4/8

def 

모델
데이터 샘플링
변수 중요도(shap)

def extract_good_features_using_shap_LGB(params, SEED):
    clf = lgb.LGBMClassifier(objective='multiclass',
                             random_state=1989,
                             silent=True, 
                             metric='multi_logloss',
                             n_jobs=-1, n_estimators=10000, 
                             class_weight='balanced',
                             
                             max_depth=params['max_depth'], 
                             learning_rate=params['learning_rate'],  
                             colsample_bytree = params['colsample_bytree'], 
                             min_split_gain= params['min_split_gain'], 
                             bagging_freq = params['bagging_freq'],
                             min_child_weight=params['min_child_weight'],
                             num_leaves = params['num_leaves'], 
                             subsample = params['subsample'],
                             reg_alpha= params['reg_alpha'],
                             reg_lambda= params['reg_lambda'],
                             num_class=len(np.unique(y)),
                             bagging_seed=SEED,
                             seed=SEED,
                            )

    kfold = 5
    kf = StratifiedKFold(n_splits=kfold, shuffle=True)
    feat_importance_df  = pd.DataFrame()

    for i, (train_index, test_index) in enumerate(kf.split(train, y)):
        print('='*30, '{} of {} folds'.format(i+1, kfold), '='*30)
        start = time.time()
        X_train, X_val = train.iloc[train_index], train.iloc[test_index]
        y_train, y_val = y.iloc[train_index], y.iloc[test_index]
        clf.fit(X_train, y_train, eval_set=[(X_train, y_train), (X_val, y_val)], eval_metric=evaluate_macroF1_lgb, categorical_feature=categorical_feats,
                early_stopping_rounds=500, verbose=500)
        shap_values = shap.TreeExplainer(clf.booster_).shap_values(X_train)
        fold_importance_df  = pd.DataFrame()
        fold_importance_df['feature'] = X_train.columns
        fold_importance_df['shap_values'] = abs(np.array(shap_values)[:, :].mean(1).mean(0))
        fold_importance_df['feat_imp'] = clf.feature_importances_
        feat_importance_df = pd.concat([feat_importance_df, fold_importance_df])
        print_execution_time(start)

    feat_importance_df_shap = feat_importance_df.groupby('feature').mean().sort_values('shap_values', ascending=False).reset_index()
#     feat_importance_df_shap['shap_cumsum'] = feat_importance_df_shap['shap_values'].cumsum() / feat_importance_df_shap['shap_values'].sum()
#     good_features = feat_importance_df_shap.loc[feat_importance_df_shap['shap_cumsum'] < 0.999].feature
    return feat_importance_df_shap