def study_khan(features, target, preprocessing=StandardScaler(), grid=True, random_seed=42, output_file='khan.json'):
results= {}
base_classifier= SVC(kernel='rbf', probability=True, random_state=random_seed)
grid_search_params= {'C': [10**i for i in range(-4, 5)]}
np.random.seed(random_seed)
# without oversampling
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='accuracy')
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=5, random_state= random_seed)
preds= evaluate(pipeline, features, target, validator)
results['without_oversampling_auc']= accuracy_score(preds['label'].values, preds['prediction'].values > 0.5)
results['without_oversampling_details']= preds.to_dict()
print('without oversampling: ', results['without_oversampling_auc'])
# with correct oversampling
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='accuracy')
classifier= OversamplingClassifier(ADASYN(random_state=random_seed), classifier)
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=5, random_state= random_seed)
preds= evaluate(classifier, features, target, validator)
results['with_oversampling_auc']= accuracy_score(preds['label'].values, preds['prediction'].values > 0.5)
results['with_oversampling_details']= preds.to_dict()
print('with oversampling: ', results['with_oversampling_auc'])
# in-sample evaluation
classifier= base_classifier
preds= classifier.fit(features.values, target.values).predict_proba(features.values)[:,1]
preds= pd.DataFrame({'fold': 0, 'label': target.values, 'prediction': preds})
results['in_sample_auc']= accuracy_score(preds['label'].values, preds['prediction'].values > 0.5)
results['in_sample_details']= preds.to_dict()
print('in sample: ', results['in_sample_auc'])
# with incorrect oversampling
X, y= ADASYN(random_state=random_seed).sample(features.values, target.values)
X= pd.DataFrame(X, columns=features.columns)
y= pd.Series(y)
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='accuracy')
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=5, random_state=random_seed)
preds= evaluate(pipeline, X, y, validator)
results['incorrect_oversampling_auc']= accuracy_score(preds['label'].values, preds['prediction'].values > 0.5)
results['incorrect_oversampling_details']= preds.to_dict()
print('incorrect oversampling: ', results['incorrect_oversampling_auc'])
json.dump(results, open(output_file, 'w'))
return results
def study_jagerlibensek(features, target, preprocessing=StandardScaler(), grid=True, random_seed=42, output_file='acharya_results.json'):
results= {}
base_classifier= QuadraticDiscriminantAnalysis()
grid_search_params= {'reg_param': [0.01, 0.1, 0.5, 0.9, 0.99]}
# without oversampling
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='roc_auc')
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=10, random_state= random_seed)
preds= evaluate(pipeline, features, target, validator)
results['without_oversampling_auc']= roc_auc_score(preds['label'].values, preds['prediction'].values)
results['without_oversampling_details']= preds.to_dict()
print('without oversampling: ', results['without_oversampling_auc'])
# with correct oversampling
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='roc_auc')
classifier= OversamplingClassifier(ADASYN(), classifier)
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=10, random_state= random_seed)
preds= evaluate(classifier, features, target, validator)
results['with_oversampling_auc']= roc_auc_score(preds['label'].values, preds['prediction'].values)
results['with_oversampling_details']= preds.to_dict()
print('with oversampling: ', results['with_oversampling_auc'])
# in-sample evaluation
classifier= base_classifier
preds= classifier.fit(features.values, target.values).predict_proba(features.values)[:,1]
preds= pd.DataFrame({'fold': 0, 'label': target.values, 'prediction': preds})
results['in_sample_auc']= roc_auc_score(preds['label'].values, preds['prediction'].values)
results['in_sample_details']= preds.to_dict()
print('in sample: ', results['in_sample_auc'])
# with incorrect oversampling
X, y= ADASYN().sample(features.values, target.values)
X= pd.DataFrame(X, columns=features.columns)
y= pd.Series(y)
classifier= base_classifier if not grid else GridSearchCV(base_classifier, grid_search_params, scoring='roc_auc')
pipeline= classifier if not preprocessing else Pipeline([('preprocessing', preprocessing), ('classifier', classifier)])
validator= StratifiedKFold(n_splits=10, random_state= random_seed)
preds= evaluate(pipeline, X, y, validator)
results['incorrect_oversampling_auc']= roc_auc_score(preds['label'].values, preds['prediction'].values)
results['incorrect_oversampling_details']= preds.to_dict()
print('incorrect oversampling: ', results['incorrect_oversampling_auc'])
json.dump(results, open(output_file, 'w'))
return results
def fit(self, X, y):
base_classifier= SVC(kernel='rbf', probability=True, random_state=self.random_state)
grid_search_params= {'C': [10**i for i in range(-4, 5)]}
classifier= base_classifier if not self.grid else GridSearchCV(base_classifier, grid_search_params, scoring='accuracy')
classifier= OversamplingClassifier(ADASYN(random_state=self.random_state), classifier)
self.pipeline= classifier if not self.preprocessing else Pipeline([('preprocessing', self.preprocessing), ('classifier', classifier)])
self.pipeline.fit(X, y)
return self
def fit(self, X, y):
base_classifier= QuadraticDiscriminantAnalysis()
grid_search_params= {'reg_param': [0.01, 0.1, 0.5, 0.9, 0.99]}
classifier= base_classifier if not self.grid else GridSearchCV(base_classifier, grid_search_params, scoring='roc_auc')
classifier= OversamplingClassifier(ADASYN(), classifier)
self.pipeline= classifier if not self.preprocessing else Pipeline([('preprocessing', self.preprocessing), ('classifier', classifier)])
self.pipeline.fit(X, y)
return self
def fit(self, X, y):
base_classifier = RandomForestClassifier(
random_state=self.random_state)
grid_search_params = {
'max_depth': [3, 5, 10, None],
'min_samples_leaf': [1, 3, 5]
}
classifier = base_classifier if not self.grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
classifier = OversamplingClassifier(
ADASYN(random_state=self.random_state), classifier)
self.pipeline = classifier if not self.preprocessing else Pipeline([
('preprocessing', self.preprocessing), ('classifier', classifier)
])
self.pipeline.fit(X, y)
return self
def fit(self, X, y):
base_classifier = SVC(kernel='linear',
random_state=self.random_state,
probability=True)
grid_search_params = {
'kernel': ['linear'],
'C': [10**i for i in range(-4, 5)],
'probability': [True],
'random_state': [self.random_state]
}
classifier = base_classifier if not self.grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
classifier = OversamplingClassifier(ADASYN(), classifier)
self.pipeline = classifier if not self.preprocessing else Pipeline([
('preprocessing', self.preprocessing), ('classifier', classifier)
])
self.pipeline.fit(X, y)
return self
def study_sadiahmed(features,
target,
preprocessing=StandardScaler(),
grid=True,
random_seed=42,
output_file='sadiahmed_results.json'):
features = features.loc[:, sadiahmed_features + ['Rectime', 'Gestation']]
mask = (features['Rectime'] >= 27) & (features['Rectime'] <= 32)
features = features.loc[mask, :]
target = target.loc[mask]
term_ix = np.arange(len(target), dtype=int)[target == 1]
preterm_ix = np.arange(len(target), dtype=int)[target == 0]
term_diffs = list(features.iloc[term_ix]['Gestation'] -
features.iloc[term_ix]['Rectime'])
preterm_diffs = list(features.iloc[preterm_ix]['Gestation'] -
features.iloc[preterm_ix]['Rectime'])
features = features.drop(['Rectime', 'Gestation'], axis=1)
# These indices seem to best resemble the reported mean and stddev...
term_sample = list(term_ix[np.argsort(term_diffs)[10:25]])
preterm_sample = list(preterm_ix[np.argsort(preterm_diffs)[2:17]])
features = features.iloc[term_sample + preterm_sample, :]
target = target.iloc[term_sample + preterm_sample]
results = {}
base_classifier = SVC(kernel='linear',
random_state=random_seed,
probability=True)
grid_search_params = {
'kernel': ['linear'],
'C': [10**i for i in range(-4, 5)],
'probability': [True],
'random_state': [random_seed]
}
# without oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(pipeline, features, target, validator)
results['without_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['without_oversampling_details'] = preds.to_dict()
print('without oversampling: ', results['without_oversampling_auc'])
# with correct oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
classifier = OversamplingClassifier(ADASYN(), classifier)
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(classifier, features, target, validator)
results['with_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['with_oversampling_details'] = preds.to_dict()
print('with oversampling: ', results['with_oversampling_auc'])
# in-sample evaluation
classifier = base_classifier
preds = classifier.fit(features.values,
target.values).predict_proba(features.values)[:, 1]
preds = pd.DataFrame({
'fold': 0,
'label': target.values,
'prediction': preds
})
results['in_sample_auc'] = roc_auc_score(preds['label'].values,
preds['prediction'].values)
results['in_sample_details'] = preds.to_dict()
print('in sample: ', results['in_sample_auc'])
# with incorrect oversampling
X, y = ADASYN().sample(features.values, target.values)
X = pd.DataFrame(X, columns=features.columns)
y = pd.Series(y)
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(pipeline, X, y, validator)
results['incorrect_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['incorrect_oversampling_details'] = preds.to_dict()
print('incorrect oversampling: ', results['incorrect_oversampling_auc'])
json.dump(results, open(output_file, 'w'))
return results
def study_acharya(features,
target,
preprocessing=StandardScaler(),
grid=True,
random_seed=42,
output_file='acharya_results.json'):
#features= features.loc[:,acharya_features]
results = {}
base_classifier = SVC(kernel='rbf',
random_state=random_seed,
probability=True)
grid_search_params = {
'kernel': ['rbf'],
'C': [10**i for i in range(-4, 5)],
'probability': [True],
'random_state': [random_seed]
}
# without oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='accuracy')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(pipeline, features, target, validator)
results['without_oversampling_acc'] = accuracy_score(
preds['label'].values, preds['prediction'].values > 0.5)
results['without_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['without_oversampling_details'] = preds.to_dict()
print('without oversampling: ', results['without_oversampling_auc'])
# with correct oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='accuracy')
classifier = OversamplingClassifier(ADASYN(), classifier)
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
correct_pipeline = pipeline
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(classifier, features, target, validator)
results['with_oversampling_acc'] = accuracy_score(
preds['label'].values, preds['prediction'].values > 0.5)
results['with_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['with_oversampling_details'] = preds.to_dict()
print('with oversampling: ', results['with_oversampling_auc'])
# with incorrect oversampling
X, y = ADASYN().sample(features.values, target.values)
X = pd.DataFrame(X, columns=features.columns)
y = pd.Series(y)
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='accuracy')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=10, random_state=random_seed)
preds = evaluate(pipeline, X, y, validator)
results['incorrect_oversampling_acc'] = accuracy_score(
preds['label'].values, preds['prediction'].values > 0.5)
results['incorrect_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['incorrect_oversampling_details'] = preds.to_dict()
print('incorrect oversampling: ', results['incorrect_oversampling_auc'])
json.dump(results, open(output_file, 'w'))
return results
def study_peng(features,
target,
preprocessing=StandardScaler(),
grid=True,
random_seed=42,
output_file='idowu.json'):
results = {}
base_classifier = RandomForestClassifier(random_state=random_seed)
grid_search_params = {
'max_depth': [3, 5, 10, None],
'min_samples_leaf': [1, 3, 5]
}
mask = features['Rectime'] <= 26
features = features.loc[mask, :]
target = target.loc[mask]
features = features.drop('Rectime', axis=1)
np.random.seed(random_seed)
# without oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=5, random_state=random_seed)
preds = evaluate(pipeline, features, target, validator)
results['without_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['without_oversampling_details'] = preds.to_dict()
print('without oversampling: ', results['without_oversampling_auc'])
# with correct oversampling
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
classifier = OversamplingClassifier(ADASYN(random_state=random_seed),
classifier)
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=5, random_state=random_seed)
preds = evaluate(classifier, features, target, validator)
results['with_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['with_oversampling_details'] = preds.to_dict()
print('with oversampling: ', results['with_oversampling_auc'])
# in-sample evaluation
classifier = base_classifier
preds = classifier.fit(features.values,
target.values).predict_proba(features.values)[:, 1]
preds = pd.DataFrame({
'fold': 0,
'label': target.values,
'prediction': preds
})
results['in_sample_auc'] = roc_auc_score(preds['label'].values,
preds['prediction'].values)
results['in_sample_details'] = preds.to_dict()
print('in sample: ', results['in_sample_auc'])
# with incorrect oversampling
X, y = ADASYN(random_state=random_seed).sample(features.values,
target.values)
X = pd.DataFrame(X, columns=features.columns)
y = pd.Series(y)
classifier = base_classifier if not grid else GridSearchCV(
base_classifier, grid_search_params, scoring='roc_auc')
pipeline = classifier if not preprocessing else Pipeline([('preprocessing',
preprocessing),
('classifier',
classifier)])
validator = StratifiedKFold(n_splits=5, random_state=random_seed)
preds = evaluate(pipeline, X, y, validator)
results['incorrect_oversampling_auc'] = roc_auc_score(
preds['label'].values, preds['prediction'].values)
results['incorrect_oversampling_details'] = preds.to_dict()
print('incorrect oversampling: ', results['incorrect_oversampling_auc'])
json.dump(results, open(output_file, 'w'))
return results