XGBoost/LightGBM 梯度提升¶

pip install xgboost lightgbm                          # 安装两个库
pip install catboost                                  # 也可装 CatBoost（第三个选择）

import xgboost as xgb                                 # 导入 XGBoost
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

# 准备数据
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

# Scikit-learn API（推荐入门）
model = xgb.XGBClassifier(
    n_estimators=200,                                 # 树的数量
    max_depth=6,                                      # 最大深度
    learning_rate=0.1,                                # 学习率
    subsample=0.8,                                    # 行采样比例
    colsample_bytree=0.8,                             # 列采样比例
    eval_metric='logloss',                            # 评估指标
    random_state=42,
    n_jobs=-1                                         # 使用所有 CPU
)

# 训练（带早停）
model.fit(
    X_train, y_train,
    eval_set=[(X_test, y_test)],                      # 验证集
    verbose=10                                        # 每 10 轮打印
)

# 预测
y_pred = model.predict(X_test)                        # 预测类别
y_prob = model.predict_proba(X_test)[:, 1]            # 预测概率
print(f"准确率: {accuracy_score(y_test, y_pred):.4f}")

import lightgbm as lgb                                # 导入 LightGBM

# Scikit-learn API
model = lgb.LGBMClassifier(
    n_estimators=200,                                 # 树的数量
    max_depth=-1,                                     # 不限深度（靠叶子数控制）
    num_leaves=31,                                    # 叶子节点数（关键参数）
    learning_rate=0.1,                                # 学习率
    subsample=0.8,                                    # 行采样
    colsample_bytree=0.8,                             # 列采样
    random_state=42,
    n_jobs=-1
)

# 训练
model.fit(
    X_train, y_train,
    eval_set=[(X_test, y_test)],                      # 验证集
    callbacks=[lgb.early_stopping(50),                # 50 轮不提升则停
               lgb.log_evaluation(10)]                # 每 10 轮打印
)

y_pred = model.predict(X_test)
print(f"准确率: {accuracy_score(y_test, y_pred):.4f}")

import matplotlib.pyplot as plt

# XGBoost 特征重要性
xgb.plot_importance(model, max_num_features=20, importance_type='gain')
plt.tight_layout()
plt.savefig('xgb_importance.png', dpi=300)

# LightGBM 特征重要性
lgb.plot_importance(model, max_num_features=20, importance_type='gain')
plt.tight_layout()
plt.savefig('lgb_importance.png', dpi=300)

import optuna                                         # 导入 Optuna

def objective(trial):
    params = {
        'n_estimators': trial.suggest_int('n_estimators', 100, 500),
        'max_depth': trial.suggest_int('max_depth', 3, 12),
        'learning_rate': trial.suggest_float('learning_rate', 0.01, 0.3, log=True),
        'subsample': trial.suggest_float('subsample', 0.6, 1.0),
        'colsample_bytree': trial.suggest_float('colsample_bytree', 0.6, 1.0),
        'num_leaves': trial.suggest_int('num_leaves', 20, 100),
    }
    model = lgb.LGBMClassifier(**params, random_state=42, n_jobs=-1)
    model.fit(X_train, y_train, eval_set=[(X_test, y_test)],
              callbacks=[lgb.early_stopping(50), lgb.log_evaluation(0)])
    y_pred = model.predict(X_test)
    return accuracy_score(y_test, y_pred)

study = optuna.create_study(direction='maximize')     # 最大化准确率
study.optimize(objective, n_trials=100)               # 搜索 100 次
print(f"最优参数: {study.best_params}")

# === XGBoost vs LightGBM 选择 ===
# XGBoost: 精度略高、社区更大、GPU 支持好
# LightGBM: 训练更快、内存更省、大数据更适合
# CatBoost: 类别特征自动处理、最不容易过拟合

# === 关键参数 ===
# n_estimators:     树的数量（100-500）
# learning_rate:    学习率（0.01-0.3）
# max_depth:        最大深度（XGB: 3-10, LGB: -1 不限）
# num_leaves:       叶子数（LGB 专有, 20-100）
# subsample:        行采样（0.6-1.0）
# colsample_bytree: 列采样（0.6-1.0）
# early_stopping:   早停轮数（30-100）

# === 核心流程 ===
# 1. 数据划分 → 2. 创建模型 → 3. fit(早停) → 4. predict → 5. 调参