def _iter_indices_mask(self, x, y, indices):
if self._shuffle:
check_random_state(self._random_state).shuffle(indices)
mask = np.zeros(len(indices), dtype=np.bool)
for i in indices:
new_mask = mask.copy()
new_mask[i] = True
yield new_mask
def _iter_indices_mask(self, x, y, indices):
if self._shuffle:
check_random_state(self._random_state).shuffle(indices)
n_splits = self._folds
fold_sizes = (len(x) // n_splits) * np.ones(n_splits, dtype=np.int)
fold_sizes[:len(x) % n_splits] += 1
current = 0
mask = np.zeros(len(indices), dtype=np.bool)
for fold_size in fold_sizes:
start, stop = current, current + fold_size
copy_mask = np.copy(mask)
copy_mask[indices[start:stop]] = True
current = stop
yield copy_mask
def __init__(self,
ratio=0.8,
times=10,
user_idx=0,
item_idx=1,
random_state=None):
super(UserHoldOut, self).__init__()
self._times = times
self._ratio = ratio
self._user_idx = user_idx
self._item_idx = item_idx
self._random_state = check_random_state(random_state)
def split(self, x, y):
data = pd.DataFrame(x)
n_samples = data.shape[0]
indices = np.arange(len(x))
for i in range(self._times):
check_random_state(self._random_state).shuffle(indices)
train_size = int(n_samples * self._ratio)
# split data according to the shuffled index and the holdout size
train_idx = indices[:train_size]
train_split = data.ix[indices[:train_size]]
test_split = data.ix[indices[train_size:]]
# remove new user and new items from the test split
train_users = train_split[self._user_idx].unique()
train_items = train_split[self._item_idx].unique()
test_idx = test_split.index[
(test_split[self._user_idx].isin(train_users))
& (test_split[self._item_idx].isin(train_items))]
yield train_idx, test_idx
def _iter_indices_mask(self, x, y, indices):
data = pd.DataFrame(x)[[self._user_idx, self._item_idx]]
mask = np.zeros(data.shape[0], dtype=np.bool)
for i in range(self._times):
copy_mask = np.copy(mask)
grouped = data.groupby(0)
for user, g in grouped:
idx_shuffled = g.index.values.reshape(-1)
n_observed = int((1 - self._ratio) * len(idx_shuffled))
check_random_state(self._random_state).shuffle(idx_shuffled)
copy_mask[idx_shuffled[0:n_observed]] = True
# cleaning
train_split = data.ix[indices[np.logical_not(copy_mask)]]
test_split = data.ix[indices[copy_mask]]
# remove new user and new items from the test split
train_items = train_split[self._item_idx].unique()
test_idx = test_split.index[~test_split[self._item_idx].
isin(train_items)]
copy_mask[test_idx] = False
print(1 - (sum(copy_mask) / len(copy_mask)))
yield copy_mask
def split(self, x, y):
data = pd.DataFrame(x)
n_samples = data.shape[0]
indices = np.arange(len(x))
for i in range(self._times):
check_random_state(self._random_state).shuffle(indices)
train_size = int(n_samples * self._ratio)
# split data according to the shuffled index and the holdout size
train_idx = indices[:train_size]
test_idx = indices[train_size:]
# This block of code checks whether the user and the item
# in the test set belong to the train_set otherwise
# they are dropped
# train_split = data.ix[indices[:train_size]]
# test_split = data.ix[indices[train_size:]]
# train_users = train_split[self._user_idx].unique()
# train_items = train_split[self._item_idx].unique()
# test_idx = test_split.index[(test_split[self._user_idx].isin(train_users)) &
# (test_split[self._item_idx].isin(train_items))]
yield train_idx, test_idx
