Module reclab.recommenders.cfnade.cfnade
Implementation of the CF-NADE recommender using Keras.
Expand source code
"""Implementation of the CF-NADE recommender using Keras."""
from keras.layers import Input, Dropout, Lambda, add
from keras.models import Model
import keras.regularizers
from keras.optimizers import Adam
import numpy as np
from .cfnade_lib.nade import NADE
from .cfnade_lib import utils
from .. import recommender
class Cfnade(recommender.PredictRecommender):
"""
A Neural Autoregressive Distribution Estimator (NADE) for collaborative filtering (CF) tasks.
Parameters
---------
num_users : int
Number of users in the environment.
num_items : int
Number of items in the environment.
train_set : np.matrix
Matrix of shape (num_users, num_items) populated with user ratings.
train_epoch : int
Number of epochs to train for each call.
batch_size : int
Batch size during initial training phase.
rating_bucket: int
number of rating buckets
rate_score: array of float
An array of corresponding rating score for each bucket
hidden_dim: int
hidden dimension to construct the layer
learning_rate: float
learning rate
"""
def __init__(
self, num_users, num_items,
batch_size=64, train_epoch=10,
rating_bucket=5, hidden_dim=500,
learning_rate=0.001, normalized_layer=False):
"""Create new Cfnade recommender."""
super().__init__()
self._num_users = num_users
self._num_items = num_items
self._batch_size = batch_size
if num_items <= batch_size:
self._batch_size = num_items
self._input_dim0 = num_users
self._rating_bucket = rating_bucket
self._rate_score = np.array(np.arange(1, rating_bucket+1), np.float32)
self._hidden_dim = hidden_dim
self._learning_rate = learning_rate
self._train_epoch = train_epoch
self._hyperparameters.update(locals())
self._new_items = np.zeros(num_items)
# We only want the function arguments so remove class related objects.
del self._hyperparameters['self']
del self._hyperparameters['__class__']
# Prepare model
input_layer = Input(shape=(self._input_dim0, self._rating_bucket), name='input_ratings')
output_ratings = Input(shape=(self._input_dim0, self._rating_bucket), name='output_ratings')
input_masks = Input(shape=(self._input_dim0,), name='input_masks')
output_masks = Input(shape=(self._input_dim0,), name='output_masks')
nade_layer = Dropout(0.0)(input_layer)
nade_layer = NADE(
hidden_dim=self._hidden_dim, activation='tanh', bias=True,
W_regularizer=keras.regularizers.l2(0.02),
V_regularizer=keras.regularizers.l2(0.02),
b_regularizer=keras.regularizers.l2(0.02),
c_regularizer=keras.regularizers.l2(0.02),
normalized_layer=normalized_layer)(nade_layer)
predicted_ratings = Lambda(
utils.prediction_layer,
output_shape=utils.prediction_output_shape,
name='predicted_ratings')(nade_layer)
func_d = Lambda(
utils.d_layer, output_shape=utils.d_output_shape,
name='func_d')(input_masks)
sum_masks = add([input_masks, output_masks])
func_d_2 = Lambda(
utils.D_layer, output_shape=utils.D_output_shape,
name='func_d_2')(sum_masks)
loss_out = Lambda(
utils.rating_cost_lambda_func, output_shape=(1, ),
name='nade_loss')([nade_layer, output_ratings,
input_masks, output_masks, func_d_2, func_d])
self._cf_nade_model = Model(
inputs=[input_layer, output_ratings, input_masks, output_masks],
outputs=[loss_out, predicted_ratings])
optimizer = Adam(self._learning_rate, 0.9, 0.999, 1e-8)
self._cf_nade_model.compile(
loss={'nade_loss': lambda y_true, y_pred: y_pred},
optimizer=optimizer)
self._cf_nade_model.save_weights('model.h5')
@property
def name(self): # noqa: D102
return 'cfnade'
def update(self, users=None, items=None, ratings=None): # noqa: D102
super().update(users, items, ratings)
self._cf_nade_model.load_weights('model.h5')
ratings_matrix = self._ratings.toarray()
ratings_matrix = np.around(ratings_matrix.transpose())
ratings_matrix = ratings_matrix.astype(int)
train_set = utils.DataSet(ratings_matrix,
num_users=self._num_users,
num_items=self._num_items,
batch_size=self._batch_size,
rating_bucket=self._rating_bucket,
mode=0)
self._cf_nade_model.fit_generator(train_set.generate(),
steps_per_epoch=(self._num_items // self._batch_size),
epochs=self._train_epoch,
callbacks=[train_set], verbose=1)
def _predict(self, user_item): # noqa: D102
ratings_matrix = self._ratings.toarray()
ratings_matrix = np.around(ratings_matrix.transpose())
ratings_matrix = ratings_matrix.astype(int)
# keep track of unseen items in ratings
ratings_matrix_total = ratings_matrix.transpose().sum(axis=1)
self._new_items = np.where(ratings_matrix_total == 0)[0]
test_set = utils.DataSet(ratings_matrix,
num_users=self._num_users,
num_items=self._num_items,
batch_size=self._batch_size,
rating_bucket=self._rating_bucket,
mode=2)
pred_rating = []
for batch in test_set.generate():
pred_matrix = self._cf_nade_model.predict(batch[0])[1]
pred_rating_batch = pred_matrix * self._rate_score[np.newaxis, np.newaxis, :]
pred_rating_batch = pred_rating_batch.sum(axis=2)
pred_rating.append(pred_rating_batch)
pred_rating = np.concatenate(pred_rating, axis=0)
predictions = []
for user, item, _ in user_item:
if item in self._new_items:
predictions.append(3)
else:
predictions.append(pred_rating[item, user])
return np.array(predictions)Classes
class Cfnade (num_users, num_items, batch_size=64, train_epoch=10, rating_bucket=5, hidden_dim=500, learning_rate=0.001, normalized_layer=False)-
A Neural Autoregressive Distribution Estimator (NADE) for collaborative filtering (CF) tasks.
Parameters
num_users:int- Number of users in the environment.
num_items:int- Number of items in the environment.
train_set:np.matrix- Matrix of shape (num_users, num_items) populated with user ratings.
train_epoch:int- Number of epochs to train for each call.
batch_size:int- Batch size during initial training phase.
rating_bucket:int- number of rating buckets
rate_score:arrayoffloat- An array of corresponding rating score for each bucket
hidden_dim:int- hidden dimension to construct the layer
learning_rate:float- learning rate
Create new Cfnade recommender.
Expand source code
class Cfnade(recommender.PredictRecommender): """ A Neural Autoregressive Distribution Estimator (NADE) for collaborative filtering (CF) tasks. Parameters --------- num_users : int Number of users in the environment. num_items : int Number of items in the environment. train_set : np.matrix Matrix of shape (num_users, num_items) populated with user ratings. train_epoch : int Number of epochs to train for each call. batch_size : int Batch size during initial training phase. rating_bucket: int number of rating buckets rate_score: array of float An array of corresponding rating score for each bucket hidden_dim: int hidden dimension to construct the layer learning_rate: float learning rate """ def __init__( self, num_users, num_items, batch_size=64, train_epoch=10, rating_bucket=5, hidden_dim=500, learning_rate=0.001, normalized_layer=False): """Create new Cfnade recommender.""" super().__init__() self._num_users = num_users self._num_items = num_items self._batch_size = batch_size if num_items <= batch_size: self._batch_size = num_items self._input_dim0 = num_users self._rating_bucket = rating_bucket self._rate_score = np.array(np.arange(1, rating_bucket+1), np.float32) self._hidden_dim = hidden_dim self._learning_rate = learning_rate self._train_epoch = train_epoch self._hyperparameters.update(locals()) self._new_items = np.zeros(num_items) # We only want the function arguments so remove class related objects. del self._hyperparameters['self'] del self._hyperparameters['__class__'] # Prepare model input_layer = Input(shape=(self._input_dim0, self._rating_bucket), name='input_ratings') output_ratings = Input(shape=(self._input_dim0, self._rating_bucket), name='output_ratings') input_masks = Input(shape=(self._input_dim0,), name='input_masks') output_masks = Input(shape=(self._input_dim0,), name='output_masks') nade_layer = Dropout(0.0)(input_layer) nade_layer = NADE( hidden_dim=self._hidden_dim, activation='tanh', bias=True, W_regularizer=keras.regularizers.l2(0.02), V_regularizer=keras.regularizers.l2(0.02), b_regularizer=keras.regularizers.l2(0.02), c_regularizer=keras.regularizers.l2(0.02), normalized_layer=normalized_layer)(nade_layer) predicted_ratings = Lambda( utils.prediction_layer, output_shape=utils.prediction_output_shape, name='predicted_ratings')(nade_layer) func_d = Lambda( utils.d_layer, output_shape=utils.d_output_shape, name='func_d')(input_masks) sum_masks = add([input_masks, output_masks]) func_d_2 = Lambda( utils.D_layer, output_shape=utils.D_output_shape, name='func_d_2')(sum_masks) loss_out = Lambda( utils.rating_cost_lambda_func, output_shape=(1, ), name='nade_loss')([nade_layer, output_ratings, input_masks, output_masks, func_d_2, func_d]) self._cf_nade_model = Model( inputs=[input_layer, output_ratings, input_masks, output_masks], outputs=[loss_out, predicted_ratings]) optimizer = Adam(self._learning_rate, 0.9, 0.999, 1e-8) self._cf_nade_model.compile( loss={'nade_loss': lambda y_true, y_pred: y_pred}, optimizer=optimizer) self._cf_nade_model.save_weights('model.h5') @property def name(self): # noqa: D102 return 'cfnade' def update(self, users=None, items=None, ratings=None): # noqa: D102 super().update(users, items, ratings) self._cf_nade_model.load_weights('model.h5') ratings_matrix = self._ratings.toarray() ratings_matrix = np.around(ratings_matrix.transpose()) ratings_matrix = ratings_matrix.astype(int) train_set = utils.DataSet(ratings_matrix, num_users=self._num_users, num_items=self._num_items, batch_size=self._batch_size, rating_bucket=self._rating_bucket, mode=0) self._cf_nade_model.fit_generator(train_set.generate(), steps_per_epoch=(self._num_items // self._batch_size), epochs=self._train_epoch, callbacks=[train_set], verbose=1) def _predict(self, user_item): # noqa: D102 ratings_matrix = self._ratings.toarray() ratings_matrix = np.around(ratings_matrix.transpose()) ratings_matrix = ratings_matrix.astype(int) # keep track of unseen items in ratings ratings_matrix_total = ratings_matrix.transpose().sum(axis=1) self._new_items = np.where(ratings_matrix_total == 0)[0] test_set = utils.DataSet(ratings_matrix, num_users=self._num_users, num_items=self._num_items, batch_size=self._batch_size, rating_bucket=self._rating_bucket, mode=2) pred_rating = [] for batch in test_set.generate(): pred_matrix = self._cf_nade_model.predict(batch[0])[1] pred_rating_batch = pred_matrix * self._rate_score[np.newaxis, np.newaxis, :] pred_rating_batch = pred_rating_batch.sum(axis=2) pred_rating.append(pred_rating_batch) pred_rating = np.concatenate(pred_rating, axis=0) predictions = [] for user, item, _ in user_item: if item in self._new_items: predictions.append(3) else: predictions.append(pred_rating[item, user]) return np.array(predictions)Ancestors
- PredictRecommender
- Recommender
- abc.ABC
Inherited members