↑ 关注 + 星标 ，每天学Python新技能
后台回复【大礼包】送你Python自学大礼包

LightGBM参数概述

1. 影响决策树结构和学习的参数
2. 影响训练速度的参数
3. 提高精度的参数
4. 防止过拟合的参数

1、控制树结构的超参数

提高准确性的超参数

更多超参数来控制过拟合

在 Optuna 中创建搜索网格

• 字典形式的参数网格
• 创建一个模型（可以配合交叉验证kfold）来尝试超参数组合集
• 用于模型训练的数据集
• 使用此模型生成预测
• 根据用户定义的指标对预测进行评分并返回

import optuna  # pip install optuna
from sklearn.metrics import log_loss
from sklearn.model_selection import StratifiedKFold

def objective(trial, X, y):
    # 后面填充
    param_grid = {}
    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=1121218)

    cv_scores = np.empty(5)
    for idx, (train_idx, test_idx) in enumerate(cv.split(X, y)):
        X_train, X_test = X.iloc[train_idx], X.iloc[test_idx]
        y_train, y_test = y[train_idx], y[test_idx]

        model = lgbm.LGBMClassifier(objective="binary", **param_grid)
        model.fit(
            X_train,
            y_train,
            eval_set=[(X_test, y_test)],
            eval_metric="binary_logloss",
            early_stopping_rounds=100,
        )
        preds = model.predict_proba(X_test)
        cv_scores[idx] = preds

    return np.mean(cv_scores)

def objective(trial, X, y):
    # 字典形式的参数网格
    param_grid = {
        "n_estimators": trial.suggest_categorical("n_estimators", [10000]),
        "learning_rate": trial.suggest_float("learning_rate", 0.01, 0.3),
        "num_leaves": trial.suggest_int("num_leaves", 20, 3000, step=20),
        "max_depth": trial.suggest_int("max_depth", 3, 12),
        "min_data_in_leaf": trial.suggest_int("min_data_in_leaf", 200, 10000, step=100),
        "max_bin": trial.suggest_int("max_bin", 200, 300),
        "lambda_l1": trial.suggest_int("lambda_l1", 0, 100, step=5),
        "lambda_l2": trial.suggest_int("lambda_l2", 0, 100, step=5),
        "min_gain_to_split": trial.suggest_float("min_gain_to_split", 0, 15),
        "bagging_fraction": trial.suggest_float(
            "bagging_fraction", 0.2, 0.95, step=0.1
        ),
        "bagging_freq": trial.suggest_categorical("bagging_freq", [1]),
        "feature_fraction": trial.suggest_float(
            "feature_fraction", 0.2, 0.95, step=0.1
        ),
    }

创建 Optuna 自动调起来

from optuna.integration import LightGBMPruningCallback

def objective(trial, X, y):
    # 参数网格
    param_grid = {
        "n_estimators": trial.suggest_categorical("n_estimators", [10000]),
        "learning_rate": trial.suggest_float("learning_rate", 0.01, 0.3),
        "num_leaves": trial.suggest_int("num_leaves", 20, 3000, step=20),
        "max_depth": trial.suggest_int("max_depth", 3, 12),
        "min_data_in_leaf": trial.suggest_int("min_data_in_leaf", 200, 10000, step=100),
        "lambda_l1": trial.suggest_int("lambda_l1", 0, 100, step=5),
        "lambda_l2": trial.suggest_int("lambda_l2", 0, 100, step=5),
        "min_gain_to_split": trial.suggest_float("min_gain_to_split", 0, 15),
        "bagging_fraction": trial.suggest_float("bagging_fraction", 0.2, 0.95, step=0.1),
        "bagging_freq": trial.suggest_categorical("bagging_freq", [1]),
        "feature_fraction": trial.suggest_float("feature_fraction", 0.2, 0.95, step=0.1),
        "random_state": 2021,
    }
    # 5折交叉验证
    cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=1121218)

    cv_scores = np.empty(5)
    for idx, (train_idx, test_idx) in enumerate(cv.split(X, y)):
        X_train, X_test = X.iloc[train_idx], X.iloc[test_idx]
        y_train, y_test = y[train_idx], y[test_idx]
        
        # LGBM建模
        model = lgbm.LGBMClassifier(objective="binary", **param_grid)
        model.fit(
            X_train,
            y_train,
            eval_set=[(X_test, y_test)],
            eval_metric="binary_logloss",
            early_stopping_rounds=100,
            callbacks=[
                LightGBMPruningCallback(trial, "binary_logloss")
            ],
        )
        # 模型预测
        preds = model.predict_proba(X_test)
        # 优化指标logloss最小
        cv_scores[idx] = log_loss(y_test, preds)

    return np.mean(cv_scores)

study = optuna.create_study(direction="minimize", study_name="LGBM Classifier")
func = lambda trial: objective(trial, X, y)
study.optimize(func, n_trials=20)

print(f"\tBest value (rmse): {study.best_value:.5f}")
print(f"\tBest params:")

for key, value in study.best_params.items():
    print(f"\t\t{key}: {value}")
    
-----------------------------------------------------
Best value (binary_logloss): 0.35738
 Best params:
  device: gpu
  lambda_l1: 7.71800699380605e-05
  lambda_l2: 4.17890272377219e-06
  bagging_fraction: 0.7000000000000001
  feature_fraction: 0.4
  bagging_freq: 5
  max_depth: 5
  num_leaves: 1007
  min_data_in_leaf: 45
  min_split_gain: 15.703519227860273
  learning_rate: 0.010784015325759629
  n_estimators: 10000

得到这个参数组合后，我们就可以拿去跑模型了，看结果再手动微调，这样就可以省很多时间了。

结语

关于Optuna的强大之处，后面会对比同类的调参工具介绍，敬请期待。

推荐阅读
程序员在深圳买千万豪宅，被领导知道后，被打2星！
个个年薪都五十万的数据分析师，是真的吗？
打架打出来的牛逼城市！
985高校的学生“夫妻宿舍”令人羡慕，网友：这福利心动了，我要去读博！
微软这个太强了

推荐一个公众号，帮助程序员自学与成长