def user_item_features_true_propensities_true_ratings(num_rows, num_cols,
num_components):
true_U = np.random.rand(num_rows, num_components)
true_V = np.random.rand(num_cols, num_components)
true_X = np.dot(true_U, true_V.T)
true_X = (true_X - true_X.min()) / (true_X.max() - true_X.min())
true_X = 1 + np.round(true_X * 4)
true_U_feature = np.random.randn(num_rows, num_components) / 8
true_V_feature = np.random.randn(num_cols, num_components) / 8
param_U = np.random.rand(num_components)
param_V = np.random.rand(num_components)
true_P = np.ones((num_rows, num_cols))
for u in range(num_rows):
for i in range(num_cols):
true_P[u, i] = \
std_logistic_function(np.inner(true_U_feature[u], param_U) +
np.inner(true_V_feature[i], param_V))
return true_P, true_X, true_U_feature, true_V_feature
def fit(self, trainset):
AlgoBase.fit(self, trainset)
M = np.zeros((trainset.n_users, trainset.n_items))
X = np.zeros((trainset.n_users, trainset.n_items))
for u, i, r in trainset.all_ratings():
M[u, i] = 1
X[u, i] = r
if type(self.propensity_scores) == np.ndarray:
P_hat = self.propensity_scores
elif self.propensity_scores == '1bitmc':
if self.verbose:
print('Estimating propensity scores via matrix completion...')
while True:
memoize_hash = hashlib.sha256(M.data.tobytes())
memoize_hash.update(
json.dumps([
self.one_bit_mc_tau, self.one_bit_mc_gamma,
self.one_bit_mc_max_rank
]).encode('utf-8'))
cache_filename = os.path.join(
cache_dir,
memoize_hash.hexdigest() + '.txt')
if not os.path.isfile(cache_filename):
os.makedirs(cache_dir, exist_ok=True)
P_hat = \
one_bit_MC_fully_observed(
M, std_logistic_function,
grad_std_logistic_function,
self.one_bit_mc_tau,
self.one_bit_mc_gamma,
max_rank=self.one_bit_mc_max_rank)
try:
np.savetxt(cache_filename, P_hat)
except:
pass
else:
try:
P_hat = np.loadtxt(cache_filename)
if P_hat.shape[0] != trainset.n_users or \
P_hat.shape[1] != trainset.n_items:
print(
'*** WARNING: Recomputing propensity scores ' +
'(mismatched dimensions in cached file)')
try:
os.remove(cache_filename)
except:
pass
continue
except ValueError:
print('*** WARNING: Recomputing propensity scores ' +
'(malformed numpy array encountered)')
try:
os.remove(cache_filename)
except:
pass
continue
break
elif self.propensity_scores == '1bitmc_mod':
if self.verbose:
print('Estimating propensity scores via matrix completion...')
while True:
memoize_hash = hashlib.sha256(M.data.tobytes())
memoize_hash.update(
json.dumps([
'1bitmc-mod', self.one_bit_mc_tau,
self.one_bit_mc_gamma, self.one_bit_mc_max_rank
]).encode('utf-8'))
cache_filename = os.path.join(
cache_dir,
memoize_hash.hexdigest() + '.txt')
if not os.path.isfile(cache_filename):
one_minus_logistic_gamma \
= 1 - std_logistic_function(self.one_bit_mc_gamma)
link = lambda x: \
mod_logistic_function(x, self.one_bit_mc_gamma,
one_minus_logistic_gamma)
grad_link = lambda x: \
grad_mod_logistic_function(x, self.one_bit_mc_gamma,
one_minus_logistic_gamma)
os.makedirs(cache_dir, exist_ok=True)
P_hat = \
one_bit_MC_mod_fully_observed(M, link, grad_link,
self.one_bit_mc_tau,
self.one_bit_mc_gamma,
max_rank=
self.one_bit_mc_max_rank)
try:
np.savetxt(cache_filename, P_hat)
except:
pass
else:
try:
P_hat = np.loadtxt(cache_filename)
if P_hat.shape[0] != trainset.n_users or \
P_hat.shape[1] != trainset.n_items:
print(
'*** WARNING: Recomputing propensity scores ' +
'(mismatched dimensions in cached file)')
try:
os.remove(cache_filename)
except:
pass
continue
except ValueError:
print('*** WARNING: Recomputing propensity scores ' +
'(malformed numpy array encountered)')
try:
os.remove(cache_filename)
except:
pass
continue
break
else:
raise Exception('Unknown weight method: ' + self.propensity_scores)
self.propensity_estimates = P_hat
if self.verbose:
print('Running debiased matrix completion...')
self.predictions = \
weighted_softimpute(X, M, 1 / P_hat, self.lmbda, self.max_rank,
self.min_value, self.max_value)
# modify training entries to be their observed values
for u, i, r in trainset.all_ratings():
self.predictions[u, i] = r
return self