def test_forward_select(self):
path = datasetsDir(self) + 'minigolf' + SEP + 'weather'
_training = training(path)
_attributes, _klass = metadata(path)
_test = test(path)
_gold = gold(path)
verify_training = copy.deepcopy(_training)
verify_attributes = copy.deepcopy(_attributes)
feat_sel = fs.FeatureSelection(_training, _attributes, _klass, _test,
_gold, ['1R', '4', '0.1'])
feat_sel.forward_selection()
self.assertEqual(1, len(_attributes))
self.assertEqual('outlook', _attributes[0].name)
self.verify_number_of_attributes(_training, 1)
self.verify_number_of_attributes(_test, 1)
self.verify_number_of_attributes(_gold, 1)
#verification
verification_cv_datasets = verify_training.cross_validation_datasets(4)
accuracies = {}
for attribute in verify_attributes:
accuracies[
attribute.name] = feat_sel.avg_accuracy_by_cross_validation(
verification_cv_datasets, 4, attr.Attributes([attribute]))
#'windy': 0.41666666666666663, 'outlook': 0.79166666666666663, 'temperature': 0.41666666666666663, 'humidity': 0.54166666666666663
self.assertAlmostEqual(0.4166666, accuracies['windy'], 6)
self.assertAlmostEqual(0.79166666, accuracies['outlook'], 6)
self.assertAlmostEqual(0.4166666, accuracies['temperature'], 6)
self.assertAlmostEqual(0.5416666, accuracies['humidity'], 6)
#outlook selected
accuracies = {}
for each in verify_attributes:
if each.name == 'outlook':
outlook = each
verify_attributes.remove(outlook)
for attribute in verify_attributes:
accuracies[(
'outlook',
attribute.name)] = feat_sel.avg_accuracy_by_cross_validation(
verification_cv_datasets, 4,
attr.Attributes([outlook, attribute]))
#{('outlook', 'humidity'): 0.79166666666666663, ('outlook', 'temperature'): 0.79166666666666663, ('outlook', 'windy'): 0.54166666666666663}
self.assertAlmostEqual(0.7916666, accuracies[('outlook', 'humidity')],
6)
self.assertAlmostEqual(0.7916666, accuracies['outlook', 'temperature'],
6)
self.assertAlmostEqual(0.5416666, accuracies[('outlook', 'windy')], 6)
def test_attributes_are_equal(self):
attrs = a.Attributes([
a.Attribute('band', ['dual', 'tri', 'quad'], 0),
a.Attribute('size', ['big', 'small', 'medium'], 1)
])
same = a.Attributes([
a.Attribute('band', ['dual', 'tri', 'quad'], 0),
a.Attribute('size', ['big', 'small', 'medium'], 1)
])
self.assertEqual(attrs, same, 'they should be the same')
other = a.Attributes([
a.Attribute('band', ['dual', 'tri', 'quad'], 0),
a.Attribute('pda', ['y', 'n'], 1)
])
self.assertNotEqual(self.attrs, other, 'shouldnt be the same')
def test_empty_freq_dists(self):
attr1 = a.Attribute("first", ['a', 'b', 'c'], 0)
attr2 = a.Attribute("second", ['d', 'e'], 1)
attrs = a.Attributes([attr1, attr2])
freq_dists = attrs.empty_freq_dists()
self.assertEqual(2, len(freq_dists))
self.assertEqual(3, len(freq_dists[attr1]))
self.assertEqual(2, len(freq_dists[attr2]))
def metadata(self, file_path):
lines = self.__get_lines(file_path, self.NAMES)
klass_values = item.NameItem(lines[0]).processed().split(',')
index,attributes = 0, []
for line in lines:
nameitem = item.NameItem(line)
processed = nameitem.processed()
if not len(processed) == 0 and nameitem.isAttribute():
attributes.append(a.Attribute(self.get_name(processed), self.get_values(processed), index))
index += 1
return (a.Attributes(attributes), klass_values)
def __select_attributes(self, max, selected, others, delta):
if others is None or len(others) == 0: return selected
max_at_level, attr_with_max_acc, fold = -1, None, self.get_fold()
datasets = self.training.cross_validation_datasets(fold)
for attribute in others:
selected.append(attribute)
avg_accuracy = self.avg_accuracy_by_cross_validation(
datasets, fold, attr.Attributes(selected))
if avg_accuracy > max_at_level:
max_at_level = avg_accuracy
attr_with_max_acc = attribute
selected.remove(attribute)
if max_at_level - max < delta: return selected
selected.append(attr_with_max_acc)
others.remove(attr_with_max_acc)
return self.__select_attributes(max_at_level, selected, others, delta)
def __eliminate_attributes(self, max, selected, delta):
if selected is None or len(selected) == 0 or len(selected) == 1:
return selected
max_at_level, selections_with_max_acc, fold = -1, None, self.get_fold()
datasets = self.training.cross_validation_datasets(fold)
selected_for_iter = selected[:]
for attribute in selected_for_iter:
selected.remove(attribute)
avg_accuracy = self.avg_accuracy_by_cross_validation(
datasets, fold, attr.Attributes(selected))
if avg_accuracy > max_at_level:
max_at_level = avg_accuracy
selections_with_max_acc = selected[:]
selected.append(attribute)
if max_at_level - max < delta: return selected
return self.__eliminate_attributes(max_at_level,
selections_with_max_acc, delta)
def test_backward_select(self):
path = datasetsDir(self) + 'minigolf' + SEP + 'weather'
_training = training(path)
_attributes, _klass = metadata(path)
_test = test(path)
_gold = gold(path)
verify_training = copy.deepcopy(_training)
verify_attributes = copy.deepcopy(_attributes)
feat_sel = fs.FeatureSelection(_training, _attributes, _klass, _test,
_gold, ['1R', '4', '0.1'])
feat_sel.backward_elimination()
self.assertEqual(3, len(_attributes))
self.verify_number_of_attributes(_training, 3)
self.verify_number_of_attributes(_test, 3)
self.verify_number_of_attributes(_gold, 3)
#verification
#level 0
avg_acc = feat_sel.avg_accuracy_by_cross_validation(
verify_training.cross_validation_datasets(4), 4, verify_attributes)
self.assertAlmostEqual(0.5416666, avg_acc, 6)
verification_cv_datasets = verify_training.cross_validation_datasets(4)
accuracies = {}
for attribute in verify_attributes:
attributes = verify_attributes[:]
attributes.remove(attribute)
accuracies[(attributes[0].name, attributes[1].name,
attributes[2].name
)] = feat_sel.avg_accuracy_by_cross_validation(
verification_cv_datasets, 4,
attr.Attributes(attributes))
# {('outlook', 'humidity', 'windy'): 0.54166666666666663,
# ('outlook', 'temperature', 'windy'): 0.54166666666666663,
# ('temperature', 'humidity', 'windy'): 0.29166666666666663,
# ('outlook', 'temperature', 'humidity'): 0.79166666666666663}
self.assertAlmostEqual(0.5416666,
accuracies[('outlook', 'humidity', 'windy')], 6)
self.assertAlmostEqual(0.5416666,
accuracies[('outlook', 'temperature', 'windy')],
6)
self.assertAlmostEqual(
0.2916666, accuracies[('temperature', 'humidity', 'windy')], 6)
self.assertAlmostEqual(
0.7916666, accuracies[('outlook', 'temperature', 'humidity')], 6)
#
#('outlook', 'temperature', 'humidity') selected
accuracies = {}
for each in verify_attributes:
if each.name == 'windy':
windy = each
verify_attributes.remove(windy)
for attribute in verify_attributes:
attributes = verify_attributes[:]
attributes.remove(attribute)
accuracies[(attributes[0].name, attributes[1].name
)] = feat_sel.avg_accuracy_by_cross_validation(
verification_cv_datasets, 4,
attr.Attributes(attributes))
self.assertAlmostEqual(0.7916666, accuracies[('outlook', 'humidity')],
6)
self.assertAlmostEqual(0.7916666, accuracies['outlook', 'temperature'],
6)
self.assertAlmostEqual(0.4166666,
accuracies[('temperature', 'humidity')], 6)
def test_to_string(self):
attr1 = a.Attribute("first", ['a', 'b', 'c'], 0)
attr2 = a.Attribute("second", ['d', 'e'], 1)
attrs = a.Attributes([attr1, attr2])
self.assertEqual('[first:[a,b,c] index:0, second:[d,e] index:1]',
str(attrs))
