def fit_storage(self, data):
if self.binarize and any(
attr.is_discrete and len(attr.values) > self.MAX_BINARIZATION
for attr in data.domain.attributes):
# No fallback in the script; widgets can prevent this error
# by providing a fallback and issue a warning about doing so
raise ValueError("Exhaustive binarization does not handle "
"attributes with more than {} values".
format(self.MAX_BINARIZATION))
active_inst = np.nonzero(~np.isnan(data.Y))[0].astype(np.int32)
root = self.build_tree(data, active_inst)
if root is None:
root = Node(None, 0, np.array([0., 0.]))
root.subset = active_inst
model = TreeModel(data, root)
return model
def fit_storage(self, data):
if self.binarize and any(
attr.is_discrete and len(attr.values) > self.MAX_BINARIZATION
for attr in data.domain.attributes):
# No fallback in the script; widgets can prevent this error
# by providing a fallback and issue a warning about doing so
raise ValueError("Exhaustive binarization does not handle "
"attributes with more than {} values".format(
self.MAX_BINARIZATION))
active_inst = np.nonzero(~np.isnan(data.Y))[0].astype(np.int32)
root = self.build_tree(data, active_inst)
if root is None:
distr = distribution.Discrete(data, data.domain.class_var)
if np.sum(distr) == 0:
distr[:] = 1
root = Node(None, 0, distr)
root.subset = active_inst
model = TreeModel(data, root)
return model
def build_tree(self, data, active_inst, level=1):
"""Induce a tree from the given data
Returns:
root node (Node)"""
node_insts = data[active_inst]
if len(node_insts) < self.min_samples_leaf:
return None
if len(node_insts) < self.min_samples_split or \
self.max_depth is not None and level > self.max_depth:
node, branches, n_children = Node(None, None, None), None, 0
else:
node, branches, n_children = self._select_attr(node_insts)
mean, var = np.mean(node_insts.Y), np.var(node_insts.Y)
node.value = np.array([mean, 1 if np.isnan(var) else var])
node.subset = active_inst
if branches is not None:
node.children = [
self.build_tree(data, active_inst[branches == br], level + 1)
for br in range(n_children)]
return node
def _build_tree(self, data, active_inst, level=1):
"""Induce a tree from the given data
Returns:
root node (Node)"""
node_insts = data[active_inst]
distr = distribution.Discrete(node_insts, data.domain.class_var)
if len(node_insts) < self.min_samples_leaf:
return None
if len(node_insts) < self.min_samples_split or \
max(distr) >= sum(distr) * self.sufficient_majority or \
self.max_depth is not None and level > self.max_depth:
node, branches, n_children = Node(None, None, distr), None, 0
else:
node, branches, n_children = self._select_attr(node_insts)
node.subset = active_inst
if branches is not None:
node.children = [
self._build_tree(data, active_inst[branches == br], level + 1)
for br in range(n_children)]
return node
def build_tree(self, data, active_inst, level=1):
"""Induce a tree from the given data
Returns:
root node (Node)"""
node_insts = data[active_inst]
if len(node_insts) < self.min_samples_leaf:
return None
if len(node_insts) < self.min_samples_split or \
self.max_depth is not None and level > self.max_depth:
node, branches, n_children = Node(None, None, None), None, 0
else:
node, branches, n_children = self._select_attr(node_insts)
mean, var = np.mean(node_insts.Y), np.var(node_insts.Y)
node.value = np.array([mean, 1 if np.isnan(var) else var])
node.subset = active_inst
if branches is not None:
node.children = [
self.build_tree(data, active_inst[branches == br], level + 1)
for br in range(n_children)]
return node
def build_tree(self, data, active_inst, level=1):
"""Induce a tree from the given data
Returns:
root node (Node)"""
node_insts = data[active_inst]
distr = distribution.Discrete(node_insts, data.domain.class_var)
if len(node_insts) < self.min_samples_leaf:
return None
if len(node_insts) < self.min_samples_split or \
max(distr) >= sum(distr) * self.sufficient_majority or \
self.max_depth is not None and level > self.max_depth:
node, branches, n_children = Node(None, None, distr), None, 0
else:
node, branches, n_children = self._select_attr(node_insts)
node.subset = active_inst
if branches is not None:
node.children = [
self.build_tree(data, active_inst[branches == br], level + 1)
for br in range(n_children)]
return node
def _select_attr(self, data):
"""Select the attribute for the next split.
Returns
-------
tuple with an instance of Node and a numpy array indicating
the branch index for each data instance, or -1 if data instance
is dropped
"""
# Prevent false warnings by pylint
attr = attr_no = None
REJECT_ATTRIBUTE = 0, None, None, 0
def _score_disc():
n_values = len(attr.values)
score = _tree_scorers.compute_grouped_MSE(col_x, col_y, n_values,
self.min_samples_leaf)
# The score is already adjusted for missing attribute values, so
# we don't do it here
if score == 0:
return REJECT_ATTRIBUTE
branches = col_x.flatten()
branches[np.isnan(branches)] = -1
return score, DiscreteNode(attr, attr_no, None), branches, n_values
def _score_disc_bin():
n_values = len(attr.values)
if n_values == 2:
return _score_disc()
score, mapping = _tree_scorers.find_binarization_MSE(
col_x, col_y, n_values, self.min_samples_leaf)
# The score is already adjusted for missing attribute values, so
# we don't do it here
if score == 0:
return REJECT_ATTRIBUTE
mapping, branches = MappedDiscreteNode.branches_from_mapping(
data.X[:, attr_no], mapping, len(attr.values))
node = MappedDiscreteNode(attr, attr_no, mapping, None)
return score, node, branches, 2
def _score_cont():
"""Scoring for numeric attributes"""
nans = np.sum(np.isnan(col_x))
non_nans = len(col_x) - nans
arginds = np.argsort(col_x)[:non_nans]
score, cut = _tree_scorers.find_threshold_MSE(
col_x, col_y, arginds, self.min_samples_leaf)
if score == 0:
return REJECT_ATTRIBUTE
score *= non_nans / len(col_x)
branches = np.full(len(col_x), -1, dtype=int)
mask = ~np.isnan(col_x)
branches[mask] = (col_x[mask] > cut).astype(int)
node = NumericNode(attr, attr_no, cut, None)
return score, node, branches, 2
#######################################
# The real _select_attr starts here
domain = data.domain
col_y = data.Y
best_score, *best_res = REJECT_ATTRIBUTE
best_res = [
Node(None, 0, None),
] + best_res[1:]
disc_scorer = _score_disc_bin if self.binarize else _score_disc
for attr_no, attr in enumerate(domain.attributes):
col_x = data[:, attr_no].X.reshape((len(data), ))
sc, *res = disc_scorer() if attr.is_discrete else _score_cont()
if res[0] is not None and sc > best_score:
best_score, best_res = sc, res
return best_res
def _select_attr(self, data):
"""Select the attribute for the next split.
Returns:
tuple with an instance of Node and a numpy array indicating
the branch index for each data instance, or -1 if data instance
is dropped
"""
# Prevent false warnings by pylint
attr = attr_no = None
col_x = None
REJECT_ATTRIBUTE = 0, None, None, 0
def _score_disc():
"""Scoring for discrete attributes, no binarization
The class computes the entropy itself, not by calling other
functions. This is to make sure that it uses the same
definition as the below classes that compute entropy themselves
for efficiency reasons."""
n_values = len(attr.values)
if n_values < 2:
return REJECT_ATTRIBUTE
cont = _tree_scorers.contingency(col_x, len(data.domain.attributes[attr_no].values),
data.Y, len(data.domain.class_var.values))
attr_distr = np.sum(cont, axis=0)
null_nodes = attr_distr < self.min_samples_leaf
# This is just for speed. If there is only a single non-null-node,
# entropy wouldn't decrease anyway.
if sum(null_nodes) >= n_values - 1:
return REJECT_ATTRIBUTE
cont[:, null_nodes] = 0
attr_distr = np.sum(cont, axis=0)
cls_distr = np.sum(cont, axis=1)
n = np.sum(attr_distr)
# Avoid log(0); <= instead of == because we need an array
cls_distr[cls_distr <= 0] = 1
attr_distr[attr_distr <= 0] = 1
cont[cont <= 0] = 1
class_entr = n * np.log(n) - np.sum(cls_distr * np.log(cls_distr))
attr_entr = np.sum(attr_distr * np.log(attr_distr))
cont_entr = np.sum(cont * np.log(cont))
score = (class_entr - attr_entr + cont_entr) / n / np.log(2)
score *= n / len(data) # punishment for missing values
branches = col_x.copy()
branches[np.isnan(branches)] = -1
if score == 0:
return REJECT_ATTRIBUTE
node = DiscreteNode(attr, attr_no, None)
return score, node, branches, n_values
def _score_disc_bin():
"""Scoring for discrete attributes, with binarization"""
n_values = len(attr.values)
if n_values <= 2:
return _score_disc()
cont = contingency.Discrete(data, attr)
attr_distr = np.sum(cont, axis=0)
# Skip instances with missing value of the attribute
cls_distr = np.sum(cont, axis=1)
if np.sum(attr_distr) == 0: # all values are missing
return REJECT_ATTRIBUTE
best_score, best_mapping = _tree_scorers.find_binarization_entropy(
cont, cls_distr, attr_distr, self.min_samples_leaf)
if best_score <= 0:
return REJECT_ATTRIBUTE
best_score *= 1 - np.sum(cont.unknowns) / len(data)
mapping, branches = MappedDiscreteNode.branches_from_mapping(
col_x, best_mapping, n_values)
node = MappedDiscreteNode(attr, attr_no, mapping, None)
return best_score, node, branches, 2
def _score_cont():
"""Scoring for numeric attributes"""
nans = np.sum(np.isnan(col_x))
non_nans = len(col_x) - nans
arginds = np.argsort(col_x)[:non_nans]
best_score, best_cut = _tree_scorers.find_threshold_entropy(
col_x, data.Y, arginds,
len(class_var.values), self.min_samples_leaf)
if best_score == 0:
return REJECT_ATTRIBUTE
best_score *= non_nans / len(col_x)
branches = np.full(len(col_x), -1, dtype=int)
mask = ~np.isnan(col_x)
branches[mask] = (col_x[mask] > best_cut).astype(int)
node = NumericNode(attr, attr_no, best_cut, None)
return best_score, node, branches, 2
#######################################
# The real _select_attr starts here
is_sparse = sp.issparse(data.X)
domain = data.domain
class_var = domain.class_var
best_score, *best_res = REJECT_ATTRIBUTE
best_res = [Node(None, None, None)] + best_res[1:]
disc_scorer = _score_disc_bin if self.binarize else _score_disc
for attr_no, attr in enumerate(domain.attributes):
col_x = data.X[:, attr_no]
if is_sparse:
col_x = col_x.toarray()
col_x = col_x.flatten()
sc, *res = disc_scorer() if attr.is_discrete else _score_cont()
if res[0] is not None and sc > best_score:
best_score, best_res = sc, res
best_res[0].value = distribution.Discrete(data, class_var)
return best_res
def setUpClass(cls):
super().setUpClass()
WidgetOutputsTestMixin.init(cls)
tree = TreeLearner()
cls.model = tree(cls.data)
cls.model.instances = cls.data
cls.signal_name = "Tree"
cls.signal_data = cls.model
# Load a dataset that contains two variables with the same entropy
data_same_entropy = Table(
path.join(path.dirname(path.dirname(path.dirname(__file__))),
"tests", "datasets", "same_entropy.tab"))
cls.data_same_entropy = tree(data_same_entropy)
cls.data_same_entropy.instances = data_same_entropy
vara = DiscreteVariable("aaa", values=("e", "f", "g"))
root = DiscreteNode(vara, 0, np.array([42, 8]))
root.subset = np.arange(50)
varb = DiscreteVariable("bbb", values=tuple("ijkl"))
child0 = MappedDiscreteNode(varb, 1, np.array([0, 1, 0, 0]), (38, 5))
child0.subset = np.arange(16)
child1 = Node(None, 0, (13, 3))
child1.subset = np.arange(16, 30)
varc = ContinuousVariable("ccc")
child2 = NumericNode(varc, 2, 42, (78, 12))
child2.subset = np.arange(30, 50)
root.children = (child0, child1, child2)
child00 = Node(None, 0, (15, 4))
child00.subset = np.arange(10)
child01 = Node(None, 0, (10, 5))
child01.subset = np.arange(10, 16)
child0.children = (child00, child01)
child20 = Node(None, 0, (90, 4))
child20.subset = np.arange(30, 35)
child21 = Node(None, 0, (70, 9))
child21.subset = np.arange(35, 50)
child2.children = (child20, child21)
domain = Domain([vara, varb, varc], ContinuousVariable("y"))
t = [[i, j, k] for i in range(3) for j in range(4) for k in (40, 44)]
x = np.array((t * 3)[:50])
data = Table.from_numpy(domain, x, np.arange(len(x)))
cls.tree = TreeModel(data, root)