def set_URM_train(self, URM_train_new, estimate_model_for_cold_users = False, topK = 100, **kwargs):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(self.RECOMMENDER_NAME)
if len(kwargs)>0:
print("{}: set_URM_train keyword arguments not supported for this recommender class. Received: {}".format(self.RECOMMENDER_NAME, kwargs))
self.URM_train = check_matrix(URM_train_new.copy(), 'csr', dtype=np.float32)
self.URM_train.eliminate_zeros()
if estimate_model_for_cold_users == "itemKNN":
print("{}: Estimating ItemKNN model from ITEM latent factors...".format(self.RECOMMENDER_NAME))
W_sparse = compute_W_sparse_from_item_latent_factors(self.ITEM_factors, topK=topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(self.URM_train)
self._ItemKNNRecommender.fit(W_sparse, topK=topK)
self._cold_user_KNN_model_available = True
self._warm_user_KNN_mask = np.ediff1d(self.URM_train.indptr) > 0
print("{}: Estimating ItemKNN model from ITEM latent factors... done!".format(self.RECOMMENDER_NAME))
elif estimate_model_for_cold_users == "mean_item_factors":
print("{}: Estimating USER latent factors from ITEM latent factors...".format(self.RECOMMENDER_NAME))
self._cold_user_mask = np.ediff1d(self.URM_train.indptr) == 0
profile_length = np.ediff1d(self.URM_train.indptr)
profile_length_sqrt = np.sqrt(profile_length)
self.USER_factors = self.URM_train.dot(self.ITEM_factors)
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if profile_length_sqrt[user_index] > 0:
self.USER_factors[user_index, :] /= profile_length_sqrt[user_index]
print("{}: Estimating USER latent factors from ITEM latent factors... done!".format(self.RECOMMENDER_NAME))
def load_model(self, folder_path, file_name=None):
super(BaseMatrixFactorizationRecommender,
self).load_model(folder_path, file_name=file_name)
if self._cold_user_KNN_model_flag:
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.fit(self._ItemKNNRecommender_W_sparse,
topK=self._ItemKNNRecommender_topK)
del self._ItemKNNRecommender_W_sparse
del self._ItemKNNRecommender_topK
def load_model(self, folder_path, file_name=None):
super(BaseMatrixFactorizationRecommender,
self).load_model(folder_path, file_name=file_name)
self.mf_recommender.load_model(folder_path,
file_name=file_name + "_warm_users")
if self._cold_user_KNN_model_flag:
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.load_model(folder_path,
file_name=file_name +
"_cold_users")
def computeMaxTheoreticalPerformance(self):
# Max performance would be if we were able to learn the content matrix having for each non-zero cell exactly
# the value that appears in the collaborative similarity
print(self.RECOMMENDER_NAME + ": Computing collaborative performance")
recommender = ItemKNNCustomSimilarityRecommender()
recommender.fit(self.S_matrix_target, self.URM_train)
results_run = self.evaluator_object(recommender)
self.writeLog(self.RECOMMENDER_NAME +
": Collaborative performance is: {}".format(results_run))
print(self.RECOMMENDER_NAME + ": Computing top structural performance")
n_items = self.ICM.shape[0]
S_optimal = sps.csr_matrix(
(self.data_list, (self.row_list, self.col_list)),
shape=(n_items, n_items))
S_optimal.eliminate_zeros()
recommender = ItemKNNCustomSimilarityRecommender()
recommender.fit(S_optimal, self.URM_train)
results_run = self.evaluator_object(recommender)
self.writeLog(
self.RECOMMENDER_NAME +
": Top structural performance is: {}".format(results_run))
class BaseMatrixFactorizationRecommender(BaseRecommender):
"""
This class refers to a BaseRecommender KNN which uses matrix factorization,
it provides functions to compute item's score as well as a function to save the W_matrix
The prediction for cold users will always be -inf for ALL items
"""
def __init__(self, URM_train):
super(BaseMatrixFactorizationRecommender, self).__init__(URM_train)
self._cold_user_mask = np.ediff1d(self.URM_train.indptr) == 0
self._cold_user_KNN_model_available = False
self._warm_user_KNN_mask = np.zeros(len(self._cold_user_mask),
dtype=np.bool)
if self._cold_user_mask.any():
print("{}: Detected {} ({:.2f} %) cold users.".format(
self.RECOMMENDER_NAME, self._cold_user_mask.sum(),
self._cold_user_mask.sum() / len(self._cold_user_mask) * 100))
def _get_cold_user_mask(self):
return self._cold_user_mask
def set_URM_train(self,
URM_train_new,
estimate_model_for_cold_users=False,
topK=100,
**kwargs):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(
self.RECOMMENDER_NAME)
if len(kwargs) > 0:
print(
"{}: set_URM_train keyword arguments not supported for this recommender class. Received: {}"
.format(self.RECOMMENDER_NAME, kwargs))
self.URM_train = check_matrix(URM_train_new.copy(),
'csr',
dtype=np.float32)
self.URM_train.eliminate_zeros()
if estimate_model_for_cold_users == "itemKNN":
print("{}: Estimating ItemKNN model from ITEM latent factors...".
format(self.RECOMMENDER_NAME))
W_sparse = compute_W_sparse_from_item_latent_factors(
self.ITEM_factors, topK=topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.fit(W_sparse, topK=topK)
self._cold_user_KNN_model_available = True
self._warm_user_KNN_mask = np.ediff1d(self.URM_train.indptr) > 0
print(
"{}: Estimating ItemKNN model from ITEM latent factors... done!"
.format(self.RECOMMENDER_NAME))
elif estimate_model_for_cold_users == "mean_item_factors":
print(
"{}: Estimating USER latent factors from ITEM latent factors..."
.format(self.RECOMMENDER_NAME))
self._cold_user_mask = np.ediff1d(self.URM_train.indptr) == 0
profile_length = np.ediff1d(self.URM_train.indptr)
profile_length_sqrt = np.sqrt(profile_length)
self.USER_factors = self.URM_train.dot(self.ITEM_factors)
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if profile_length_sqrt[user_index] > 0:
self.USER_factors[
user_index, :] /= profile_length_sqrt[user_index]
print(
"{}: Estimating USER latent factors from ITEM latent factors... done!"
.format(self.RECOMMENDER_NAME))
def _compute_item_score(self, user_id_array, items_to_compute=None):
"""
USER_factors is n_users x n_factors
ITEM_factors is n_items x n_factors
The prediction for cold users will always be -inf for ALL items
:param user_id_array:
:param items_to_compute:
:return:
"""
assert self.USER_factors.shape[1] == self.ITEM_factors.shape[1], \
"{}: User and Item factors have inconsistent shape".format(self.RECOMMENDER_NAME)
assert self.USER_factors.shape[0] > user_id_array.max(),\
"{}: Cold users not allowed. Users in trained model are {}, requested prediction for users up to {}".format(
self.RECOMMENDER_NAME, self.USER_factors.shape[0], user_id_array.max())
if items_to_compute is not None:
item_scores = -np.ones(
(len(user_id_array), self.ITEM_factors.shape[0]),
dtype=np.float32) * np.inf
item_scores[:, items_to_compute] = np.dot(
self.USER_factors[user_id_array],
self.ITEM_factors[items_to_compute, :].T)
else:
item_scores = np.dot(self.USER_factors[user_id_array],
self.ITEM_factors.T)
cold_users_MF_mask = self._get_cold_user_mask()[user_id_array]
if cold_users_MF_mask.any():
if self._cold_user_KNN_model_available:
# Add KNN scores for users cold for MF but warm in KNN model
cold_users_in_MF_warm_in_KNN_mask = np.logical_and(
cold_users_MF_mask,
self._warm_user_KNN_mask[user_id_array])
item_scores[
cold_users_in_MF_warm_in_KNN_mask, :] = self._ItemKNNRecommender._compute_item_score(
user_id_array[cold_users_in_MF_warm_in_KNN_mask],
items_to_compute=items_to_compute)
# Set cold users as those neither in MF nor in KNN
cold_users_MF_mask = np.logical_and(
cold_users_MF_mask,
np.logical_not(cold_users_in_MF_warm_in_KNN_mask))
# Set as -inf all remaining cold user scores
item_scores[cold_users_MF_mask, :] = -np.ones_like(
item_scores[cold_users_MF_mask, :]) * np.inf
return item_scores
def saveModel(self, folder_path, file_name=None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
print("{}: Saving model in file '{}'".format(self.RECOMMENDER_NAME,
folder_path + file_name))
dictionary_to_save = {
"USER_factors": self.USER_factors,
"ITEM_factors": self.ITEM_factors,
"_cold_user_mask": self._cold_user_mask
}
pickle.dump(dictionary_to_save,
open(folder_path + file_name, "wb"),
protocol=pickle.HIGHEST_PROTOCOL)
print("{}: Saving complete".format(self.RECOMMENDER_NAME,
folder_path + file_name))
class BaseMatrixFactorizationRecommender(BaseRecommender):
"""
This class refers to a BaseRecommender KNN which uses matrix factorization,
it provides functions to compute item's score as well as a function to save the W_matrix
The prediction for cold users will always be -inf for ALL items
"""
def __init__(self, URM_train):
super(BaseMatrixFactorizationRecommender, self).__init__(URM_train)
self.use_bias = False
self._cold_user_KNN_model_available = False
self._warm_user_KNN_mask = np.zeros(len(self._get_cold_user_mask()),
dtype=np.bool)
def set_URM_train(self,
URM_train_new,
estimate_model_for_cold_users=False,
topK=100,
**kwargs):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(
self.RECOMMENDER_NAME)
if len(kwargs) > 0:
print(
"{}: set_URM_train keyword arguments not supported for this recommender class. Received: {}"
.format(self.RECOMMENDER_NAME, kwargs))
self.URM_train = check_matrix(URM_train_new.copy(),
'csr',
dtype=np.float32)
self.URM_train.eliminate_zeros()
if estimate_model_for_cold_users == "itemKNN":
print("{}: Estimating ItemKNN model from ITEM latent factors...".
format(self.RECOMMENDER_NAME))
W_sparse = compute_W_sparse_from_item_latent_factors(
self.ITEM_factors, topK=topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.fit(W_sparse, topK=topK)
self._cold_user_KNN_model_available = True
self._warm_user_KNN_mask = np.ediff1d(self.URM_train.indptr) > 0
print(
"{}: Estimating ItemKNN model from ITEM latent factors... done!"
.format(self.RECOMMENDER_NAME))
elif estimate_model_for_cold_users == "mean_item_factors":
print(
"{}: Estimating USER latent factors from ITEM latent factors..."
.format(self.RECOMMENDER_NAME))
self._cold_user_mask = np.ediff1d(self.URM_train.indptr) == 0
profile_length = np.ediff1d(self.URM_train.indptr)
profile_length_sqrt = np.sqrt(profile_length)
self.USER_factors = self.URM_train.dot(self.ITEM_factors)
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if profile_length_sqrt[user_index] > 0:
self.USER_factors[
user_index, :] /= profile_length_sqrt[user_index]
print(
"{}: Estimating USER latent factors from ITEM latent factors... done!"
.format(self.RECOMMENDER_NAME))
#########################################################################################################
########## ##########
########## COMPUTE ITEM SCORES ##########
########## ##########
#########################################################################################################
def _compute_item_score(self, user_id_array, items_to_compute=None):
"""
USER_factors is n_users x n_factors
ITEM_factors is n_items x n_factors
The prediction for cold users will always be -inf for ALL items
:param user_id_array:
:param items_to_compute:
:return:
"""
assert self.USER_factors.shape[1] == self.ITEM_factors.shape[1], \
"{}: User and Item factors have inconsistent shape".format(self.RECOMMENDER_NAME)
assert self.USER_factors.shape[0] > user_id_array.max(),\
"{}: Cold users not allowed. Users in trained model are {}, requested prediction for users up to {}".format(
self.RECOMMENDER_NAME, self.USER_factors.shape[0], user_id_array.max())
if items_to_compute is not None:
item_scores = -np.ones(
(len(user_id_array), self.ITEM_factors.shape[0]),
dtype=np.float32) * np.inf
item_scores[:, items_to_compute] = np.dot(
self.USER_factors[user_id_array],
self.ITEM_factors[items_to_compute, :].T)
else:
item_scores = np.dot(self.USER_factors[user_id_array],
self.ITEM_factors.T)
# No need to select only the specific negative items or warm users because the -inf score will not change
if self.use_bias:
item_scores += self.ITEM_bias + self.GLOBAL_bias
item_scores = (item_scores.T + self.USER_bias[user_id_array]).T
item_scores = self._compute_item_score_postprocess_for_cold_users(
user_id_array, item_scores)
item_scores = self._compute_item_score_postprocess_for_cold_items(
item_scores)
return item_scores
def _compute_item_score_postprocess_for_cold_users(self, user_id_array,
item_scores):
"""
Remove cold users from the computed item scores, setting them to -inf
Or estimate user factors with specified method
:param user_id_array:
:param item_scores:
:return:
"""
cold_users_batch_mask = self._get_cold_user_mask()[user_id_array]
# Set as -inf all cold user scores
if cold_users_batch_mask.any():
if self._cold_user_KNN_model_available:
# Add KNN scores for users cold for MF but warm in KNN model
cold_users_in_MF_warm_in_KNN_mask = np.logical_and(
cold_users_batch_mask,
self._warm_user_KNN_mask[user_id_array])
item_scores[
cold_users_in_MF_warm_in_KNN_mask, :] = self._ItemKNNRecommender._compute_item_score(
user_id_array[cold_users_in_MF_warm_in_KNN_mask],
items_to_compute=items_to_compute)
# Set cold users as those neither in MF nor in KNN
cold_users_batch_mask = np.logical_and(
cold_users_batch_mask,
np.logical_not(cold_users_in_MF_warm_in_KNN_mask))
# Set as -inf all remaining cold user scores
item_scores[cold_users_batch_mask, :] = -np.ones_like(
item_scores[cold_users_batch_mask, :]) * np.inf
return item_scores
#########################################################################################################
########## ##########
########## LOAD AND SAVE ##########
########## ##########
#########################################################################################################
def saveModel(self, folder_path, file_name=None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
print("{}: Saving model in file '{}'".format(self.RECOMMENDER_NAME,
folder_path + file_name))
data_dict_to_save = {
"USER_factors": self.USER_factors,
"ITEM_factors": self.ITEM_factors,
"use_bias": self.use_bias,
"_cold_user_mask": self._cold_user_mask
}
if self.use_bias:
data_dict_to_save["ITEM_bias"] = self.ITEM_bias
data_dict_to_save["USER_bias"] = self.USER_bias
data_dict_to_save["GLOBAL_bias"] = self.GLOBAL_bias
dataIO = DataIO(folder_path=folder_path)
dataIO.save_data(file_name=file_name,
data_dict_to_save=data_dict_to_save)
print("{}: Saving complete".format(self.RECOMMENDER_NAME,
folder_path + file_name))
class MF_cold_user_wrapper(BaseMatrixFactorizationRecommender):
""" MF_cold_user_wrapper"""
RECOMMENDER_NAME = "MF_cold_user_wrapper"
def __init__(self,
MF_recommender_class: BaseMatrixFactorizationRecommender,
*posargs, **kwargs):
"""
Creates an instance of the MF algorithm with the given hyperparameters and data
:param MF_recommender_class:
:param posargs:
:param kwargs:
"""
super(MF_cold_user_wrapper, self).__init__(*posargs, **kwargs)
self.mf_recommender = MF_recommender_class(*posargs, **kwargs)
self.RECOMMENDER_NAME = self.mf_recommender.RECOMMENDER_NAME # + "_cold_user_Wrapper"
self.estimate_model_for_cold_users = False
self._cold_user_KNN_model_flag = False
self._cold_user_KNN_estimated_factors_flag = False
self._warm_user_KNN_mask = np.zeros(len(self._get_cold_user_mask()),
dtype=np.bool)
def _compute_item_score(self, user_id_array, items_to_compute=None):
"""
Compute the items scores using the native function for the MF algorithm
:param posargs:
:param kwargs:
:return:
"""
# item_scores = self.mf_recommender._compute_item_score(user_id_array, items_to_compute = items_to_compute)
item_scores = super(MF_cold_user_wrapper, self)._compute_item_score(
user_id_array, items_to_compute=items_to_compute)
item_scores = self._compute_item_score_for_cold_users(
user_id_array, item_scores, items_to_compute=items_to_compute)
return item_scores
def _compute_item_score_for_cold_users(self,
user_id_array,
item_scores,
items_to_compute=None):
"""
Compute item scores with the ItemKNN model
:param user_id_array:
:param item_scores:
:return:
"""
cold_users_batch_mask = self._get_cold_user_mask()[user_id_array]
if cold_users_batch_mask.any(
) and not self._cold_user_KNN_estimated_factors_flag:
if self._cold_user_KNN_model_flag:
# Add KNN scores for users cold for MF but warm in KNN model
cold_users_in_MF_warm_in_KNN_mask = np.logical_and(
cold_users_batch_mask,
self._warm_user_KNN_mask[user_id_array])
item_scores[
cold_users_in_MF_warm_in_KNN_mask, :] = self._ItemKNNRecommender._compute_item_score(
user_id_array[cold_users_in_MF_warm_in_KNN_mask],
items_to_compute=items_to_compute)
return item_scores
def set_URM_train(self, URM_train_new):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(
self.RECOMMENDER_NAME)
URM_train_new = check_matrix(URM_train_new, 'csr', dtype=np.float32)
profile_length_new = np.ediff1d(URM_train_new.indptr)
if self.estimate_model_for_cold_users == "itemKNN":
self._print("Generating ItemKNN model from ITEM latent factors...")
W_sparse = compute_W_sparse_from_item_latent_factors(
self.ITEM_factors,
topK=self.estimate_model_for_cold_users_topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
URM_train_new)
self._ItemKNNRecommender.fit(W_sparse, topK=None)
self._cold_user_KNN_model_flag = True
self._cold_user_KNN_model_flag = True
self._warm_user_KNN_mask = profile_length_new > 0
self._print(
"Generating ItemKNN model from ITEM latent factors... done!")
elif self.estimate_model_for_cold_users == "mean_item_factors":
self._print(
"Estimating USER latent factors from ITEM latent factors...")
cold_user_mask_previous = self._get_cold_user_mask()
profile_length_sqrt = np.sqrt(profile_length_new)
self.USER_factors[cold_user_mask_previous, :] = URM_train_new.dot(
self.ITEM_factors)[cold_user_mask_previous, :]
self._cold_user_KNN_estimated_factors_flag = True
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if cold_user_mask_previous[
user_index] and profile_length_sqrt[user_index] > 0:
self.USER_factors[
user_index, :] /= profile_length_sqrt[user_index]
self._print(
"Estimating USER latent factors from ITEM latent factors... done!"
)
self.URM_train = sps.csr_matrix(URM_train_new.copy())
self.URM_train.eliminate_zeros()
def fit(self,
*posargs,
estimate_model_for_cold_users=None,
estimate_model_for_cold_users_topK=100,
**kwargs):
"""
Fits the MF model with the given hyper-parameters and sets the kwarg "estimate_model_for_cold_users"
:param posargs:
:param kwargs:
:return:
"""
self.estimate_model_for_cold_users = estimate_model_for_cold_users
self.estimate_model_for_cold_users_topK = estimate_model_for_cold_users_topK
self.mf_recommender.fit(*posargs, **kwargs)
self.USER_factors = self.mf_recommender.USER_factors.copy()
self.ITEM_factors = self.mf_recommender.ITEM_factors.copy()
def save_model(self, folder_path, file_name=None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
self._print("Saving model in file '{}'".format(folder_path +
file_name))
self.mf_recommender.save_model(folder_path,
file_name=file_name + "_warm_users")
data_dict_to_save = {
"_cold_user_KNN_model_flag":
self._cold_user_KNN_model_flag,
"_cold_user_KNN_estimated_factors_flag":
self._cold_user_KNN_estimated_factors_flag
}
if self._cold_user_KNN_model_flag:
self._ItemKNNRecommender.save_model(folder_path,
file_name=file_name +
"_cold_users")
dataIO = DataIO(folder_path=folder_path)
dataIO.save_data(file_name=file_name,
data_dict_to_save=data_dict_to_save)
self._print("Saving complete")
def load_model(self, folder_path, file_name=None):
super(BaseMatrixFactorizationRecommender,
self).load_model(folder_path, file_name=file_name)
self.mf_recommender.load_model(folder_path,
file_name=file_name + "_warm_users")
if self._cold_user_KNN_model_flag:
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.load_model(folder_path,
file_name=file_name +
"_cold_users")
def set_URM_train(self, URM_train_new):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(
self.RECOMMENDER_NAME)
URM_train_new = check_matrix(URM_train_new, 'csr', dtype=np.float32)
profile_length_new = np.ediff1d(URM_train_new.indptr)
if self.estimate_model_for_cold_users == "itemKNN":
self._print("Generating ItemKNN model from ITEM latent factors...")
W_sparse = compute_W_sparse_from_item_latent_factors(
self.ITEM_factors,
topK=self.estimate_model_for_cold_users_topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
URM_train_new)
self._ItemKNNRecommender.fit(W_sparse, topK=None)
self._cold_user_KNN_model_flag = True
self._cold_user_KNN_model_flag = True
self._warm_user_KNN_mask = profile_length_new > 0
self._print(
"Generating ItemKNN model from ITEM latent factors... done!")
elif self.estimate_model_for_cold_users == "mean_item_factors":
self._print(
"Estimating USER latent factors from ITEM latent factors...")
cold_user_mask_previous = self._get_cold_user_mask()
profile_length_sqrt = np.sqrt(profile_length_new)
self.USER_factors[cold_user_mask_previous, :] = URM_train_new.dot(
self.ITEM_factors)[cold_user_mask_previous, :]
self._cold_user_KNN_estimated_factors_flag = True
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if cold_user_mask_previous[
user_index] and profile_length_sqrt[user_index] > 0:
self.USER_factors[
user_index, :] /= profile_length_sqrt[user_index]
self._print(
"Estimating USER latent factors from ITEM latent factors... done!"
)
self.URM_train = sps.csr_matrix(URM_train_new.copy())
self.URM_train.eliminate_zeros()
class BaseMatrixFactorizationRecommender(BaseRecommender):
"""
This class refers to a BaseRecommender KNN which uses matrix factorization,
it provides functions to compute item's score as well as a function to save the W_matrix
The prediction for cold users will always be -inf for ALL items
"""
def __init__(self,
URM_train,
verbose=True,
evaluation_block_size: int = 1):
super(BaseMatrixFactorizationRecommender,
self).__init__(URM_train,
verbose=verbose,
evaluation_block_size=evaluation_block_size)
self.use_bias = False
self._cold_user_KNN_model_flag = False
self._cold_user_KNN_estimated_factors_flag = False
self._warm_user_KNN_mask = np.zeros(len(self._get_cold_user_mask()),
dtype=np.bool)
def set_URM_train(self,
URM_train_new,
estimate_model_for_cold_users=False,
topK=100,
**kwargs):
"""
:param URM_train_new:
:param estimate_item_similarity_for_cold_users: Set to TRUE if you want to estimate the item-item similarity for cold users to be used as in a KNN algorithm
:param topK: 100
:param kwargs:
:return:
"""
assert self.URM_train.shape == URM_train_new.shape, "{}: set_URM_train old and new URM train have different shapes".format(
self.RECOMMENDER_NAME)
if len(kwargs) > 0:
self._print(
"set_URM_train keyword arguments not supported for this recommender class. Received: {}"
.format(kwargs))
URM_train_new = check_matrix(URM_train_new, 'csr', dtype=np.float32)
profile_length_new = np.ediff1d(URM_train_new.indptr)
if estimate_model_for_cold_users == "itemKNN":
self._print("Estimating ItemKNN model from ITEM latent factors...")
W_sparse = compute_W_sparse_from_item_latent_factors(
self.ITEM_factors, topK=topK)
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
URM_train_new)
self._ItemKNNRecommender.fit(W_sparse, topK=topK)
self._ItemKNNRecommender_topK = topK
self._cold_user_KNN_model_flag = True
self._warm_user_KNN_mask = profile_length_new > 0
self._print(
"Estimating ItemKNN model from ITEM latent factors... done!")
elif estimate_model_for_cold_users == "mean_item_factors":
self._print(
"Estimating USER latent factors from ITEM latent factors...")
cold_user_mask_previous = self._get_cold_user_mask()
profile_length_sqrt = np.sqrt(profile_length_new)
self.USER_factors[cold_user_mask_previous, :] = URM_train_new.dot(
self.ITEM_factors)[cold_user_mask_previous, :]
self._cold_user_KNN_estimated_factors_flag = True
#Divide every row for the sqrt of the profile length
for user_index in range(self.n_users):
if cold_user_mask_previous[
user_index] and profile_length_sqrt[user_index] > 0:
self.USER_factors[
user_index, :] /= profile_length_sqrt[user_index]
self._print(
"Estimating USER latent factors from ITEM latent factors... done!"
)
self.URM_train = check_matrix(URM_train_new.copy(),
'csr',
dtype=np.float32)
self.URM_train.eliminate_zeros()
#########################################################################################################
########## ##########
########## COMPUTE ITEM SCORES ##########
########## ##########
#########################################################################################################
def _compute_item_score(self, user_id_array, items_to_compute=None):
"""
USER_factors is n_users x n_factors
ITEM_factors is n_items x n_factors
The prediction for cold users will always be -inf for ALL items
:param user_id_array:
:param items_to_compute:
:return:
"""
assert self.USER_factors.shape[1] == self.ITEM_factors.shape[1], \
"{}: User and Item factors have inconsistent shape".format(self.RECOMMENDER_NAME)
assert self.USER_factors.shape[0] > np.max(user_id_array),\
"{}: Cold users not allowed. Users in trained model are {}, requested prediction for users up to {}".format(
self.RECOMMENDER_NAME, self.USER_factors.shape[0], np.max(user_id_array))
if items_to_compute is not None:
item_scores = -np.ones(
(len(user_id_array), self.ITEM_factors.shape[0]),
dtype=np.float32) * np.inf
item_scores[:, items_to_compute] = np.dot(
self.USER_factors[user_id_array],
self.ITEM_factors[items_to_compute, :].T)
else:
item_scores = np.dot(self.USER_factors[user_id_array],
self.ITEM_factors.T)
# No need to select only the specific negative items or warm users because the -inf score will not change
if self.use_bias:
item_scores += self.ITEM_bias + self.GLOBAL_bias
item_scores = (item_scores.T + self.USER_bias[user_id_array]).T
return item_scores
#########################################################################################################
########## ##########
########## LOAD AND SAVE ##########
########## ##########
#########################################################################################################
def save_model(self, folder_path, file_name=None):
if file_name is None:
file_name = self.RECOMMENDER_NAME
logger.info(f"Saving model in file '{folder_path + file_name}'")
data_dict_to_save = {
"USER_factors":
self.USER_factors,
"ITEM_factors":
self.ITEM_factors,
"use_bias":
self.use_bias,
"_cold_user_mask":
self._cold_user_mask,
"_cold_user_KNN_model_flag":
self._cold_user_KNN_model_flag,
"_cold_user_KNN_estimated_factors_flag":
self._cold_user_KNN_estimated_factors_flag
}
if self.use_bias:
data_dict_to_save["ITEM_bias"] = self.ITEM_bias
data_dict_to_save["USER_bias"] = self.USER_bias
data_dict_to_save["GLOBAL_bias"] = self.GLOBAL_bias
if self._cold_user_KNN_model_flag:
data_dict_to_save[
"_ItemKNNRecommender_W_sparse"] = self._ItemKNNRecommender.W_sparse
data_dict_to_save[
"_ItemKNNRecommender_topK"] = self._ItemKNNRecommender_topK
dataIO = DataIO(folder_path=folder_path)
dataIO.save_data(file_name=file_name,
data_dict_to_save=data_dict_to_save)
self._print("Saving complete")
def load_model(self, folder_path, file_name=None):
super(BaseMatrixFactorizationRecommender,
self).load_model(folder_path, file_name=file_name)
if self._cold_user_KNN_model_flag:
self._ItemKNNRecommender = ItemKNNCustomSimilarityRecommender(
self.URM_train)
self._ItemKNNRecommender.fit(self._ItemKNNRecommender_W_sparse,
topK=self._ItemKNNRecommender_topK)
del self._ItemKNNRecommender_W_sparse
del self._ItemKNNRecommender_topK
