def find_threshold(self, data_attr, data_target):
'''
Find threshold of data_attr with respect to data_target
'''
attr_name = data_attr.columns[0]
target_name = data_target.columns[0]
data = pd.concat([data_attr, data_target], axis=1)
data = data.sort_values(attr_name).reset_index(drop=True)
# Retrieve all indexes with different value of target attribute
diff_index = list()
for i in range(len(data) - 1):
if data[target_name].iloc[i] != data[target_name].iloc[i + 1]:
diff_index.append(i)
best_point = 0
best_idx = -1
for i in diff_index:
if (self.gain_ratio):
point = Calculate.gain_ratio(data[attr_name],
data[target_name],
is_continue=True,
split_index=i)
else:
point = Calculate.info_gain(data[attr_name],
data[target_name],
is_continue=True,
split_index=i)
if point > best_point:
best_point = point
best_idx = i
best_splitter = (data[attr_name].iloc[best_idx] +
data[attr_name].iloc[best_idx + 1]) / 2
return [best_splitter, best_point]
def __fit_without_prune(self, data, features, target):
'''
Built entire decision tree without pruning
'''
continuous_features = list()
discrete_features = list()
for feature in features:
if len(list(data[feature])) > 0:
is_continue = self.is_attr_continue(list(data[feature]))
if is_continue:
continuous_features.append(feature)
else:
discrete_features.append(feature)
if not continuous_features:
return MyID3(self.gain_ratio).fit(data, features, target)
# Continuous attribute
# If only one value exist
entropy_data_target = Calculate.entropy(data[target])
if entropy_data_target == 0:
value_list = Calculate.get_unique_data(data, target)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
return Tree(
Node(
None,
0.0, # Entropy must be 0 since only one value exist
value_dict,
result=data[target][0],
is_leaf=True))
if (len(features) == 0):
value_list = Calculate.get_unique_data(data, target)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
return Tree(
Node(None,
entropy_data_target,
value_dict,
result=Calculate.most_label(data[target]),
is_leaf=True))
# Find best attribute and build tree recursively
best_attr = ''
best_point = 0
is_discrete = False
best_splitter = 0
chosen_edge = list(['', ''])
for feature in continuous_features:
best_treshold = self.find_threshold(data[[feature]],
data[[target]])
if best_treshold[1] > best_point:
best_attr = str(feature)
chosen_edge[0] = best_attr + ' > ' + str(best_treshold[0])
chosen_edge[1] = best_attr + ' <= ' + str(best_treshold[0])
best_point = best_treshold[1]
best_splitter = best_treshold[0]
for feature in discrete_features:
point = Calculate.info_gain(data[feature], data[target])
if point > best_point:
best_point = point
best_attr = str(feature)
is_discrete = True
value_list = Calculate.get_unique_data(data, target)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
dtree = Tree(Node(best_attr, best_point, value_dict))
# Scan all posible value to be generated subtree
if is_discrete:
list_attribute = Calculate.get_unique_data(data, best_attr)
else:
list_attribute = Calculate.split_by_threshold(
data, best_attr, best_splitter)
i = 0
for attribute in list_attribute:
data = pd.DataFrame(data=list_attribute[attribute]).reset_index(
drop=True)
dtree.add_child(self.__fit_without_prune(data, features, target))
if is_discrete:
dtree.children[i].value.edge = attribute
else:
dtree.children[i].value.edge = chosen_edge[i]
i += 1
return dtree
def fit(self, data, attributes, target_name):
'''
Built and return decision tree using ID3 algorithm
'''
data_target = data[target_name]
# Data target contains one label
entropy_data_target = Calculate.entropy(data_target)
if entropy_data_target == 0:
value_list = Calculate.get_unique_data(data, target_name)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
# Set current_node, info_gain, values
tree = Tree(
Node(None,
entropy_data_target,
value_dict,
result=data_target[0],
is_leaf=True))
return tree
# Nothing attribute shall be chosen
if len(attributes) == 0:
# Set current_node, info_gain, values
value_list = Calculate.get_unique_data(data, target_name)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
tree = Tree(
Node(None,
entropy_data_target,
value_dict,
result=Calculate.most_label(data_target),
is_leaf=True))
return tree
else:
# Find best attribute to be node using either info gain or gain ratio
best_attr = ''
best_point = 0 # Could be Info gain or Gain ratio
for attr in attributes:
if self.gain_ratio:
point = Calculate.gain_ratio(data[attr], data_target)
if point > best_point:
best_point = point
best_attr = attr
else:
point = Calculate.info_gain(data[attr], data_target)
if point > best_point:
best_point = point
best_attr = attr
value_list = Calculate.get_unique_data(data, target_name)
value_dict = dict()
for key, value in value_list.items():
value_dict[key] = len(value_list[key])
# Build decision tree recursively
dtree = Tree(Node(best_attr, best_point, value_dict))
# Delete usage attribute in attributes
attributes.remove(best_attr)
# Scan all posible value to be generated subtree
list_attribute = Calculate.get_unique_data(data, best_attr)
i = 0
for attribute in list_attribute:
data = pd.DataFrame(
data=list_attribute[attribute]).reset_index(drop=True)
data.drop(best_attr, axis=1, inplace=True)
dtree.add_child(self.fit(data, attributes, target_name))
dtree.children[i].value.edge = attribute
i += 1
return dtree
