class BaggingSK(PoolGenerator):
""""
This class should not be used, use brew.generation.bagging.Bagging instead.
"""
def __init__(self,
base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote'):
self.base_classifier = base_classifier
self.n_classifiers = n_classifiers
# using the sklearn implementation of bagging for now
self.sk_bagging = BaggingClassifier(base_estimator=base_classifier,
n_estimators=n_classifiers,
max_samples=1.0,
max_features=1.0)
self.ensemble = Ensemble()
self.combiner = Combiner(rule=combination_rule)
def fit(self, X, y):
self.sk_bagging.fit(X, y)
self.ensemble.add_classifiers(self.sk_bagging.estimators_)
# self.classes_ = set(y)
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class Bagging(PoolGenerator):
def __init__(self,
base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote'):
self.base_classifier = base_classifier
self.n_classifiers = n_classifiers
self.ensemble = None
self.combiner = Combiner(rule=combination_rule)
def fit(self, X, y):
self.ensemble = Ensemble()
for _ in range(self.n_classifiers):
# bootstrap
idx = np.random.choice(X.shape[0], X.shape[0], replace=True)
data, target = X[idx, :], y[idx]
classifier = sklearn.base.clone(self.base_classifier)
classifier.fit(data, target)
self.ensemble.add(classifier)
return
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class RandomSubspace(PoolGenerator):
def __init__(self,
base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote',
max_features=0.5):
self.base_classifier = base_classifier
self.n_classifiers = n_classifiers
self.combiner = Combiner(rule=combination_rule)
self.classifiers = None
self.ensemble = None
self.max_features = max_features
def fit(self, X, y):
self.ensemble = Ensemble()
for i in range(self.n_classifiers):
chosen_features = np.random.choice(
X.shape[1],
int(np.ceil(X.shape[1] * self.max_features)),
replace=False)
transformer = FeatureSubsamplingTransformer(
features=chosen_features)
classifier = BrewClassifier(classifier=sklearn.base.clone(
self.base_classifier),
transformer=transformer)
classifier.fit(X, y)
self.ensemble.add(classifier)
return
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class ICSBagging(PoolGenerator):
def __init__(self,
K=10,
alpha=0.75,
base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote',
diversity_metric='e',
positive_label=1):
self.K = K
self.alpha = alpha
self.base_classifier = base_classifier
self.n_classifiers = n_classifiers
self.combination_rule = combination_rule
self.positive_label = positive_label
self.classifiers = None
self.ensemble = None
self.combiner = Combiner(rule=combination_rule)
self.diversity_metric = diversity_metric
self.diversity = Diversity(metric=diversity_metric)
self.validation_X = None
self.validation_y = None
def set_validation(self, X, y):
self.validation_X = X
self.validation_y = y
def fitness(self, classifier):
'''
# TODO normalize diversity metric.
'''
self.ensemble.add(classifier)
out = self.ensemble.output(self.validation_X)
y_pred = self.combiner.combine(out)
y_true = self.validation_y
auc = evaluation.auc_score(y_true, y_pred)
div = self.diversity.calculate(self.ensemble, self.validation_X,
self.validation_y)
# diversity = entropy_measure_e(self.ensemble,
# self.validation_X, self.validation_y)
self.ensemble.classifiers.pop()
return self.alpha * auc + (1.0 - self.alpha) * div
def _calc_pos_prob(self):
y_pred = self.combiner.combine(self.ensemble.output(self.validation_X))
mask = self.positive_label == self.validation_y
pos_acc = float(sum(y_pred[mask] == self.validation_y[mask])) / len(
self.validation_y[mask])
neg_acc = float(sum(y_pred[~mask] == self.validation_y[~mask])) / len(
self.validation_y[~mask])
return 1.0 - (pos_acc / (pos_acc + neg_acc))
def bootstrap_classifiers(self, X, y, K, pos_prob):
mask = self.positive_label == y
negative_label = y[~mask][0]
clfs = []
sets_cX, sets_cy = [], []
for i in range(K):
cX, cy = [], []
for j in range(X.shape[0]):
if np.random.random() < pos_prob:
idx = np.random.random_integers(0, len(X[mask]) - 1)
cX = cX + [X[mask][idx]]
cy = cy + [self.positive_label]
else:
idx = np.random.random_integers(0, len(X[~mask]) - 1)
cX = cX + [X[~mask][idx]]
cy = cy + [negative_label]
if self.positive_label not in cy:
idx_1 = np.random.random_integers(0, len(cX) - 1)
idx_2 = np.random.random_integers(0, len(X[mask]) - 1)
cX[idx_1] = X[mask][idx_2]
cy[idx_1] = self.positive_label
elif negative_label not in cy:
idx_1 = np.random.random_integers(0, len(cX) - 1)
idx_2 = np.random.random_integers(0, len(X[~mask]) - 1)
cX[idx_1] = X[~mask][idx_2]
cy[idx_1] = negative_label
# print len(cX), len(cy), X.shape[0], len(X), np.bincount(cy)
sets_cX, sets_cy = sets_cX + [cX], sets_cy + [cy]
clf = sklearn.base.clone(self.base_classifier)
clfs = clfs + [clf.fit(cX, cy)]
return clfs
def fit(self, X, y):
# if self.validation_X == None and self.validation_y == None:
self.validation_X = X
self.validation_y = y
self.classes_ = set(y)
self.ensemble = Ensemble()
clfs = self.bootstrap_classifiers(X, y, self.K, 0.5)
self.ensemble.add(np.random.choice(clfs))
for _ in range(1, self.n_classifiers):
clfs = self.bootstrap_classifiers(X, y, self.K,
self._calc_pos_prob())
self.ensemble.add(max(clfs, key=lambda clf: self.fitness(clf)))
self.validation_X = None
self.validation_y = None
return self
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class ICSBaggingNew(PoolGenerator):
def __init__(self,
K=10,
alpha=0.75,
base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote',
diversity_metric='e',
positive_label=1):
self.K = K
self.alpha = alpha
self.base_classifier = base_classifier
self.n_classifiers = n_classifiers
self.positive_label = positive_label
self.ensemble = None
self.combiner = Combiner(rule=combination_rule)
self.diversity = Diversity(metric=diversity_metric)
self.validation_X = None
self.validation_y = None
def set_validation(self, X, y):
self.validation_X = X
self.validation_y = y
def fitness(self, classifier):
'''
# TODO normalize diversity metric.
'''
self.ensemble.add(classifier)
y_pred = self.predict(self.validation_X)
y_true = self.validation_y
auc = evaluation.auc_score(y_true, y_pred)
div = self.diversity.calculate(self.ensemble, self.validation_X,
y_true)
self.ensemble.classifiers.pop() # create interface for this later
return self.alpha * auc + (1.0 - self.alpha) * div
def _calc_pos_prob(self):
y_pred = self.predict(self.validation_X)
y_true = self.validation_y
# obtaining recall scores for each label (assuming the labels are
# binary)
pos_acc = recall_score(y_true,
y_pred,
average='binary',
pos_label=self.positive_label)
neg_acc = recall_score(y_true,
y_pred,
average='binary',
pos_label=int(not self.positive_label))
return neg_acc / (pos_acc + neg_acc)
def bootstrap_classifiers(self, X, y, K, pos_prob):
pos_idx = (y == self.positive_label)
neg_idx = (y == int(not self.positive_label))
X_pos, _ = X[pos_idx, :], y[pos_idx] # positive examples
X_neg, _ = X[neg_idx, :], y[neg_idx] # negative examples
classifiers = []
for i in range(K):
X_new = np.zeros(X.shape)
y_new = np.zeros(y.shape)
for j in range(X.shape[0]):
if pos_prob > np.random.random():
# add a randomly chosen positive example
idx = np.random.randint(X_pos.shape[0])
X_new[j, :] = X_pos[idx, :]
y_new[j] = self.positive_label
else:
# add a randomly chosen negative example
idx = np.random.randint(X_neg.shape[0])
X_new[j, :] = X_neg[idx, :]
y_new[j] = int(not self.positive_label)
# if no positive example is present, make sure you insert at least
# one
if not np.any(y_new == self.positive_label):
# chosen spot for replacement on new array
idx_new = np.random.randint(X_new.shape[0])
# chosen positive example index
idx_pos = np.random.randint(X_pos.shape[0])
X_new[idx_new, :] = X_pos[idx_pos, :]
y_new[idx_new] = self.positive_label
# if no negative example is present, make sure you insert at least
# one
elif not np.any(y_new == int(not self.positive_label)):
# chosen spot for replacement on new array
idx_new = np.random.randint(X_new.shape[0])
# chosen positive example index
idx_neg = np.random.randint(X_neg.shape[0])
X_new[idx_new, :] = X_neg[idx_neg, :]
y_new[idx_new] = int(not self.positive_label)
# train classifier with the bootstrapped data
clf = sklearn.base.clone(self.base_classifier)
clf.fit(X_new, y_new)
classifiers.append(clf)
return classifiers
def fit(self, X, y):
# if self.validation_X == None and self.validation_y == None:
self.validation_X = X
self.validation_y = y
self.classes_ = set(y)
self.ensemble = Ensemble()
clfs = self.bootstrap_classifiers(X, y, self.K, 0.5)
self.ensemble.add(np.random.choice(clfs))
for i in range(1, self.n_classifiers):
clfs = self.bootstrap_classifiers(X, y, self.K,
self._calc_pos_prob())
self.ensemble.add(max(clfs, key=lambda clf: self.fitness(clf)))
self.validation_X = None
self.validation_y = None
return self
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class SmoteBagging(PoolGenerator):
def __init__(self, base_classifier=None,
n_classifiers=100,
combination_rule='majority_vote', k=5):
# self.b = b
self.k = k
self.n_classifiers = n_classifiers
self.base_classifier = base_classifier
self.ensemble = None
self.combiner = Combiner(rule=combination_rule)
def smote_bootstrap_sample(self, X, y, b, k):
classes = np.unique(y)
count = np.bincount(y) # number of instances of each class
majority_class = count.argmax() # majority clas
majority_count = count.max() # majority class
data = np.empty((0, X.shape[1]))
target = np.empty((0,))
for i in classes:
class_data = X[(y == i), :]
if i == majority_class: # majority class
# regular bootstrap (i.e. 100% sampling rate)
idx = np.random.choice(majority_count, (majority_count,))
data = np.concatenate((data, class_data[idx, :]))
target = np.concatenate(
(target, i * np.ones((majority_count,))))
# print('original class data = {}'.format(class_data.shape))
# print('sampled class data = {}'.format(class_data[idx,:].shape)) # noqa
# print()
else: # minority classes
# bootstrap the class data with defined sampling rate
sample_rate = (majority_count /
class_data.shape[0]) * (b / 100)
idx = np.random.choice(
class_data.shape[0], (int(sample_rate * class_data.shape[0]),)) # noqa
sampled_class_data = class_data[idx, :]
# print('original class data = {}'.format(class_data.shape))
# print('majority_count = {}'.format(majority_count))
# print('class data = {}'.format(class_data.shape))
# print('b = {}'.format(b))
# print('sample rate = {}'.format(sample_rate))
# print('sampled class data = {}'.format(sampled_class_data.shape)) # noqa
# run smote on bootstrapped data to obtain synthetic samples
# ceil to make sure N_smote is a multiple of 100, and the small
# value to avoid a zero
N_smote = int(np.ceil(
(majority_count / sampled_class_data.shape[0]) * (1 - b / 100 + 10e-8)) * 100) # noqa
# print(N_smote)
# print('----------')
# print('smote parameters:')
# print('T : {}'.format(sampled_class_data.shape))
# print('N : {}'.format(N_smote))
synthetic = smote(sampled_class_data, N=N_smote, k=self.k)
# print('synthetic data = {})'.format(synthetic.shape))
# print(synthetic)
# add synthetic samples to sampled class data
n_missing = majority_count - sampled_class_data.shape[0]
idx = np.random.choice(synthetic.shape[0], (n_missing,))
new_class_data = np.concatenate(
(sampled_class_data, synthetic[idx, :]))
# print('new class data = {})'.format(new_class_data.shape))
# print()
data = np.concatenate((data, new_class_data))
target = np.concatenate(
(target, i * np.ones((new_class_data.shape[0],))))
return data, target
def fit(self, X, y):
self.ensemble = Ensemble()
# this parameter should change between [10, 100] with
# increments of 10, for every classifier in the ensemble
b = 10
for i in range(self.n_classifiers):
# print()
# print('classifier : {}'.format(i))
# print('------------------------')
# print('b = {}'.format(b))
data, target = self.smote_bootstrap_sample(
X, y, b=float(b), k=self.k)
# print('data = {}'.format(data.shape))
# print()
classifier = sklearn.base.clone(self.base_classifier)
classifier.fit(data, target)
self.ensemble.add(classifier)
if b >= 100:
b = 10
else:
b += 10
return
def predict(self, X):
out = self.ensemble.output(X)
return self.combiner.combine(out)
class EnsembleClassifier(object):
def __init__(self, ensemble=None, selector=None, combiner=None):
self.ensemble = ensemble
self.selector = selector
if combiner is None:
self.combiner = Combiner(rule='majority_vote')
elif isinstance(combiner, str):
self.combiner = Combiner(rule=combiner)
elif isinstance(combiner, Combiner):
self.combiner = combiner
else:
raise ValueError('Invalid parameter combiner')
def fit(self, X, y):
self.ensemble.fit(X, y)
def predict(self, X):
# TODO: warn the user if mode of ensemble
# output excludes the chosen combiner?
if self.selector is None:
out = self.ensemble.output(X)
y = self.combiner.combine(out)
else:
y = []
for i in range(X.shape[0]):
ensemble, weights = self.selector.select(
self.ensemble, X[i, :][np.newaxis, :])
if weights is not None: # use the ensemble with weights
if self.combiner.combination_rule == 'majority_vote':
out = ensemble.output(X[i, :][np.newaxis, :])
else:
out = ensemble.output(X[i, :][np.newaxis, :],
mode='probs')
# apply weights
for i in range(out.shape[2]):
out[:, :, i] = out[:, :, i] * weights[i]
[tmp] = self.combiner.combine(out)
y.append(tmp)
else: # use the ensemble, but ignore the weights
if self.combiner.combination_rule == 'majority_vote':
out = ensemble.output(X[i, :][np.newaxis, :])
else:
out = ensemble.output(X[i, :][np.newaxis, :],
mode='probs')
[tmp] = self.combiner.combine(out)
y.append(tmp)
return np.asarray(y)
def predict_proba(self, X):
# TODO: warn the user if mode of ensemble
# output excludes the chosen combiner?
if self.selector is None:
out = self.ensemble.output(X, mode='probs')
return np.mean(out, axis=2)
else:
out_full = []
for i in range(X.shape[0]):
ensemble, weights = self.selector.select(
self.ensemble, X[i, :][np.newaxis, :])
if weights is not None: # use the ensemble with weights
out = ensemble.output(X[i, :][np.newaxis, :])
# apply weights
for i in range(out.shape[2]):
out[:, :, i] = out[:, :, i] * weights[i]
# [tmp] = self.combiner.combine(out)
out_full.extend(list(np.mean(out, axis=2)))
else: # use the ensemble, but ignore the weights
out = ensemble.output(X[i, :][np.newaxis, :])
out_full.extend(list(np.mean(out, axis=2)))
# return np.asarray(y)
return np.array(out_full)
def score(self, X, y, sample_weight=None):
return accuracy_score(y, self.predict(X), sample_weight=sample_weight)