Module reclab.recommenders.sparse
An implementation of SLIM and EASE sparse linear recommenders.
For details, see: - http://glaros.dtc.umn.edu/gkhome/node/774 - https://arxiv.org/pdf/1905.03375.pdf
Expand source code
"""An implementation of SLIM and EASE sparse linear recommenders.
For details, see:
- http://glaros.dtc.umn.edu/gkhome/node/774
- https://arxiv.org/pdf/1905.03375.pdf
"""
import warnings
import numpy as np
import scipy.sparse
import sklearn.linear_model
from sklearn.exceptions import ConvergenceWarning
from . import recommender
warnings.simplefilter('ignore', category=ConvergenceWarning)
class SLIM(recommender.PredictRecommender):
"""The SLIM recommendation model which is a sparse linear method.
Parameters
----------
binarize : boolean
Determines whether to binarize ratings before fitting a model.
alpha : float
Constant that multiplies the regularization terms.
l1_ratio : float
The ratio of the L1 regularization term with respect to the L2 regularization.
max_iter : int
The maximum number of iterations to train the model for.
tol : float
The tolerance below which the optimization will stop.
seed : int
The random seed to use when training the model.
"""
def __init__(self,
binarize=False,
alpha=1.0,
l1_ratio=0.1,
positive=True,
max_iter=100,
tol=1e-4,
seed=0):
"""Create a SLIM recommender."""
super().__init__()
self._binarize = binarize
self._model = sklearn.linear_model.ElasticNet(alpha=alpha,
l1_ratio=l1_ratio,
positive=positive,
fit_intercept=False,
copy_X=False,
precompute=True,
selection='random',
max_iter=max_iter,
tol=tol,
random_state=seed)
self._weights = None
self._hyperparameters.update(locals())
# We only want the function arguments so remove class related objects.
del self._hyperparameters['self']
del self._hyperparameters['__class__']
@property
def name(self): # noqa: D102
return 'slim'
def update(self, users=None, items=None, ratings=None): # noqa: D102
super().update(users, items, ratings)
num_items = len(self._items)
self._weights = scipy.sparse.dok_matrix((num_items, num_items))
if self._binarize:
row, col = self._ratings.nonzero()
data = np.ones(len(row))
ratings = scipy.sparse.csr_matrix((data, (row, col)), shape=self._ratings.shape).tolil()
else:
ratings = self._ratings.tolil()
for item_id in range(num_items):
target = ratings[:, item_id].toarray()
# Zero out the column of the current item to prevent a trivial solution.
ratings[:, item_id] = 0
# Fit the mode and save the weights
# This currently takes 0.02s/item on ML100k
self._model.fit(ratings, target)
self._weights[:, item_id] = self._model.sparse_coef_.T
self._weights[item_id, item_id] = 0
# Restore the rating column.
ratings[:, item_id] = target
self._weights = scipy.sparse.csr_matrix(self._weights)
@property
def dense_predictions(self): # noqa: D102
if self._dense_predictions is None:
self._dense_predictions = (self._ratings @ self._weights).todense()
return self._dense_predictions
def _predict(self, user_item): # noqa: D102
# Predict on all user-item pairs.
all_predictions = self.dense_predictions
predictions = []
for user_id, item_id, _ in user_item:
predictions.append(all_predictions[user_id, item_id])
return np.array(predictions)
class EASE(recommender.PredictRecommender):
"""The EASE recommendation model which is a simple linear method.
Parameters
----------
binarize : boolean
Determines whether to binarize ratings before fitting a model.
lam : float
Constant that multiplies the regularization terms.
"""
def __init__(self,
binarize=False,
lam=1.0):
"""Create an EASE recommender."""
super().__init__()
self._binarize = binarize
self._lam = lam
self._weights = None
self._hyperparameters.update(locals())
# We only want the function arguments so remove class related objects.
del self._hyperparameters['self']
del self._hyperparameters['__class__']
@property
def name(self): # noqa: D102
return 'ease'
def update(self, users=None, items=None, ratings=None): # noqa: D102
super().update(users, items, ratings)
if self._binarize:
row, col = self._ratings.nonzero()
data = np.ones(len(row))
ratings = scipy.sparse.csr_matrix((data, (row, col)), shape=self._ratings.shape)
else:
ratings = self._ratings
item_products = ratings.T @ ratings
diag_ind = np.diag_indices(item_products.shape[0])
item_products[diag_ind] += self._lam
inverse_mat = np.linalg.inv(item_products.todense())
self._weights = inverse_mat / (-np.diag(inverse_mat))
self._weights[diag_ind] = 0
@property
def dense_predictions(self): # noqa: D102
if self._dense_predictions is None:
self._dense_predictions = (self._ratings @ self._weights)
return self._dense_predictions
def _predict(self, user_item): # noqa: D102
# Predict on all user-item pairs.
all_predictions = self.dense_predictions
predictions = []
for user_id, item_id, _ in user_item:
predictions.append(all_predictions[user_id, item_id])
return np.array(predictions)Classes
class EASE (binarize=False, lam=1.0)-
The EASE recommendation model which is a simple linear method.
Parameters
binarize:boolean- Determines whether to binarize ratings before fitting a model.
lam:float- Constant that multiplies the regularization terms.
Create an EASE recommender.
Expand source code
class EASE(recommender.PredictRecommender): """The EASE recommendation model which is a simple linear method. Parameters ---------- binarize : boolean Determines whether to binarize ratings before fitting a model. lam : float Constant that multiplies the regularization terms. """ def __init__(self, binarize=False, lam=1.0): """Create an EASE recommender.""" super().__init__() self._binarize = binarize self._lam = lam self._weights = None self._hyperparameters.update(locals()) # We only want the function arguments so remove class related objects. del self._hyperparameters['self'] del self._hyperparameters['__class__'] @property def name(self): # noqa: D102 return 'ease' def update(self, users=None, items=None, ratings=None): # noqa: D102 super().update(users, items, ratings) if self._binarize: row, col = self._ratings.nonzero() data = np.ones(len(row)) ratings = scipy.sparse.csr_matrix((data, (row, col)), shape=self._ratings.shape) else: ratings = self._ratings item_products = ratings.T @ ratings diag_ind = np.diag_indices(item_products.shape[0]) item_products[diag_ind] += self._lam inverse_mat = np.linalg.inv(item_products.todense()) self._weights = inverse_mat / (-np.diag(inverse_mat)) self._weights[diag_ind] = 0 @property def dense_predictions(self): # noqa: D102 if self._dense_predictions is None: self._dense_predictions = (self._ratings @ self._weights) return self._dense_predictions def _predict(self, user_item): # noqa: D102 # Predict on all user-item pairs. all_predictions = self.dense_predictions predictions = [] for user_id, item_id, _ in user_item: predictions.append(all_predictions[user_id, item_id]) return np.array(predictions)Ancestors
- PredictRecommender
- Recommender
- abc.ABC
Inherited members
class SLIM (binarize=False, alpha=1.0, l1_ratio=0.1, positive=True, max_iter=100, tol=0.0001, seed=0)-
The SLIM recommendation model which is a sparse linear method.
Parameters
binarize:boolean- Determines whether to binarize ratings before fitting a model.
alpha:float- Constant that multiplies the regularization terms.
l1_ratio:float- The ratio of the L1 regularization term with respect to the L2 regularization.
max_iter:int- The maximum number of iterations to train the model for.
tol:float- The tolerance below which the optimization will stop.
seed:int- The random seed to use when training the model.
Create a SLIM recommender.
Expand source code
class SLIM(recommender.PredictRecommender): """The SLIM recommendation model which is a sparse linear method. Parameters ---------- binarize : boolean Determines whether to binarize ratings before fitting a model. alpha : float Constant that multiplies the regularization terms. l1_ratio : float The ratio of the L1 regularization term with respect to the L2 regularization. max_iter : int The maximum number of iterations to train the model for. tol : float The tolerance below which the optimization will stop. seed : int The random seed to use when training the model. """ def __init__(self, binarize=False, alpha=1.0, l1_ratio=0.1, positive=True, max_iter=100, tol=1e-4, seed=0): """Create a SLIM recommender.""" super().__init__() self._binarize = binarize self._model = sklearn.linear_model.ElasticNet(alpha=alpha, l1_ratio=l1_ratio, positive=positive, fit_intercept=False, copy_X=False, precompute=True, selection='random', max_iter=max_iter, tol=tol, random_state=seed) self._weights = None self._hyperparameters.update(locals()) # We only want the function arguments so remove class related objects. del self._hyperparameters['self'] del self._hyperparameters['__class__'] @property def name(self): # noqa: D102 return 'slim' def update(self, users=None, items=None, ratings=None): # noqa: D102 super().update(users, items, ratings) num_items = len(self._items) self._weights = scipy.sparse.dok_matrix((num_items, num_items)) if self._binarize: row, col = self._ratings.nonzero() data = np.ones(len(row)) ratings = scipy.sparse.csr_matrix((data, (row, col)), shape=self._ratings.shape).tolil() else: ratings = self._ratings.tolil() for item_id in range(num_items): target = ratings[:, item_id].toarray() # Zero out the column of the current item to prevent a trivial solution. ratings[:, item_id] = 0 # Fit the mode and save the weights # This currently takes 0.02s/item on ML100k self._model.fit(ratings, target) self._weights[:, item_id] = self._model.sparse_coef_.T self._weights[item_id, item_id] = 0 # Restore the rating column. ratings[:, item_id] = target self._weights = scipy.sparse.csr_matrix(self._weights) @property def dense_predictions(self): # noqa: D102 if self._dense_predictions is None: self._dense_predictions = (self._ratings @ self._weights).todense() return self._dense_predictions def _predict(self, user_item): # noqa: D102 # Predict on all user-item pairs. all_predictions = self.dense_predictions predictions = [] for user_id, item_id, _ in user_item: predictions.append(all_predictions[user_id, item_id]) return np.array(predictions)Ancestors
- PredictRecommender
- Recommender
- abc.ABC
Inherited members