indices = np.argsort(importances)[::-1]
print('{}特征权值分布为: '.format(model_name))
for f in range(X_train.shape[1]):
print("%d. feature %d [%s] (%f)" %
(f + 1, indices[f], features_to_train[indices[f]],
importances[indices[f]]))
features_distribution.append(
(f + 1, indices[f], features_to_train[indices[f]],
importances[indices[f]]))
important_features.append(features_to_train[indices[f]])
print(important_features)
except AttributeError:
print('{} has no feture_importances_'.format(model_name))
try:
y_score = clf.decision_function(X_test)[:, 1]
except AttributeError:
print('{} has no decision_function, use predict func.'.format(
model_name))
y_score = clf.predict_proba(X_test)[:, 1]
# Compute ROC curve and ROC area for each class
roc_auc = roc_auc_score(y_test, y_score, sample_weight=None)
# Plot ROC curve
print('{} ROC curve (area = {})'.format(model_name, roc_auc))
print('Saving model %s to %s......' % (model_name, model_file))
joblib.dump(clf, model_file)
model_metainfo = {
'sub_file': sub_file,
def cvgrid_search(X, y, X_test, model_type):
X_train, X_val, y_train, y_val = train_test_split(X,
y,
stratify=y,
shuffle=True,
test_size=0.2,
random_state=42)
if model_type == 'cat':
model = CatBoostClassifier(**PARAMS['cat_def'])
grid_search_result = model.grid_search(PARAMS['cat_cv'],
X=X_train,
y=y_train,
plot=False,
cv=5,
stratified=True,
verbose=0)
best_model = CatBoostClassifier(**PARAMS['cat_def'],
**grid_search_result['params'])
best_model.fit(X_train, y_train, eval_set=(X_val, y_val), verbose=0)
get_preds = lambda model, x: model.predict(
x, prediction_type='Probability')[:, 1]
if model_type == 'rf':
rf = RandomForestClassifier()
rf_random = RandomizedSearchCV(estimator=rf,
param_distributions=PARAMS['rf_cv'],
n_iter=50,
cv=5,
verbose=0,
random_state=42,
n_jobs=-1)
# Tune hyperparams
rf_random.fit(X_train, y_train)
best_model = rf_random.best_estimator_
# Train set only
best_model.fit(X_train, y_train)
get_preds = lambda model, x: model.predict_proba(x)[:, 1]
if model_type == 'lr':
lr = linear_model.LogisticRegression()
lr_random = RandomizedSearchCV(estimator=lr,
param_distributions=PARAMS['lr_cv'],
n_iter=50,
cv=5,
verbose=0,
random_state=42,
n_jobs=-1)
# Tune hyperparams
lr_random.fit(X_train, y_train)
best_model = lr_random.best_estimator_
# Train set only
best_model.fit(X_train, y_train)
get_preds = lambda model, x: model.predict_proba(x)[:, 1]
if model_type == 'svm':
svc = SVC()
svc_random = RandomizedSearchCV(estimator=svc,
param_distributions=PARAMS['svm_cv'],
n_iter=50,
cv=5,
verbose=0,
random_state=42,
n_jobs=-1)
# Tune hyperparams
svc_random.fit(X_train, y_train)
best_model = svc_random.best_estimator_
# Train set only
best_model.fit(X_train, y_train)
get_preds = lambda model, x: model.decision_function(x)
y_pred_train = get_preds(best_model, X_train)
y_pred_val = get_preds(best_model, X_val)
train_auc = roc_auc_score(y_train, y_pred_train)
val_auc = roc_auc_score(y_val, y_pred_val)
train_ap = average_precision_score(y_train, y_pred_train)
val_ap = average_precision_score(y_val, y_pred_val)
# Final fit
best_model.fit(X, y)
return (get_preds(best_model, X),
get_preds(best_model,
X_test)), (train_auc, val_auc), (train_ap, val_ap)
