"""Tree inducers: SKL and Orange's own inducer"""
import numpy as np
import scipy.sparse as sp
import sklearn.tree as skl_tree
from Orange.base import TreeModel as TreeModelInterface
from Orange.classification import SklLearner, SklModel, Learner
from Orange.classification import _tree_scorers
from Orange.statistics import distribution, contingency
from Orange.tree import Node, DiscreteNode, MappedDiscreteNode, \
NumericNode, TreeModel
__all__ = ["SklTreeLearner", "TreeLearner"]
[docs]
class TreeLearner(Learner):
"""
Tree inducer with proper handling of nominal attributes and binarization.
The inducer can handle missing values of attributes and target.
For discrete attributes with more than two possible values, each value can
get a separate branch (`binarize=False`), or values can be grouped into
two groups (`binarize=True`, default).
The tree growth can be limited by the required number of instances for
internal nodes and for leafs, the sufficient proportion of majority class,
and by the maximal depth of the tree.
If the tree is not binary, it can contain zero-branches.
Args:
binarize (bool):
if `True` the inducer will find optimal split into two
subsets for values of discrete attributes. If `False` (default),
each value gets its branch.
min_samples_leaf (float):
the minimal number of data instances in a leaf
min_samples_split (float):
the minimal nubmer of data instances that is
split into subgroups
max_depth (int): the maximal depth of the tree
sufficient_majority (float):
a majority at which the data is not split
further
Returns:
instance of OrangeTreeModel
"""
__returns__ = TreeModel
# Binarization is exhaustive, so we set a limit on the number of values
MAX_BINARIZATION = 16
def __init__(
self, *args, binarize=False, max_depth=None,
min_samples_leaf=1, min_samples_split=2, sufficient_majority=0.95,
preprocessors=None, **kwargs):
super().__init__(preprocessors=preprocessors)
self.params = {}
self.binarize = self.params['binarize'] = binarize
self.min_samples_leaf = self.params['min_samples_leaf'] = min_samples_leaf
self.min_samples_split = self.params['min_samples_split'] = min_samples_split
self.sufficient_majority = self.params['sufficient_majority'] = sufficient_majority
self.max_depth = self.params['max_depth'] = max_depth
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 _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
[docs]
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
class SklTreeClassifier(SklModel, TreeModelInterface):
"""Wrapper for SKL's tree classifier with the interface API for
visualizations"""
def __init__(self, *args, **kwargs):
SklModel.__init__(self, *args, **kwargs)
self._cached_sample_assignments = None
[docs]
class SklTreeLearner(SklLearner):
"""Wrapper for SKL's tree inducer"""
__wraps__ = skl_tree.DecisionTreeClassifier
__returns__ = SklTreeClassifier
name = 'tree'
supports_weights = True
def __init__(self, criterion="gini", splitter="best", max_depth=None,
min_samples_split=2, min_samples_leaf=1,
max_features=None,
random_state=None, max_leaf_nodes=None,
preprocessors=None):
super().__init__(preprocessors=preprocessors)
self.params = vars()