def split(self, dataset):
super(ColdItemsKFold, self).split(dataset)
folds = []
split_belonging = np.random.choice(self.n_folds, dataset.n_items, replace=True)
for i in range(self.n_folds):
urm = {}
urm_mappers = {}
mask = split_belonging != i
for URM_name in dataset.get_URM_names():
URM = dataset.get_URM(URM_name).tocsc(copy=True)
# Sort nnz values by row and column indices, in order to remain consistent in the splits of different URMs
for j in np.arange(URM.shape[1])[mask].tolist():
URM.data[URM.indptr[j]:URM.indptr[j + 1]] = 0.0
URM.eliminate_zeros()
urm[URM_name] = URM.tocsr()
urm_mappers[URM_name] = dataset.get_URM_mapper(URM_name)
folds.append(
Dataset(dataset.get_name(), base_folder=dataset.get_base_folder(),
postprocessings=dataset.get_postprocessings(),
URM_dict=urm, URM_mappers_dict=urm_mappers,
ICM_dict=dataset.get_ICM_dict(), ICM_mappers_dict=dataset.get_ICM_mappers_dict(),
UCM_dict=dataset.get_UCM_dict(), UCM_mappers_dict=dataset.get_UCM_mappers_dict()
)
)
r = []
for i in range(self.n_folds):
urm = {}
urm_mappers = {}
for URM_name in folds[i].get_URM_names():
# Keep i-th fold as test and merge the others as train
urm[URM_name] = folds[(i + 1) % self.n_folds].get_URM(URM_name)
urm_mappers[URM_name] = folds[(i + 1) % self.n_folds].get_URM_mapper(URM_name)
for j in range(2, self.n_folds):
urm[URM_name] += folds[(i + j) % self.n_folds].get_URM(URM_name)
train = Dataset(folds[i].get_name(), base_folder=folds[i].get_base_folder(),
postprocessings=folds[i].get_postprocessings(),
URM_dict=urm, URM_mappers_dict=urm_mappers,
ICM_dict=folds[i].get_ICM_dict(), ICM_mappers_dict=folds[i].get_ICM_mappers_dict(),
UCM_dict=folds[i].get_UCM_dict(), UCM_mappers_dict=folds[i].get_UCM_mappers_dict())
urm = {}
test_urm = folds[i].get_URM()
test_urm.sort_indices()
mask = test_urm.data <= self.test_rating_threshold
for URM_name in folds[i].get_URM_names():
urm[URM_name] = folds[i].get_URM(URM_name)
urm[URM_name].sort_indices()
urm[URM_name].data[mask] = 0.0
urm[URM_name].eliminate_zeros()
test = Dataset(folds[i].get_name(), base_folder=folds[i].get_base_folder(),
postprocessings=folds[i].get_postprocessings(),
URM_dict=urm, URM_mappers_dict=folds[i].get_URM_mappers_dict(),
ICM_dict=folds[i].get_ICM_dict(), ICM_mappers_dict=folds[i].get_ICM_mappers_dict(),
UCM_dict=folds[i].get_UCM_dict(), UCM_mappers_dict=folds[i].get_UCM_mappers_dict())
if not self.allow_cold_users:
users_to_remove = np.arange(train.n_users)[np.ediff1d(train.get_URM().tocsr().indptr) <= 0]
train.remove_users(users_to_remove)
test.remove_users(users_to_remove)
r.append((train, test))
return r
def split(self, dataset):
super(WarmItemsKFold, self).split(dataset)
# I can do the kfold of a slice of the initial URM!
if self.percentage_initial_data_to_split < 1.0:
h = Holdout(train_perc=self.percentage_initial_data_to_split, test_perc=1-self.percentage_initial_data_to_split)
dataset = h.split(dataset)[0]
folds = []
URM = dataset.get_URM().tocoo()
split_belonging = np.random.choice(self.n_folds, URM.data.size, replace=True)
for i in range(self.n_folds):
urm = {}
urm_mappers = {}
mask = split_belonging == i
for URM_name in dataset.get_URM_names():
URM = dataset.get_URM(URM_name).tocoo()
# Sort nnz values by row and column indices, in order to remain consistent in the splits of different URMs
row, col, data = zip(*sorted(zip(URM.row, URM.col, URM.data), key=lambda x: (x[0], x[1])))
urm[URM_name] = sps.csr_matrix((np.array(data)[mask], (np.array(row)[mask], np.array(col)[mask])),
shape=URM.shape)
urm_mappers[URM_name] = dataset.get_URM_mapper(URM_name)
folds.append(
Dataset(dataset.get_name(), base_folder=dataset.get_base_folder(),
postprocessings=dataset.get_postprocessings(),
URM_dict=urm, URM_mappers_dict=urm_mappers,
ICM_dict=dataset.get_ICM_dict(), ICM_mappers_dict=dataset.get_ICM_mappers_dict(),
UCM_dict=dataset.get_UCM_dict(), UCM_mappers_dict=dataset.get_UCM_mappers_dict()
)
)
r = []
for i in range(self.n_folds):
urm = {}
urm_mappers = {}
for URM_name in folds[i].get_URM_names():
# Keep i-th fold as test and merge the others as train
urm[URM_name] = folds[(i + 1) % self.n_folds].get_URM(URM_name)
urm_mappers[URM_name] = folds[(i + 1) % self.n_folds].get_URM_mapper(URM_name)
for j in range(2, self.n_folds):
urm[URM_name] += folds[(i + j) % self.n_folds].get_URM(URM_name)
train = Dataset(folds[i].get_name(), base_folder=folds[i].get_base_folder(),
postprocessings=folds[i].get_postprocessings(),
URM_dict=urm, URM_mappers_dict=urm_mappers,
ICM_dict=folds[i].get_ICM_dict(), ICM_mappers_dict=folds[i].get_ICM_mappers_dict(),
UCM_dict=folds[i].get_UCM_dict(), UCM_mappers_dict=folds[i].get_UCM_mappers_dict())
urm = {}
test_urm = folds[i].get_URM()
test_urm.sort_indices()
mask = test_urm.data <= self.test_rating_threshold
for URM_name in folds[i].get_URM_names():
urm[URM_name] = folds[i].get_URM(URM_name)
urm[URM_name].sort_indices()
urm[URM_name].data[mask] = 0.0
urm[URM_name].eliminate_zeros()
test = Dataset(folds[i].get_name(), base_folder=folds[i].get_base_folder(),
postprocessings=folds[i].get_postprocessings(),
URM_dict=urm, URM_mappers_dict=folds[i].get_URM_mappers_dict(),
ICM_dict=folds[i].get_ICM_dict(), ICM_mappers_dict=folds[i].get_ICM_mappers_dict(),
UCM_dict=folds[i].get_UCM_dict(), UCM_mappers_dict=folds[i].get_UCM_mappers_dict())
if not self.allow_cold_users:
users_to_remove = np.arange(train.n_users)[np.ediff1d(train.get_URM().tocsr().indptr) <= 0]
train.remove_users(users_to_remove)
test.remove_users(users_to_remove)
r.append((train, test))
return r
