def _fit(self, recommender_data, users, items, matrix):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
seed=254938879,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(negative_rate=0))
negative_sample_generator.set_train_matrix(matrix)
negative_sample_generator.set_items(items)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner = rs.OfflineIteratingImplicitLearner(**self.parameter_defaults(
seed=254938879,
number_of_iterations=9,
))
learner.set_gradient_computer(gradient_computer)
learner.set_negative_sample_generator(negative_sample_generator)
learner.set_model(model)
learner.set_recommender_data(recommender_data)
learner.add_gradient_updater(updater)
return (model, learner)
def _fit(self, recommender_data, users, items, matrix):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
seed=self.parameter_default('factor_seed', 67439852),
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
gradient_computer.add_gradient_updater(updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0, initialize_all=False, max_item=-1))
negative_sample_generator.set_train_matrix(matrix)
negative_sample_generator.set_items(items)
negative_sample_generator.add_updater(gradient_computer)
learner = rs.OfflineIteratingOnlineLearnerWrapper(
**self.parameter_defaults(
seed=self.parameter_default('learner_seed', 254938879),
number_of_iterations=9,
shuffle=True,
))
learner.add_iterate_updater(negative_sample_generator)
return (model, learner)
def _config(self, top_k, seed):
model = rs.AsymmetricFactorModel(
**self.parameter_defaults(begin_min=-0.01,
begin_max=0.01,
dimension=10,
use_sigmoid=False,
norm_type="exponential",
gamma=0.8,
initialize_all=False))
gradient_updater = rs.AsymmetricFactorModelGradientUpdater(
**self.parameter_defaults(
learning_rate=0.05,
cumulative_item_updates=False,
))
gradient_updater.set_model(model)
simple_updater = rs.AsymmetricFactorModelUpdater()
simple_updater.set_model(model)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
gradient_computer.add_gradient_updater(gradient_updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=20,
initialize_all=False,
seed=928357823,
))
negative_sample_generator.add_updater(gradient_computer)
return (model, [negative_sample_generator, simple_updater], [])
def _config(self, top_k, seed):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
learner = rs.ImplicitGradientLearner()
learner.add_gradient_updater(updater)
learner.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0.0,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
learner.set_negative_sample_generator(negative_sample_generator)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
return (model, learner, [])
def config(self, elems):
config = self.parameter_defaults(
top_k=100,
min_time=0,
seed=0,
out_file=None,
filters=[],
loggers=[],
)
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
learner = rs.ImplicitGradientLearner()
learner.add_gradient_updater(updater)
learner.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0.0,
initialize_all=False,
seed=0,
))
learner.set_negative_sample_generator(negative_sample_generator)
pointWise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(pointWise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
fmfilter = rs.FactorModelFilter()
fmfilter.set_model(model)
prediction_creator = rs.PredictionCreatorGlobal(
**self.parameter_defaults(
top_k=10000,
# initial_threshold=1000,
lookback=0))
prediction_creator.set_model(model)
prediction_creator.set_filter(fmfilter)
online_predictor = rs.OnlinePredictor(**self.parameter_defaults(
min_time=0, time_frame=86400, file_name=""))
online_predictor.set_prediction_creator(prediction_creator)
config['loggers'].append(online_predictor)
return {'config': config, 'model': model, 'learner': learner}
def _config(self, top_k, seed):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
learner_parameters = self.parameter_defaults(
number_of_iterations=3,
start_time=-1,
period_length=86400,
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name="",
timeframe_length=0,
)
if (learner_parameters['timeframe_length'] == 0):
learner_parameters.pop('timeframe_length', None)
learner = rs.OfflineImplicitGradientLearner(**learner_parameters)
else:
learner = rs.PeriodicTimeframeImplicitGradientLearner(
**learner_parameters)
learner.set_model(model)
learner.add_gradient_updater(updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
learner.set_negative_sample_generator(negative_sample_generator)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
return (model, learner, [], [])
def config(self, elems):
config = self.parameter_defaults(
top_k=100,
min_time=0,
seed=0,
out_file=None,
filters=[],
loggers=[],
)
model = rs.SvdppModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
use_sigmoid=False,
norm_type="exponential",
gamma=0.8,
user_vector_weight=0.5,
history_weight=0.5
))
gradient_updater = rs.SvdppModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05,
cumulative_item_updates=False,
))
gradient_updater.set_model(model)
simple_updater = rs.SvdppModelUpdater()
simple_updater.set_model(model)
learner = rs.ImplicitGradientLearner()
learner.add_gradient_updater(gradient_updater)
learner.add_simple_updater(simple_updater)
learner.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(**self.parameter_defaults(
negative_rate=20,
initialize_all=False,
seed=928357823,
))
learner.set_negative_sample_generator(negative_sample_generator)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
return {
'config': config,
'model': model,
'learner': learner
}
def _config(self, top_k, seed):
model = rs.RandomChoosingCombinedModel()
updater = rs.RandomChoosingCombinedModelExpertUpdater(
**self.parameter_defaults(
eta=0.1,
top_k=top_k,
loss_type="abs",
))
updater.set_model(model)
pop_model = rs.PopularityModel()
model.add_model(pop_model)
pop_updater = rs.PopularityModelUpdater()
pop_updater.set_model(pop_model)
pop_model2 = rs.PopularityModel(
) #not updated popularity model, predicts 0 for all items and users
model.add_model(pop_model2)
pop_model3 = rs.PopularityModel(
) #not updated popularity model, predicts 0 for all items and users
model.add_model(pop_model3)
factor_model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
model.add_model(factor_model)
factor_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=0.05,
regularization_rate=0.0))
factor_updater.set_model(factor_model)
objective = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(objective)
gradient_computer.set_model(factor_model)
gradient_computer.add_gradient_updater(factor_updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=10,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
negative_sample_generator.add_updater(gradient_computer)
return (model, [updater, pop_updater,
negative_sample_generator], [], [])
def config(self, elems):
config = self.parameter_defaults(
top_k=100,
min_time=0,
loggers=[],
)
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
learner = rs.OfflineImplicitGradientLearner(
**self.parameter_defaults(number_of_iterations=3,
start_time=-1,
period_length=86400,
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name=""))
learner.set_model(model)
# learner.set_recommender_data_iterator(elems['recommender_data_iterator'])
learner.add_gradient_updater(updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0,
initialize_all=False,
))
# negative_sample_generator.set_train_matrix(elems['train_matrix'])
# negative_sample_generator.set_items(elems['items'])
learner.set_negative_sample_generator(negative_sample_generator)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
return {'config': config, 'model': model, 'learner': learner}
def _fit(self, recommender_data, users, items, matrix):
model = rs.SvdppModel(**self.parameter_defaults(
begin_min=self.parameter_default("begin_min", -0.01),
begin_max=self.parameter_default("begin_max", 0.01),
dimension=self.parameter_default("dimension", 10),
use_sigmoid=False,
norm_type="constant",
gamma=1,
user_vector_weight=0.5,
history_weight=0.5))
gradient_updater = rs.SvdppModelGradientUpdater(
**self.parameter_defaults(
learning_rate=0.05,
cumulative_item_updates=False,
))
gradient_updater.set_model(model)
simple_updater = rs.SvdppModelUpdater()
simple_updater.set_model(model)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
gradient_computer.add_gradient_updater(gradient_updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=9, initialize_all=False, max_item=-1))
negative_sample_generator.set_train_matrix(matrix)
negative_sample_generator.set_items(items)
negative_sample_generator.add_updater(gradient_computer)
learner = rs.OfflineIteratingOnlineLearnerWrapper(
**self.parameter_defaults(
seed=254938879,
number_of_iterations=20,
shuffle=True,
))
learner.add_early_updater(simple_updater)
learner.add_iterate_updater(negative_sample_generator)
return (model, learner)
def config(self, elems):
config = self.parameter_defaults(
top_k=100,
min_time=0,
seed=0,
out_file=None,
filters=[],
loggers=[],
)
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
learner = rs.ImplicitGradientLearner()
learner.add_gradient_updater(updater)
learner.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0.0,
initialize_all=False,
seed=0,
))
learner.set_negative_sample_generator(negative_sample_generator)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner.set_gradient_computer(gradient_computer)
return {'config': config, 'model': model, 'learner': learner}
def _config(self, top_k, seed):
model = rs.CombinedModel(**self.parameter_defaults(
los_file_name="my_log_file",
log_frequency=100000,
use_user_weights=False,
))
pop_model = rs.PopularityModel()
model.add_model(pop_model)
pop_updater = rs.PopularityModelUpdater()
pop_updater.set_model(pop_model)
factor_model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
model.add_model(factor_model)
factor_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=0.05,
regularization_rate=0.0))
factor_updater.set_model(factor_model)
objective = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(objective)
gradient_computer.set_model(factor_model)
gradient_computer.add_gradient_updater(factor_updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=10,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
negative_sample_generator.add_updater(gradient_computer)
return (model, [pop_updater, negative_sample_generator], [], [])
def _fit(self, recommender_data, users, items, matrix):
model = rs.AsymmetricFactorModel(
begin_min=self.parameter_default("begin_min", -0.01),
begin_max=self.parameter_default("begin_max", 0.01),
dimension=self.parameter_default("dimension", 10),
use_sigmoid=False,
norm_type="disabled",
gamma=1)
updater = rs.AsymmetricFactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=0.05,
regularization_rate=0.0))
updater.set_model(model)
simple_updater = rs.AsymmetricFactorModelUpdater()
simple_updater.set_model(model)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(negative_rate=0))
negative_sample_generator.set_train_matrix(matrix)
negative_sample_generator.set_items(items)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
learner = rs.OfflineIteratingImplicitLearner(**self.parameter_defaults(
seed=254938879,
number_of_iterations=9,
))
learner.set_gradient_computer(gradient_computer)
learner.set_negative_sample_generator(negative_sample_generator)
learner.set_model(model)
learner.set_recommender_data(recommender_data)
learner.add_gradient_updater(updater)
learner.add_early_simple_updater(simple_updater)
return (model, learner)
def _config(self, top_k, seed):
#config = self.parameter_defaults(
# top_k=100,
# evaluation_start_time=0,
# seed=0,
# out_file=None,
# filters=[],
# loggers=[],
#)
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
gradient_computer.add_gradient_updater(updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=100,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
negative_sample_generator.add_updater(gradient_computer)
return (model, [negative_sample_generator], [])
def _config(self, top_k, seed):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
updater = rs.FactorModelGradientUpdater(**self.parameter_defaults(
learning_rate=0.05, regularization_rate=0.0))
updater.set_model(model)
point_wise = rs.ObjectiveMSE()
gradient_computer = rs.GradientComputerPointWise()
gradient_computer.set_objective(point_wise)
gradient_computer.set_model(model)
gradient_computer.add_gradient_updater(updater)
negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=0,
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
negative_sample_generator.add_updater(gradient_computer)
offline_learner = rs.OfflineIteratingOnlineLearnerWrapper(
**self.parameter_defaults(
seed=254938879,
number_of_iterations=3,
shuffle=True,
))
offline_learner.add_iterate_updater(negative_sample_generator)
online_learner = rs.PeriodicOfflineLearnerWrapper(
**self.parameter_defaults(
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name="",
))
online_learner.set_model(model)
online_learner.add_offline_learner(offline_learner)
data_generator_parameters = self.parameter_defaults(
timeframe_length=0, )
if (data_generator_parameters['timeframe_length'] == 0):
data_generator = rs.CompletePastDataGenerator()
else:
data_generator = rs.TimeframeDataGenerator(
**data_generator_parameters)
online_learner.set_data_generator(data_generator)
period_computer = rs.PeriodComputer(**self.parameter_defaults(
period_length=86400,
start_time=-1,
period_mode="time",
))
online_learner.set_period_computer(period_computer)
return (model, online_learner, [])
def _config(self, top_k, seed):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
#
# batch
#
# updater
batch_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=self.parameter_default(
'batch_learning_rate', 0.05),
regularization_rate=0.0))
batch_updater.set_model(model)
# objective
point_wise = rs.ObjectiveMSE()
batch_gradient_computer = rs.GradientComputerPointWise()
batch_gradient_computer.set_objective(point_wise)
batch_gradient_computer.set_model(model)
batch_gradient_computer.add_gradient_updater(batch_updater)
# negative sample generator
batch_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('batch_negative_rate',
70),
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
batch_negative_sample_generator.add_updater(batch_gradient_computer)
batch_offline_learner = rs.OfflineIteratingOnlineLearnerWrapper(
**self.parameter_defaults(
seed=254938879,
number_of_iterations=3,
shuffle=True,
))
batch_offline_learner.add_iterate_updater(
batch_negative_sample_generator)
batch_online_learner = rs.PeriodicOfflineLearnerWrapper(
**self.parameter_defaults(
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name="",
))
batch_online_learner.set_model(model)
batch_online_learner.add_offline_learner(batch_offline_learner)
batch_data_generator_parameters = self.parameter_defaults(
timeframe_length=0, )
if (batch_data_generator_parameters['timeframe_length'] == 0):
print("Full experiment")
batch_data_generator = rs.CompletePastDataGenerator()
else:
print("Timeframe experiment")
batch_data_generator = rs.TimeframeDataGenerator(
**batch_data_generator_parameters)
batch_online_learner.set_data_generator(batch_data_generator)
batch_period_computer = rs.PeriodComputer(**self.parameter_defaults(
period_length=86400,
start_time=-1,
period_mode="time",
))
batch_online_learner.set_period_computer(batch_period_computer)
#
# online
#
# updater
online_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=self.parameter_default(
'online_learning_rate', 0.2),
regularization_rate=0.0))
online_updater.set_model(model)
# objective
point_wise = rs.ObjectiveMSE()
online_gradient_computer = rs.GradientComputerPointWise()
online_gradient_computer.set_objective(point_wise)
online_gradient_computer.set_model(model)
online_gradient_computer.add_gradient_updater(online_updater)
# negative sample generator
online_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('online_negative_rate',
100),
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
online_negative_sample_generator.add_updater(online_gradient_computer)
learner = [batch_online_learner, online_negative_sample_generator]
return (model, learner, [])
def _config(self, top_k, seed):
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
#
# batch
#
# updater
batch_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=self.parameter_default(
'batch_learning_rate', 0.05),
regularization_rate=0.0))
batch_updater.set_model(model)
# negative sample generator
batch_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('batch_negative_rate',
70),
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
# objective
point_wise = rs.ObjectiveMSE()
batch_gradient_computer = rs.GradientComputerPointWise()
batch_gradient_computer.set_objective(point_wise)
batch_gradient_computer.set_model(model)
# learner
batch_learner_parameters = self.parameter_defaults(
number_of_iterations=9,
start_time=-1,
period_length=86400,
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name="",
timeframe_length=0,
)
if (batch_learner_parameters['timeframe_length'] == 0):
batch_learner_parameters.pop('timeframe_length', None)
batch_learner = rs.OfflineImplicitGradientLearner(
**batch_learner_parameters)
else:
batch_learner = rs.PeriodicTimeframeImplicitGradientLearner(
**batch_learner_parameters)
batch_learner.set_model(model)
batch_learner.add_gradient_updater(batch_updater)
batch_learner.set_gradient_computer(batch_gradient_computer)
batch_learner.set_negative_sample_generator(
batch_negative_sample_generator)
#
# online
#
# updater
online_updater = rs.FactorModelGradientUpdater(
**self.parameter_defaults(learning_rate=self.parameter_default(
'online_learning_rate', 0.2),
regularization_rate=0.0))
online_updater.set_model(model)
# negative sample generator
online_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('online_negative_rate',
100),
initialize_all=False,
seed=67439852,
filter_repeats=False,
))
# objective
point_wise = rs.ObjectiveMSE()
online_gradient_computer = rs.GradientComputerPointWise()
online_gradient_computer.set_objective(point_wise)
online_gradient_computer.set_model(model)
# learner
online_learner = rs.ImplicitGradientLearner()
online_learner.add_gradient_updater(online_updater)
online_learner.set_model(model)
online_learner.set_negative_sample_generator(
online_negative_sample_generator)
online_learner.set_gradient_computer(online_gradient_computer)
learner = [batch_learner, online_learner]
return (model, learner, [], [])
def config(self, elems):
config = self.parameter_defaults(
top_k=100,
min_time=0,
loggers=[],
)
model = rs.FactorModel(**self.parameter_defaults(
begin_min=-0.01,
begin_max=0.01,
dimension=10,
initialize_all=False,
))
#
# batch
#
# updater
batch_updater = rs.FactorModelGradientUpdater(
learning_rate=self.parameter_default('batch_learning_rate', 0.05),
regularization_rate=0.0)
batch_updater.set_model(model)
# negative sample generator
batch_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('batch_negative_rate',
70),
initialize_all=False,
))
# objective
point_wise = rs.ObjectiveMSE()
batch_gradient_computer = rs.GradientComputerPointWise()
batch_gradient_computer.set_objective(point_wise)
batch_gradient_computer.set_model(model)
# learner
batch_learner = rs.OfflineImplicitGradientLearner(
**self.parameter_defaults(number_of_iterations=9,
start_time=-1,
period_length=86400,
write_model=False,
read_model=False,
clear_model=False,
learn=True,
base_out_file_name="",
base_in_file_name=""))
batch_learner.set_model(model)
batch_learner.add_gradient_updater(batch_updater)
batch_learner.set_gradient_computer(batch_gradient_computer)
batch_learner.set_negative_sample_generator(
batch_negative_sample_generator)
#
# online
#
# updater
online_updater = rs.FactorModelGradientUpdater(
learning_rate=self.parameter_default('online_learning_rate', 0.2),
regularization_rate=0.0)
online_updater.set_model(model)
# negative sample generator
online_negative_sample_generator = rs.UniformNegativeSampleGenerator(
**self.parameter_defaults(
negative_rate=self.parameter_default('online_negative_rate',
100),
initialize_all=False,
))
# objective
point_wise = rs.ObjectiveMSE()
online_gradient_computer = rs.GradientComputerPointWise()
online_gradient_computer.set_objective(point_wise)
online_gradient_computer.set_model(model)
# learner
online_learner = rs.ImplicitGradientLearner()
online_learner.add_gradient_updater(online_updater)
online_learner.set_model(model)
online_learner.set_negative_sample_generator(
online_negative_sample_generator)
online_learner.set_gradient_computer(online_gradient_computer)
learner = [batch_learner, online_learner]
return {'config': config, 'model': model, 'learner': learner}
