Quiz tổng hợp

1import optuna
2from xgboost import XGBClassifier
3from sklearn.model_selection import cross_val_score
4from sklearn.datasets import make_classification
5
6X, y = make_classification(n_samples=1000, n_features=20, random_state=42)
7
8def objective(trial):
9    params = {
10        'max_depth': trial.suggest_int('max_depth', 3, 10),
11        'learning_rate': trial.suggest_float('learning_rate', 0.01, 0.3, log=True),
12        'n_estimators': trial.suggest_int('n_estimators', 100, 1000, step=100),
13        'subsample': trial.suggest_float('subsample', 0.6, 1.0),
14        'colsample_bytree': trial.suggest_float('colsample_bytree', 0.6, 1.0),
15        'reg_alpha': trial.suggest_float('reg_alpha', 1e-8, 10.0, log=True),
16        'reg_lambda': trial.suggest_float('reg_lambda', 1e-8, 10.0, log=True),
17        'use_label_encoder': False,
18        'eval_metric': 'logloss',
19    }
20    model = XGBClassifier(**params, random_state=42)
21    score = cross_val_score(model, X, y, cv=5, scoring='f1').mean()
22    return score
23
24study = optuna.create_study(direction='maximize', study_name='xgb_tuning')
25study.optimize(objective, n_trials=50, show_progress_bar=True)
26
27print(f"Best F1 Score: {study.best_value:.4f}")
28print(f"Best Params: {study.best_params}")
29
30# Visualization
31optuna.visualization.plot_optimization_history(study)
32optuna.visualization.plot_param_importances(study)

1from sklearn.ensemble import RandomForestClassifier, StackingClassifier
2from sklearn.linear_model import LogisticRegression
3from sklearn.model_selection import cross_val_score
4from sklearn.datasets import make_classification
5from xgboost import XGBClassifier
6from lightgbm import LGBMClassifier
7
8X, y = make_classification(n_samples=2000, n_features=20, random_state=42)
9
10# Base models
11base_models = [
12    ('rf', RandomForestClassifier(n_estimators=200, max_depth=8, random_state=42)),
13    ('xgb', XGBClassifier(n_estimators=200, max_depth=6, learning_rate=0.1,
14                          use_label_encoder=False, eval_metric='logloss', random_state=42)),
15    ('lgbm', LGBMClassifier(n_estimators=200, max_depth=6, learning_rate=0.1,
16                            random_state=42, verbose=-1)),
17]
18
19# Meta-learner
20meta_learner = LogisticRegression(max_iter=1000)
21
22# Stacking
23stacking_clf = StackingClassifier(
24    estimators=base_models,
25    final_estimator=meta_learner,
26    cv=5,
27    stack_method='predict_proba',
28    n_jobs=-1
29)
30
31# So sánh performance
32models = {
33    'Random Forest': base_models[0][1],
34    'XGBoost': base_models[1][1],
35    'LightGBM': base_models[2][1],
36    'Stacking': stacking_clf,
37}
38
39for name, model in models.items():
40    scores = cross_val_score(model, X, y, cv=5, scoring='f1')
41    print(f"{name:20s} | F1: {scores.mean():.4f} ± {scores.std():.4f}")
42# Stacking thường cho F1 cao hơn từng model đơn lẻ

1import numpy as np
2import pandas as pd
3from scipy import stats
4
5def calculate_psi(expected, actual, bins=10):
6    """Tính Population Stability Index (PSI) giữa 2 phân phối."""
7    breakpoints = np.linspace(
8        min(expected.min(), actual.min()),
9        max(expected.max(), actual.max()),
10        bins + 1
11    )
12    expected_counts = np.histogram(expected, bins=breakpoints)[0] / len(expected)
13    actual_counts = np.histogram(actual, bins=breakpoints)[0] / len(actual)
14
15    # Tránh chia cho 0
16    expected_counts = np.clip(expected_counts, 1e-6, None)
17    actual_counts = np.clip(actual_counts, 1e-6, None)
18
19    psi = np.sum((actual_counts - expected_counts) * np.log(actual_counts / expected_counts))
20    return psi
21
22# Giả lập data
23np.random.seed(42)
24df_train = pd.DataFrame({
25    'age': np.random.normal(35, 10, 5000),
26    'income': np.random.lognormal(10, 1, 5000),
27    'credit_score': np.random.normal(650, 80, 5000),
28})
29df_prod = pd.DataFrame({
30    'age': np.random.normal(32, 12, 1000),       # drift nhẹ
31    'income': np.random.lognormal(10.5, 1.2, 1000),  # drift rõ
32    'credit_score': np.random.normal(645, 85, 1000),  # ổn định
33})
34
35# Kiểm tra drift cho từng feature
36print("=" * 60)
37print(f"{'Feature':15s} | {'PSI':>8s} | {'KS Stat':>8s} | {'p-value':>8s} | Drift?")
38print("=" * 60)
39
40for col in df_train.columns:
41    psi = calculate_psi(df_train[col], df_prod[col])
42    ks_stat, p_value = stats.ks_2samp(df_train[col], df_prod[col])
43    drift = "⚠️ YES" if psi > 0.2 or p_value < 0.05 else "✅ NO"
44    print(f"{col:15s} | {psi:8.4f} | {ks_stat:8.4f} | {p_value:8.4f} | {drift}")
45
46# PSI < 0.1: không có drift
47# 0.1 ≤ PSI < 0.2: drift nhẹ, theo dõi
48# PSI ≥ 0.2: drift đáng kể, cần retrain model