
positive = np.where(train[u].data > 0)[0]
hidden = random.sample(positive,positive.shape[0]/2)
if hidden:
train[u].data[hidden] = 0
#index_value_list = []
#for index in train[u].indices[hidden]:
# index_value_list.append((index, 0))
validation[u] = train[u].indices[hidden]
#validation[u] = index_value_list
return max_iters,validation_iters,validation
