def split_superinstance(self, si, k):
# the actual splitting
pred = kshape(self.data[si.indices, :], k)
# making sure that all of the super-instances contain at least one training instance
# super-instances without training instance are merged with the closest one that does contain a
# training instance
training = []
no_training = []
for new_si_centroid, new_si_idx in pred:
# go from super instance indices to global ones
cur_indices = [si.indices[idx] for idx in new_si_idx]
si_train_indices = [x for x in cur_indices if x in self.train_indices]
if len(si_train_indices) != 0:
training.append(SuperInstance_kShape(self.data, cur_indices, self.train_indices, new_si_centroid, si))
else:
no_training.append((cur_indices, new_si_centroid))
for indices, centroid in no_training:
# sets of indices without a training point are merged with their closest super-instance
# closeness is based on the SBD centroid
closest_train = None
closest_train_dist = np.inf
for training_si in training:
cur_dist, _ = _sbd(training_si.sbd_centroid, centroid)
if cur_dist < closest_train_dist:
closest_train_dist = cur_dist
closest_train = training_si
closest_train.indices.extend(indices)
si.children = training
return training
def get_prototype(A, indices, prototype):
max_dist_to_others = np.inf
prototype_idx = None
for idx in indices:
cur_dist, _ = _sbd(A[idx, :], prototype)
if cur_dist < max_dist_to_others:
max_dist_to_others = cur_dist
prototype_idx = idx
return prototype_idx
def distance_between_instances(x1, x2):
d, _ = _sbd(x1, x2)
return d
def distance_to(self, other_superinstance):
d, _ = _sbd(self.sbd_centroid, other_superinstance.sbd_centroid)
return d