def __init__(self, ml_type='classifier', metric=None, ml_p=None):
"""Initialize recommendation system."""
super().__init__(ml_type, metric, ml_p)
# store results
self.results_df = pd.DataFrame()
# reader for translating btw PennAI results and Suprise training set
self.reader = Reader()
# algo is the online Surprise-based rec system
self.algo = mySVD(n_factors=20,
n_epochs=20,
biased=True,
init_mean=0,
init_std_dev=.2,
lr_all=.01,
reg_all=.02,
lr_bu=None,
lr_bi=None,
lr_pu=None,
lr_qi=None,
reg_bu=None,
reg_bi=None,
reg_pu=None,
reg_qi=None,
random_state=None,
verbose=False)
self.first_fit = True
self.max_epochs = 100
def set_algo(self, surprise_kwargs={}):
alg_kwargs = {
'n_factors': 20,
'biased': True,
'init_mean': 0,
'init_std_dev': .2,
'lr_all': .01,
'reg_all': .02,
'verbose': False
}
alg_kwargs.update(surprise_kwargs)
self.algo = mySVD(**alg_kwargs)
class SVDRecommender(SurpriseRecommender):
"""SVD recommender.
see https://surprise.readthedocs.io/en/stable/matrix_factorization.html
Recommends machine learning algorithms and parameters using the SVD algorithm.
- stores ML + P and every dataset.
- learns a matrix factorization on the non-missing data.
- given a dataset, estimates the rankings of all ML+P and returns the top
n_recs.
Note that we use a custom online version of SVD found here:
https://github.com/lacava/surprise
"""
algo = mySVD(n_factors=20,
biased=True,
init_mean=0,
init_std_dev=.2,
lr_all=.01,
reg_all=.02,
lr_bu=None,
lr_bi=None,
lr_pu=None,
lr_qi=None,
reg_bu=None,
reg_bi=None,
reg_pu=None,
reg_qi=None,
random_state=None,
verbose=False)
def update_model(self, results_data):
"""Stores new results and updates SVD."""
logger.info('updating SVD model')
# shuffle the results data the first time
if self.first_fit:
results_data = results_data.sample(frac=1)
self.update_training_data(results_data)
logger.debug('fitting self.algo...')
# set the number of training iterations proportionally to the amount of
# results_data
# self.algo.n_epochs = min(len(results_data),self.max_epochs)
# self.algo.n_epochs = max(10,self.algo.n_epochs)
self.algo.n_epochs = min(len(results_data), self.max_epochs)
self.algo.n_epochs = max(self.algo.n_epochs, self.min_epochs)
self.algo.partial_fit(self.trainset)
logger.debug('done.')
if self.first_fit:
self.init_results_data = results_data
self.first_fit = False
logger.debug('model SVD updated')