🪪 HOME CREDIT #3:Modeling+Result

# --- [Step 1] 데이터 준비 및 분할 ---
X, y = df.drop(columns='TARGET'), df['TARGET']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
feature_list = cat_cols + num_cols

# --- [Step 2] BinningProcess 설정 및 학습 ---
selection_criteria = {
    "iv": {"min": 0.025, "max": 0.7, "strategy": "highest", "top": 30},
    "quality_score": {"min": 0.01}
}

binning_process = BinningProcess(
    variable_names=feature_list,
    categorical_variables=cat_cols,
    selection_criteria=selection_criteria
)

# Train 데이터로 Rule 학습
binning_process.fit(X_train, y_train)

def remove_high_correlation(df_woe, iv_summary, threshold=0.8):
    corr_matrix = df_woe.corr().abs()
    upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k=1).astype(bool))
    
    to_drop = set() # 중복 제거를 위해 set 사용
    for column in upper.columns:
        high_corr_vars = upper.index[upper[column] > threshold].tolist()
        for var in high_corr_vars:
            # IV 값 추출 시 예외 처리 추가
            iv_col = iv_summary.loc[iv_summary['name'] == column, 'iv'].values[0]
            iv_var = iv_summary.loc[iv_summary['name'] == var, 'iv'].values[0]
            
            # IV가 낮은 쪽을 drop
            drop_target = column if iv_col < iv_var else var
            to_drop.add(drop_target)
    return list(to_drop)

low_iv_corr_vars = remove_high_correlation(X_train_woe, iv_summary, threshold=0.8)
X_train_reduced = X_train_woe.drop(columns=low_iv_corr_vars)

def calculate_vif(df):
    vif_data = pd.DataFrame()
    vif_data["variable"] = df.columns
    # 상수항(intercept)에 의한 영향을 배제하기 위해 보통 WoE 데이터 그대로 사용
    vif_data["VIF"] = [variance_inflation_factor(df.values, i) for i in range(df.shape[1])]
    return vif_data

def iterative_vif_reduction(df, threshold=10):
    while True:
        vif_df = calculate_vif(df)
        max_vif = vif_df['VIF'].max()
        if max_vif > threshold:
            drop_var = vif_df.sort_values('VIF', ascending=False)['variable'].iloc[0]
            print(f"Dropping '{drop_var}' with VIF: {max_vif:.2f}")
            df = df.drop(columns=[drop_var])
        else:
            break
    return df

X_train_final_woe = iterative_vif_reduction(X_train_reduced, threshold=10)
final_selection_list = X_train_final_woe.columns.tolist()

'CODE_GENDER', 'NAME_INCOME_TYPE', 'NAME_EDUCATION_TYPE', 'OCCUPATION_TYPE', 'ORGANIZATION_TYPE', 'REGION_RATING_CLIENT_W_CITY', 'REG_CITY_NOT_WORK_CITY', 'FLAG_DOCUMENT_3', 'DAYS_EMPLOYED_ANOM', 'AMT_ANNUITY', 'AMT_GOODS_PRICE', 'REGION_POPULATION_RELATIVE', 'DAYS_REGISTRATION', 'EXT_SOURCE_1', 'EXT_SOURCE_2', 'EXT_SOURCE_3', 'LIVINGAREA_AVG', 'DAYS_LAST_PHONE_CHANGE', 'AGE', 'YEARS_EMPLOYED', 'PAYMENT_RATE'

Test AUC: 0.7491
Test Gini: 0.4983
Test K-S Statistic: 0.3726

Optimal Threshold: 0.4918

[Optimized Classification Report]
              precision    recall  f1-score   support

           0       0.96      0.67      0.79     56538
           1       0.16      0.70      0.26      4965

    accuracy                           0.67     61503
   macro avg       0.56      0.69      0.52     61503
weighted avg       0.90      0.67      0.75     61503


[Decile Analysis]
        count  bad_count  bad_rate
decile                            
0        6151       1634  0.265648
1        6150        860  0.139837
2        6150        675  0.109756
3        6150        473  0.076911
4        6151        411  0.066818
5        6150        267  0.043415
6        6150        226  0.036748
7        6150        200  0.032520
8        6150        137  0.022276
9        6151         82  0.013331

params = {
    'objective': 'binary:logistic',
    'eval_metric': 'auc',
    'learning_rate': 0.05,
    'max_depth': 4,
    'subsample': 0.8,
    'colsample_bytree': 0.8,
    'random_state': 42,
    'scale_pos_weight': (y_train == 0).sum() / (y_train == 1).sum() # 클래스 불균형 해소
}

[XGBoost Final Performance]
Test AUC: 0.7570
Test Gini: 0.5140
Test K-S Statistic: 0.3806

Optimal Threshold (Youden's J): 0.4888

[Optimized Classification Report]
              precision    recall  f1-score   support

           0       0.96      0.69      0.80     56538
           1       0.16      0.69      0.26      4965

    accuracy                           0.69     61503
   macro avg       0.56      0.69      0.53     61503
weighted avg       0.90      0.69      0.76     61503


[Decile Analysis - Risk Ranking]
        count  bad_count  bad_rate
decile                            
9        6151       1680  0.273126
8        6150        899  0.146179
7        6150        638  0.103740
6        6150        477  0.077561
5        6150        376  0.061138
4        6151        306  0.049748
3        6150        204  0.033171
2        6150        196  0.031870
1        6150        113  0.018374
0        6151         76  0.012356