def balance_classes(X, X_y, X_strat, target, train_cols, method):
y = X_strat.astype(str)
x = X.copy()
# x[target] = X_y
if method == "over":
from imblearn.over_sampling import RandomOverSampler
x, y = RandomOverSampler().fit_sample(x, y)
elif method == "under":
from imblearn.under_sampling import RandomUnderSampler
x, y = RandomUnderSampler().fit_sample(x, y)
elif method == "smote":
from imblearn.over_sampling import SMOTE
features = list(
set(
x.select_dtypes(include=[
"uint8", "int16", "int32", "int64", "float16", "float32",
"float64"
]).columns) - set(X_strat.name))
no_action = list(set(x.columns) - set(features))
x_sm = x[features].copy()
x_no_action = x[no_action].copy()
x, y = SMOTE().fit_sample(x_sm, y)
x[no_action] = x_no_action[no_action]
x[X_strat.name] = y
X = x.copy()
X_y = x[target]
gc.collect()
return X, X_y
def train(self):
fit_params = {
'early_stopping_rounds': 50,
'eval_metric': 'auc',
'eval_names': ['valid'],
'verbose': 100,
'categorical_feature': 'auto'
}
param_test = {
'num_leaves': sp_randint(6, 50),
'min_child_samples': sp_randint(100, 500),
'min_child_weight':
[1e-5, 1e-3, 1e-2, 1e-1, 1, 1e1, 1e2, 1e3, 1e4],
'subsample': sp_uniform(loc=0.2, scale=0.8),
'colsample_bytree': sp_uniform(loc=0.4, scale=0.6),
'reg_alpha': [0, 1e-1, 1, 2, 5, 7, 10, 50, 100],
'reg_lambda': [0, 1e-1, 1, 5, 10, 20, 50, 100]
}
# split data
X_train, X_test, y_train, y_test = train_test_split(
self.df_train_features,
self.df_train_label,
test_size=self.test_size)
# random oversampling
ros = RandomOverSampler(random_state=0)
X_train, y_train = RandomOverSampler().fit_resample(X_train, y_train)
clf = lgb.LGBMClassifier(max_depth=-1,
random_state=314,
silent=True,
metric='None',
n_jobs=4,
n_estimators=900)
# set randomsearch optimization params
random_search = RandomizedSearchCV(estimator=clf,
param_distributions=param_test,
scoring='roc_auc',
cv=5,
refit=True,
random_state=314,
verbose=False)
fit_params = {**fit_params, **{'eval_set': [(X_test, y_test)]}}
# fit random search
random_search.fit(X_train.copy(), y_train.copy(), **fit_params)
# build final classifier
clf_final = lgb.LGBMClassifier(
**random_search.best_estimator_.get_params())
clf_final.fit(
X_train,
y_train,
**fit_params,
callbacks=[
lgb.reset_parameter(
learning_rate=self.learning_rate_010_decay_power_0995)
])
# save model
joblib.dump(clf_final, os.path.join(MODELS_DIR, "{}.pkl".format("ML")))
predicted = clf_final.predict(X_test)
return {
"accuracy":
accuracy_score(y_test, predicted),
"roc":
roc_auc_score(y_test, predicted),
"cm":
confusion_matrix(y_test, predicted, normalize="true"),
"feature_importance":
pd.DataFrame(sorted(
zip(clf_final.feature_importances_, X_test.columns)),
columns=['Value', 'Feature'])
}
def train(self):
# split data
X_train, X_test, y_train, y_test = train_test_split(
self.df_train_features,
self.df_train_label,
test_size=self.test_size
)
# random oversampling
ros = RandomOverSampler(random_state=0)
X_train, y_train = RandomOverSampler().fit_resample(X_train, y_train)
clf = lgb.LGBMClassifier(
max_depth=-1,
random_state=314,
silent=True,
metric='None',
n_jobs=4,
n_estimators=10000)
# set randomsearch optimization params
random_search = RandomizedSearchCV(
estimator=clf,
param_distributions=config.param_test,
scoring='roc_auc',
cv=5,
refit=True,
random_state=314,
verbose=False)
fit_params = {
** config.fit_params,
**{'eval_set': [(X_test, y_test)]}}
# fit random search
random_search.fit(
X_train.copy(),
y_train.copy(),
**fit_params)
# build final classifier
clf_final = lgb.LGBMClassifier(**random_search.best_estimator_.get_params())
clf_final.fit(
X_train,
y_train,
**fit_params,
callbacks=[
lgb.reset_parameter(
learning_rate=self.learning_rate_010_decay_power_0995
)])
# save model
joblib.dump(clf_final, os.path.join(
config.MODELS_DIR,
"{}.pkl".format("ML")))
predicted = clf_final.predict(X_test)
return {
"accuracy": accuracy_score(y_test, predicted),
"roc": roc_auc_score(y_test, predicted),
"cm": confusion_matrix(y_test, predicted, normalize="true"),
"feature_importance": pd.DataFrame(
sorted(
zip(
clf_final.feature_importances_,
X_test.columns)),
columns=['Value','Feature'])
}
