def oversample(self):
"""Function tranforming train data using undersampling and oversampling. Uses undersampling as well as oversampling if the
ratio between classes is highly imbalanced. Otherwise only oversampling will be used
Keyword args:
None
Returns:
os_X_train, os_y_train. As matrices (as returned by SMOTE). os_X_train and os_y_train are saved to class and will be automatically
applied during the search method, if oversample method is run first.
"""
unique, counts = np.unique(self.y_train, return_counts=True)
count_dict = dict(zip(unique, counts))
y_keys, y_counts = count_dict.keys(), count_dict.values()
sorted_counts = sorted(y_counts, reverse=True)
highest = sorted_counts[0]
second_highest = sorted_counts[1]
ratio = highest / second_highest
# under sample then oversample
if (ratio > 4):
under = RandomUnderSampler(sampling_strategy=0.5)
over = SMOTE()
pipeline_steps = [('under_sample', under), ('over_sample', over)]
pipeline = Pipeline(steps=pipeline_steps)
os_X, os_y = pipeline.fit_resample(self.X_train, self.y_train)
else:
over = SMOTE()
pipeline_steps = [('over_sample', over)]
pipeline = Pipeline(steps=pipeline_steps)
os_X, os_y = pipeline.fit_resample(self.X_train, self.y_train)
self.imb_pipeline_steps = pipeline_steps
# osx is returned if you choose to use it yourself, however pipeline will be used in grid search automatically
# https://stackoverflow.com/questions/50245684/using-smote-with-gridsearchcv-in-scikit-learn
return os_X, os_y
}
model = RandomForestClassifier()
conf_matrix_list_of_arrays = []
scores=[]
for i in range(10):
for fold_ind, (train_index, test_index) in enumerate(stratified_group_k_fold(X, y, ids, k=8)):
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
train_groups, test_groups = ids[train_index], ids[test_index]
ipca = IncrementalPCA(n_components=X_train.shape[1]//5, batch_size=120)
ipca.fit(X_train)
X_train=ipca.transform(X_train)
X_test=ipca.transform(X_test)
X_train, y_train = pipeline.fit_resample(X_train, y_train)#Smote
clf = GridSearchCV(model, parameters,cv=5, n_jobs=4)
clf.fit(X_train, y_train)
pred = clf.predict(X_test)
conf_matrix = confusion_matrix(y_test, pred)
conf_matrix_list_of_arrays.append(conf_matrix)
score=accuracy_score(y_test, pred)
scores.append(score)
mean_of_conf_matrix_arrays = np.mean(conf_matrix_list_of_arrays, axis=0)
print(mean_of_conf_matrix_arrays,file=f)
print('Accuracy: %.7f (%.7f)' % (np.mean(scores), np.std(scores)),file=f)
f.close()
conf_matrix_list_of_arrays = []
scores=[]
for i in range(10):
for fold_ind, (train_index, test_index) in enumerate(stratified_group_k_fold(Xlog, y, ids, k=8)):
print("Fold ", fold_ind)
Xlog_train, Xlog_test = Xlog[train_index], Xlog[test_index]
Xeye_train, Xeye_test = Xeye[train_index], Xeye[test_index]
y_train, y_test = y[train_index], y[test_index]
train_groups, test_groups = ids[train_index], ids[test_index]
# print(data,X_train.shape)
pca = PCA(n_components=0.9999)
pca.fit(Xlog_train)
Xlog_train=pca.transform(Xlog_train)
Xlog_test=pca.transform(Xlog_test)
Xlog_train, ylog_train = pipeline.fit_resample(Xlog_train, y_train)#Smote
# print(data,X_train.shape)
pca = PCA(n_components=0.9999)
pca.fit(Xeye_train)
Xeye_train=pca.transform(Xeye_train)
Xeye_test=pca.transform(Xeye_test)
Xeye_train, yeye_train = pipeline.fit_resample(Xeye_train, y_train)#Smote
clf1 = GridSearchCV(rf, parameters_rf,cv=5, n_jobs=4)
clf1.fit(Xlog_train, ylog_train)
y_pred = clf1.best_estimator_.predict(Xlog_train)
m1=confusion_matrix(ylog_train, y_pred)
pred1 = clf1.predict_proba(Xlog_test)
