def min_entropy_split(dataset):
#print 'in min_entropy_split'
# Assume inputs are continuous, and are column vectors.
hypotheses = []
entropies = []
# For each attribute find the best split point.
for attribute in xrange(dataset.num_attributes()):
values = ds.unique_values(dataset.inputs, attribute)
#Iterate over the possible values of split & calculate entropy for each split.
for split_point in values:
def calc_entropy(data_set):
num_points = data_set.num_examples()
#return (num_points / float(dataset.num_examples())) * data_set.entropy_discrete()
return (num_points / float(dataset.num_examples())
) * ds.dataset_entropy_discrete(dataset)
split_entropy = map(
calc_entropy,
ds.split_continuous(dataset, attribute, split_point))
hypotheses.append((attribute, split_point))
entropies.append(sum(split_entropy))
# Select the attribute split pair that has the lowest entropy.
entropies = np.matrix(entropies)
min_idx = np.argmin(entropies)
return hypotheses[min_idx]
def train(self, dataset, splitting_func=min_entropy_split):
if not self.make_leaf(dataset):
#print 'in train.splitting', dataset.num_examples()
self.split_attribute, self.split_point = splitting_func(dataset)
#print 'self.split_attribute, self.split_point', self.split_attribute, self.split_point
data_sets = ds.split_continuous(dataset, self.split_attribute, self.split_point)
if len(data_sets) < 2:
self.prediction = dataset.outputs
return
def tree_split(set):
#print 'tree', set.num_examples()
return DecisionTree(set, splitting_func=splitting_func)
# Create & train child decision nodes
tests = create_binary_tests(self.split_attribute, self.split_point)
self.children = zip(tests, map(tree_split, data_sets))
def train(self, dataset, splitting_func=min_entropy_split):
if not self.make_leaf(dataset):
#print 'in train.splitting', dataset.num_examples()
self.split_attribute, self.split_point = splitting_func(dataset)
#print 'self.split_attribute, self.split_point', self.split_attribute, self.split_point
data_sets = ds.split_continuous(dataset, self.split_attribute,
self.split_point)
if len(data_sets) < 2:
self.prediction = dataset.outputs
return
def tree_split(set):
#print 'tree', set.num_examples()
return DecisionTree(set, splitting_func=splitting_func)
# Create & train child decision nodes
tests = create_binary_tests(self.split_attribute, self.split_point)
self.children = zip(tests, map(tree_split, data_sets))
def min_entropy_split(dataset):
#print 'in min_entropy_split'
# Assume inputs are continuous, and are column vectors.
hypotheses = []
entropies = []
# For each attribute find the best split point.
for attribute in xrange(dataset.num_attributes()):
values = ds.unique_values(dataset.inputs, attribute)
#Iterate over the possible values of split & calculate entropy for each split.
for split_point in values:
def calc_entropy(data_set):
num_points = data_set.num_examples()
#return (num_points / float(dataset.num_examples())) * data_set.entropy_discrete()
return (num_points / float(dataset.num_examples())) * ds.dataset_entropy_discrete(dataset)
split_entropy = map(calc_entropy, ds.split_continuous(dataset, attribute, split_point))
hypotheses.append((attribute, split_point))
entropies.append(sum(split_entropy))
# Select the attribute split pair that has the lowest entropy.
entropies = np.matrix(entropies)
min_idx = np.argmin(entropies)
return hypotheses[min_idx]
