def test_ola_proba():
pool_classifiers, X_dsel, y_dsel, X_test, y_test = setup_classifiers()
ola = OLA(pool_classifiers, DFP=True)
ola.fit(X_dsel, y_dsel)
probas = ola.predict_proba(X_test)
expected = np.load('deslib/tests/expected_values/ola_proba_DFP.npy')
assert np.allclose(probas, expected)
def test_ola_proba(knn_methods):
pool_classifiers, X_dsel, y_dsel, X_test, y_test = setup_classifiers()
ola = OLA(pool_classifiers, knn_classifier=knn_methods)
ola.fit(X_dsel, y_dsel)
probas = ola.predict_proba(X_test)
expected = np.load(
'deslib/tests/expected_values/ola_proba_integration.npy')
assert np.allclose(probas, expected)
def test_estimate_competence_batch(example_estimate_competence):
_, _, neighbors, distances, dsel_processed, _ = example_estimate_competence
expected = np.array([[0.57142857, 0.71428571, 0.71428571],
[0.71428571, 0.85714286, 0.71428571],
[0.57142857, 0.71428571, 0.57142857]])
ola_test = OLA()
ola_test.DSEL_processed_ = dsel_processed
ola_test.DFP_mask = np.ones((3, 3))
competences = ola_test.estimate_competence(neighbors, distances=distances)
assert np.allclose(competences, expected)
def test_estimate_competence_batch():
expected = np.array([[0.57142857, 0.71428571, 0.71428571],
[0.71428571, 0.85714286, 0.71428571],
[0.57142857, 0.71428571, 0.57142857]])
ola_test = OLA(create_pool_classifiers())
ola_test.DSEL_processed_ = dsel_processed_ex1
neighbors = neighbors_ex1
distances = distances_ex1
ola_test.DFP_mask = np.ones((3, 3))
query = np.array([[1, 1], [1, 1], [1, 1]])
competences = ola_test.estimate_competence(query, neighbors, distances=distances)
assert np.allclose(competences, expected)
def train(train_index, test_index):
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = Y[train_index], Y[test_index]
#train_perc = 0.7
#split_point = int(train_perc*len(train_index))
# valid_index = train_index[split_point:]
# train_index = train_index[:split_point]
# X_train, X_valid, X_test = X[train_index], X[valid_index], X[test_index]
# y_train, y_valid, y_test = Y[train_index], Y[valid_index], Y[test_index]
#print("TRAIN:", train_index, "VALID:", valid_index, "TEST:", test_index)
X_train, X_valid, y_train, y_valid = train_test_split(
X_train, y_train, test_size=0.3, random_state=seed)
pool_classifiers.fit(X_train, y_train)
validation_data, validation_labels = get_validation_data(
X_valid, y_valid, 0.5, hardness=hardness)
dynamic_selection_algorithm = None
try:
if args.dynamic_selection == True and args.dynamic_algorithm is None:
raise ValueError(
'Dynamic selection requires you provide an algorithm.')
elif args.dynamic_selection == True and args.dynamic_algorithm is not None:
if args.dynamic_algorithm == 'ola':
dynamic_selection_algorithm = OLA(pool_classifiers,
random_state=seed)
elif args.dynamic_algorithm == 'lca':
dynamic_selection_algorithm = LCA(pool_classifiers,
random_state=seed)
elif args.dynamic_algorithm == 'mcb':
dynamic_selection_algorithm = MCB(pool_classifiers,
random_state=seed)
elif args.dynamic_algorithm == 'knorau':
dynamic_selection_algorithm = KNORAU(pool_classifiers,
random_state=seed)
elif args.dynamic_algorithm == 'kne':
dynamic_selection_algorithm = KNORAE(pool_classifiers,
random_state=seed)
dynamic_selection_algorithm.fit(validation_data,
validation_labels)
preds = dynamic_selection_algorithm.predict(X_test)
else:
# Static combination by voting
preds = voting(X_test, pool_classifiers)
except Exception as error:
raise error
acc = get_accuracy_score(y_test, preds)
g1 = get_g1_score(y_test, preds, average='macro')
f1 = get_f1_score(y_test, preds)
roc = roc_auc_score(y_test, preds, average='macro')
return dict(f1=f1, g1=g1, acc=acc, roc=roc)
def DCS(self, x_train, y_train,X_test, Y_test, dsel):
pool_classifiers = BaggingClassifier(linear_model.Perceptron(max_iter=5), self.pool_size)
pool_classifiers.fit(x_train, y_train)
# Initialize the DES model
lca = LCA(pool_classifiers)
ola = OLA(pool_classifiers)
# Preprocess the Dynamic Selection dataset (DSEL)
score1 = lca.fit(x_train[dsel], y_train[dsel])
score2 = ola.fit(x_train[dsel], y_train[dsel])
# Predict new examples:
# print (lca.score(X_test, Y_test), ola.score(X_test, Y_test))
return (score1, score2, ) + self.calc_metrics(X_test, Y_test) # dependendo da base formato nao suportado
def fit(self, x_sel, y_sel, P, k):
'''
metodo para chamar o tipo de DS
:param: x_sel: dados de treinamento da janela de validacao
:param: y_sel: rotulos da janela de validacao
:param: P: pool de classificadores
:param: k: vizinhanca
'''
# escolhendo a tecnica de selecao de classificadores
if (self.TYPE == 'knorae'):
DS = KNORAE(P, k)
elif (self.TYPE == 'knorau'):
DS = KNORAU(P, k)
elif (self.TYPE == 'ola'):
DS = OLA(P, k)
elif (self.TYPE == 'lca'):
DS = LCA(P, k)
elif (self.TYPE == 'posteriori'):
DS = APosteriori(P, k)
elif (self.TYPE == 'priori'):
DS = APriori(P, k)
# encontrando os classificadores competentes do DS escolhido
self.DS = copy.deepcopy(DS)
self.DS.fit(x_sel, y_sel)
def escolher_modelo(nome, x_sel, y_sel, P, k):
'''
metodo para chamar o tipo de DS
:param: x_sel: dados de treinamento da janela de validacao
:param: y_sel: rotulos da janela de validacao
:param: P: pool de classificadores
:param: k: vizinhanca
'''
# escolhendo a tecnica de selecao de classificadores
if(nome=='OLA'):
DS = OLA(P, k)
number_model = 0
elif(nome=='LCA'):
DS = LCA(P, k)
number_model = 1
elif(nome=='KNORAE'):
DS = KNORAE(P, k)
number_model = 2
elif(nome=='KNORAU'):
DS = KNORAU(P, k)
number_model = 3
# encontrando os classificadores competentes do DS escolhido
DS.fit(x_sel, y_sel)
# retornando a tecnica de DS
return DS, number_model
def test_estimate_competence(index, expected):
ola_test = OLA(create_pool_classifiers())
ola_test.processed_dsel = dsel_processed_ex1
ola_test.neighbors = neighbors_ex1[index, :]
ola_test.distances = distances_ex1[index, :]
ola_test.DFP_mask = [1, 1, 1]
query = np.array([1, 1])
competences = ola_test.estimate_competence(query.reshape(1, -1))
assert np.isclose(competences, expected).all()
def test_ola():
pool_classifiers, X_dsel, y_dsel, X_test, y_test = setup_classifiers()
ola = OLA(pool_classifiers, DFP=True)
ola.fit(X_dsel, y_dsel)
assert np.isclose(ola.score(X_test, y_test), 0.88181818181818183)
def main():
###############################################################################
# Preparing the dataset
# ---------------------
# In this part we load the breast cancer dataset from scikit-learn and
# preprocess it in order to pass to the DS models. An important point here is
# to normalize the data so that it has zero mean and unit variance, which is
# a common requirement for many machine learning algorithms.
# This step can be easily done using the StandardScaler class.
rng = np.random.RandomState(123)
data = load_breast_cancer()
X = data.data
y = data.target
# split the data into training and test data
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=rng)
# Scale the variables to have 0 mean and unit variance
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
# Split the data into training and DSEL for DS techniques
X_train, X_dsel, y_train, y_dsel = train_test_split(X_train,
y_train,
test_size=0.5,
random_state=rng)
# Train a pool of 100 base classifiers
pool_classifiers = BaggingClassifier(Perceptron(max_iter=10),
n_estimators=100,
random_state=rng)
pool_classifiers.fit(X_train, y_train)
# Initialize the DS techniques
knorau = KNORAU(pool_classifiers)
kne = KNORAE(pool_classifiers)
desp = DESP(pool_classifiers)
ola = OLA(pool_classifiers)
mcb = MCB(pool_classifiers)
###############################################################################
# Calibrating base classifiers
# -----------------------------
# Some dynamic selection techniques requires that the base classifiers estimate
# probabilities in order to estimate its competence level. Since the Perceptron
# model is not a probabilistic classifier (does not implements the
# predict_proba method, it needs to be calibrated for
# probability estimation before being used by such DS techniques. This step can
# be conducted using the CalibrateClassifierCV class from scikit-learn. Note
# that in this example we pass a prefited pool of classifiers to the
# calibration method in order to use exactly the same pool used in the other
# DS methods.
calibrated_pool = []
for clf in pool_classifiers:
calibrated = CalibratedClassifierCV(base_estimator=clf, cv='prefit')
calibrated.fit(X_dsel, y_dsel)
calibrated_pool.append(calibrated)
apriori = APriori(calibrated_pool)
meta = METADES(calibrated_pool)
knorau.fit(X_dsel, y_dsel)
kne.fit(X_dsel, y_dsel)
desp.fit(X_dsel, y_dsel)
ola.fit(X_dsel, y_dsel)
mcb.fit(X_dsel, y_dsel)
apriori.fit(X_dsel, y_dsel)
meta.fit(X_dsel, y_dsel)
###############################################################################
# Evaluating the methods
# -----------------------
# Let's now evaluate the methods on the test set. We also use the performance
# of Bagging (pool of classifiers without any selection) as a baseline
# comparison. We can see that the majority of DS methods achieve higher
# classification accuracy.
print('Evaluating DS techniques:')
print('Classification accuracy KNORA-Union: ',
knorau.score(X_test, y_test))
print('Classification accuracy KNORA-Eliminate: ',
kne.score(X_test, y_test))
print('Classification accuracy DESP: ', desp.score(X_test, y_test))
print('Classification accuracy OLA: ', ola.score(X_test, y_test))
print('Classification accuracy A priori: ', apriori.score(X_test, y_test))
print('Classification accuracy MCB: ', mcb.score(X_test, y_test))
print('Classification accuracy META-DES: ', meta.score(X_test, y_test))
print('Classification accuracy Bagging: ',
pool_classifiers.score(X_test, y_test))
def _generate_local_pool(self, query):
"""
Local pool generation.
This procedure populates the "pool_classifiers" based on the query sample's neighborhood.
Thus, for each query sample, a different pool is created.
In each iteration, the training samples near the query sample are singled out and a
subpool is generated using the Self-Generating Hyperplanes (SGH) method.
Then, the DCS technique selects the best classifier in the generated subpool and it is added to the local pool.
In the following iteration, the neighborhood is increased and another SGH-generated subpool is obtained
over the new neighborhood, and again the DCS technique singles out the best in it, which is then added to the local pool.
This process is repeated until the pool reaches "n_classifiers".
Parameters
----------
query : array of shape = [n_features]
The test sample.
Returns
-------
self
References
----------
M. A. Souza, G. D. Cavalcanti, R. M. Cruz, R. Sabourin, On the characterization of the
oracle for dynamic classi
er selection, in: International Joint Conference on Neural Networks,
IEEE, 2017, pp. 332-339.
"""
n_samples, _ = self.DSEL_data.shape
self.pool_classifiers = []
n_err = 0
max_err = 2 * self.n_classifiers
curr_k = self.k
# Classifier count
n = 0
while n < self.n_classifiers and n_err < max_err:
subpool = SGH()
included_samples = np.zeros((n_samples), int)
if self.knne:
idx_neighb = np.array([], dtype=int)
# Obtain neighbors of each class individually
for j in np.arange(0, self.n_classes):
# Obtain neighbors from the classes in the RoC
if np.any(self.classes[j] == self.DSEL_target[
self.neighbors[0][np.arange(0, curr_k)]]):
nc = np.where(self.classes[j] == self.DSEL_target[
self.neighbors[0]])
idx_nc = self.neighbors[0][nc]
idx_nc = idx_nc[np.arange(
0, np.minimum(curr_k, len(idx_nc)))]
idx_neighb = np.concatenate((idx_neighb, idx_nc),
axis=0)
else:
idx_neighb = np.asarray(self.neighbors)[0][np.arange(
0, curr_k)]
# Indicate participating instances in the training of the subpool
included_samples[idx_neighb] = 1
curr_classes = np.unique(self.DSEL_target[idx_neighb])
# If there are +1 classes in the local region
if len(curr_classes) > 1:
# Obtain SGH pool
subpool.fit(self.DSEL_data, self.DSEL_target, included_samples)
# Adjust chosen DCS technique parameters
if self.ds_tech == 'ola':
ds = OLA(subpool, k=len(idx_neighb)) # change for self.k
elif self.ds_tech == 'lca':
ds = LCA(subpool, k=len(idx_neighb))
elif self.ds_tech == 'mcb':
ds = MCB(subpool, k=len(idx_neighb))
elif self.ds_tech == 'mla':
ds = MLA(subpool, k=len(idx_neighb))
elif self.ds_tech == 'a_priori':
ds = APriori(subpool, k=len(idx_neighb))
elif self.ds_tech == 'a_posteriori':
ds = APosteriori(subpool, k=len(idx_neighb))
# Fit ds technique
ds.fit(self.DSEL_data, self.DSEL_target)
neighb = np.in1d(
self.neighbors,
idx_neighb) # True/False vector of selected neighbors
# Set distances and neighbors of the query sample (already calculated)
ds.distances = np.asarray([self.distances[0][neighb]
]) # Neighborhood
ds.neighbors = np.asarray([self.neighbors[0][neighb]
]) # Neighborhood
ds.DFP_mask = np.ones(ds.n_classifiers)
# Estimate competence
comp = ds.estimate_competence(query, ds._predict_base(query))
# Select best classifier in subpool
sel_c = ds.select(comp)
# Add to local pool
self.pool_classifiers.append(copy.deepcopy(subpool[sel_c[0]]))
n += 1
# else:
# # Exception: fewer than 2 classes in the neighborhood
# print('OPS! Next!')
# Increase neighborhood size
curr_k += 2
n_err += 1
return self
# Split the data into training and DSEL for DS techniques
X_train, X_dsel, y_train, y_dsel = train_test_split(X_train,
y_train,
test_size=0.5,
random_state=rng)
# Considering a pool composed of 10 base classifiers
pool_classifiers = RandomForestClassifier(n_estimators=10,
random_state=rng,
max_depth=10)
pool_classifiers.fit(X_train, y_train)
# DS techniques without DFP
apriori = APriori(pool_classifiers)
aposteriori = APosteriori(pool_classifiers)
ola = OLA(pool_classifiers)
lca = LCA(pool_classifiers)
desp = DESP(pool_classifiers)
meta = METADES(pool_classifiers)
apriori.fit(X_dsel, y_dsel)
aposteriori.fit(X_dsel, y_dsel)
ola.fit(X_dsel, y_dsel)
lca.fit(X_dsel, y_dsel)
desp.fit(X_dsel, y_dsel)
meta.fit(X_dsel, y_dsel)
print('Evaluating DS techniques:')
print('Classification accuracy of OLA: ', ola.score(X_test, y_test))
print('Classification accuracy of LCA: ', lca.score(X_test, y_test))
print('Classification accuracy of A priori: ', apriori.score(X_test, y_test))
# Training a random forest to be used as the pool of classifiers.
# We set the maximum depth of the tree so that it
# can estimate probabilities
pool_classifiers = RandomForestClassifier(n_estimators=100, max_depth=5,
random_state=rng)
pool_classifiers.fit(X_train, y_train)
stacked = StackedClassifier(pool_classifiers, LogisticRegression())
stacked.fit(X_dsel, y_dsel)
# Initialize a DS technique. Here we specify the size of
# the region of competence (5 neighbors)
knorau = KNORAU(pool_classifiers, random_state=rng)
kne = KNORAE(pool_classifiers, k=5, random_state=rng)
desp = DESP(pool_classifiers, k=5, random_state=rng)
ola = OLA(pool_classifiers, k=5, random_state=rng)
mcb = MCB(pool_classifiers, k=5, random_state=rng)
meta = METADES(pool_classifiers, k=5, random_state=rng)
# Fit the DS techniques
knorau.fit(X_dsel, y_dsel)
kne.fit(X_dsel, y_dsel)
desp.fit(X_dsel, y_dsel)
meta.fit(X_dsel, y_dsel)
ola.fit(X_dsel, y_dsel)
mcb.fit(X_dsel, y_dsel)
###############################################################################
# Plotting the results
# -----------------------
# Let's now evaluate the methods on the test set.
def test_ola(knn_methods):
pool_classifiers, X_dsel, y_dsel, X_test, y_test = setup_classifiers()
ola = OLA(pool_classifiers, knn_classifier=knn_methods)
ola.fit(X_dsel, y_dsel)
assert np.isclose(ola.score(X_test, y_test), 0.96808510638297873)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25)
RF = RandomForestClassifier()
RF.fit(X_train, y_train)
X_train, X_dsel, y_train, y_dsel = train_test_split(X, y, test_size=0.50)
# Training a random forest to be used as the pool of classifiers. We set the maximum depth of the tree so that it
# can estimate probabilities
pool_classifiers = RandomForestClassifier(n_estimators=10, max_depth=5)
pool_classifiers.fit(X_train, y_train)
# Initialize a DS technique. Here we specify the size of the region of competence (5 neighbors)
knorau = KNORAU(pool_classifiers)
kne = KNORAE(pool_classifiers, k=5)
desp = DESP(pool_classifiers, k=5)
ola = OLA(pool_classifiers, k=5)
mcb = MCB(pool_classifiers, k=5)
meta = METADES(pool_classifiers, k=5)
# Fit the DS techniques
knorau.fit(X_dsel, y_dsel)
kne.fit(X_dsel, y_dsel)
desp.fit(X_dsel, y_dsel)
meta.fit(X_dsel, y_dsel)
ola.fit(X_dsel, y_dsel)
mcb.fit(X_dsel, y_dsel)
# Calculate classification accuracy of each technique
print('Classification accuracy RF: ', RF.score(X_test, y_test))
print('Evaluating DS techniques:')
print('Classification accuracy KNORAU: ', knorau.score(X_test, y_test))
ax.set_ylim((0, 1))
plt.show()
plt.tight_layout()
###############################################################################
# Comparison with Dynamic Selection techniques
# --------------------------------------------
#
# We will now consider four DS methods: k-Nearest Oracle-Eliminate (KNORA-E),
# Dynamic Ensemble Selection performance (DES-P), Overall Local Accuracy (OLA)
# and Rank. Let's train the classifiers and plot their decision boundaries:
knora_e = KNORAE(pool_classifiers).fit(X_train, y_train)
desp = DESP(pool_classifiers).fit(X_train, y_train)
ola = OLA(pool_classifiers).fit(X_train, y_train)
rank = Rank(pool_classifiers).fit(X_train, y_train)
# Plotting the Decision Border of the DS methods.
fig2, sub = plt.subplots(2, 2, figsize=(15, 10))
plt.subplots_adjust(wspace=0.4, hspace=0.4)
titles = [
'KNORA-Eliminate', 'DES-P', 'Overall Local Accuracy (OLA)', 'Modified Rank'
]
classifiers = [knora_e, desp, ola, rank]
for clf, ax, title in zip(classifiers, sub.flatten(), titles):
plot_classifier_decision(ax, clf, X_train, mode='filled', alpha=0.4)
plot_dataset(X_test, y_test, ax=ax)
ax.set_xlim(np.min(X[:, 0]), np.max(X[:, 0]))
ax.set_ylim(np.min(X[:, 1]), np.max(X[:, 1]))
RF.fit(X_train, y_train)
X_train, X_dsel, y_train, y_dsel = train_test_split(X_train,
y_train,
test_size=0.50,
random_state=rng)
stacked = StackedClassifier(RF, LogisticRegression())
stacked.fit(X_dsel, y_dsel)
# Initialize a DS technique. Here we specify the size of
# the region of competence (5 neighbors)
knorau = KNORAU(RF, k=5, random_state=rng)
kne = KNORAE(RF, k=5, random_state=rng)
desp = DESP(RF, k=5, random_state=rng)
ola = OLA(RF, k=5, random_state=rng)
mcb = MCB(RF, k=5, random_state=rng)
meta = METADES(RF, k=5, random_state=rng)
# Fit the DS techniques
knorau.fit(X_dsel, y_dsel)
kne.fit(X_dsel, y_dsel)
desp.fit(X_dsel, y_dsel)
meta.fit(X_dsel, y_dsel)
ola.fit(X_dsel, y_dsel)
mcb.fit(X_dsel, y_dsel)
###############################################################################
# Plotting the results
# -----------------------
# Let's now evaluate the methods on the test set.
def test_predict_proba(create_X_y):
X, y = create_X_y
clf1 = Perceptron()
clf1.fit(X, y)
OLA([clf1, clf1]).fit(X, y)
def test_predict_proba():
X = X_dsel_ex1
y = y_dsel_ex1
clf1 = Perceptron()
clf1.fit(X, y)
OLA([clf1, clf1])
def test_check_estimator():
check_estimator(OLA())
def test_ola():
pool_classifiers, X_dsel, y_dsel, X_test, y_test = setup_classifiers()
ola = OLA(pool_classifiers, DFP=True, with_IH=True, IH_rate=0.1)
ola.fit(X_dsel, y_dsel)
assert np.isclose(ola.score(X_test, y_test), 0.9030303030303031)