实战对比用Sklearn接口搞定XGBoost和LightGBM的交叉验证与特征筛选附完整代码在机器学习项目实践中XGBoost和LightGBM作为当前最强大的梯度提升框架已经成为解决结构化数据问题的标配工具。但对于习惯Scikit-learn工作流的开发者来说这两个框架的原生API总显得有些格格不入——直到它们提供了完整的Sklearn兼容接口。本文将带您用完全Sklearn风格的代码实现从交叉验证、网格搜索到特征筛选的全流程实战让您在不改变现有工作习惯的前提下充分发挥这两个框架的性能优势。1. 环境配置与基础用法对比1.1 安装与初始化设置两种框架的安装都非常简单但需要注意版本兼容性pip install xgboost1.7.0 lightgbm3.3.5 scikit-learn1.2.2初始化模型时XGBoost和LightGBM的Sklearn接口几乎可以无缝替换# XGBoost分类器 from xgboost import XGBClassifier xgb_clf XGBClassifier(n_estimators100, learning_rate0.1, max_depth3) # LightGBM分类器 from lightgbm import LGBMClassifier lgb_clf LGBMClassifier(n_estimators100, learning_rate0.1, max_depth3)关键参数对照表参数功能XGBoost参数名LightGBM参数名典型值范围树的数量n_estimatorsn_estimators50-1000学习率learning_ratelearning_rate0.01-0.3最大深度max_depthmax_depth3-10特征采样比例colsample_bytreefeature_fraction0.6-1.0样本采样比例subsamplebagging_fraction0.6-1.01.2 数据准备与基础训练我们以乳腺癌数据集为例展示基础训练流程from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split data load_breast_cancer() X, y data.data, data.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # XGBoost训练 xgb_clf.fit(X_train, y_train) print(fXGBoost测试集准确率: {xgb_clf.score(X_test, y_test):.4f}) # LightGBM训练 lgb_clf.fit(X_train, y_train) print(fLightGBM测试集准确率: {lgb_clf.score(X_test, y_test):.4f})注意LightGBM默认会自动处理类别型特征而XGBoost需要手动进行独热编码等处理2. 交叉验证实战对比2.1 使用cross_val_score的基础验证最直接的方式是使用Sklearn的cross_val_scorefrom sklearn.model_selection import cross_val_score # 5折交叉验证 xgb_scores cross_val_score(xgb_clf, X, y, cv5, scoringaccuracy) print(fXGBoost交叉验证平均准确率: {xgb_scores.mean():.4f} (±{xgb_scores.std():.4f})) lgb_scores cross_val_score(lgb_clf, X, y, cv5, scoringaccuracy) print(fLightGBM交叉验证平均准确率: {lgb_scores.mean():.4f} (±{lgb_scores.std():.4f}))2.2 自定义交叉验证策略对于需要更复杂验证策略的场景可以使用KFold自定义from sklearn.model_selection import KFold import numpy as np kf KFold(n_splits5, shuffleTrue, random_state42) xgb_scores, lgb_scores [], [] for train_idx, val_idx in kf.split(X): X_train, X_val X[train_idx], X[val_idx] y_train, y_val y[train_idx], y[val_idx] xgb_clf.fit(X_train, y_train) xgb_scores.append(xgb_clf.score(X_val, y_val)) lgb_clf.fit(X_train, y_train) lgb_scores.append(lgb_clf.score(X_val, y_val)) print(fXGBoost自定义交叉验证结果: {np.mean(xgb_scores):.4f}) print(fLightGBM自定义交叉验证结果: {np.mean(lgb_scores):.4f})2.3 早停法(early stopping)实现两种框架都支持早停机制但实现方式略有差异# XGBoost早停 xgb_clf.fit(X_train, y_train, eval_set[(X_test, y_test)], early_stopping_rounds10, verboseTrue) # LightGBM早停 lgb_clf.fit(X_train, y_train, eval_set[(X_test, y_test)], early_stopping_rounds10, eval_metricaccuracy, verbose10)提示早停轮数(early_stopping_rounds)通常设为10-50太小可能导致提前终止太大则失去早停意义3. 网格搜索超参数优化3.1 基础网格搜索实现使用GridSearchCV进行参数调优from sklearn.model_selection import GridSearchCV # XGBoost参数网格 xgb_params { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1, 0.2], n_estimators: [100, 200], subsample: [0.6, 0.8, 1.0] } # LightGBM参数网格 lgb_params { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1, 0.2], n_estimators: [100, 200], feature_fraction: [0.6, 0.8, 1.0] } # 并行网格搜索 xgb_grid GridSearchCV(xgb_clf, xgb_params, cv5, n_jobs-1, verbose1) xgb_grid.fit(X_train, y_train) lgb_grid GridSearchCV(lgb_clf, lgb_params, cv5, n_jobs-1, verbose1) lgb_grid.fit(X_train, y_train) print(fXGBoost最佳参数: {xgb_grid.best_params_}) print(fLightGBM最佳参数: {lgb_grid.best_params_})3.2 随机搜索与贝叶斯优化对于大型参数空间推荐使用更高效的搜索方法from sklearn.model_selection import RandomizedSearchCV from scipy.stats import uniform, randint # XGBoost随机搜索 xgb_random RandomizedSearchCV( xgb_clf, { max_depth: randint(3, 10), learning_rate: uniform(0.01, 0.3), n_estimators: randint(100, 500), subsample: uniform(0.6, 0.4) }, n_iter20, cv5, n_jobs-1 ) xgb_random.fit(X_train, y_train)对于追求极致性能的场景可以尝试贝叶斯优化库如optunaimport optuna def objective(trial): params { max_depth: trial.suggest_int(max_depth, 3, 10), learning_rate: trial.suggest_float(learning_rate, 0.01, 0.3), n_estimators: trial.suggest_int(n_estimators, 100, 500), subsample: trial.suggest_float(subsample, 0.6, 1.0) } model XGBClassifier(**params) return cross_val_score(model, X_train, y_train, cv5, scoringaccuracy).mean() study optuna.create_study(directionmaximize) study.optimize(objective, n_trials50)4. 特征工程与筛选实战4.1 基于重要性的特征筛选两种框架都提供了特征重要性评估import matplotlib.pyplot as plt # XGBoost特征重要性 plt.figure(figsize(10, 6)) xgb.plot_importance(xgb_grid.best_estimator_, height0.8) plt.title(XGBoost特征重要性) plt.show() # LightGBM特征重要性 plt.figure(figsize(10, 6)) lgb.plot_importance(lgb_grid.best_estimator_, height0.8) plt.title(LightGBM特征重要性) plt.show()4.2 使用SelectFromModel自动化筛选Sklearn的SelectFromModel可以基于重要性阈值自动筛选特征from sklearn.feature_selection import SelectFromModel # XGBoost特征选择 selector SelectFromModel(xgb_grid.best_estimator_, thresholdmedian) X_train_selected selector.fit_transform(X_train, y_train) print(f原始特征数: {X_train.shape[1]}, 筛选后特征数: {X_train_selected.shape[1]}) # 获取被选中的特征索引 selected_features selector.get_support(indicesTrue) print(重要特征:, [data.feature_names[i] for i in selected_features])4.3 特征筛选与模型训练的Pipeline将特征筛选和模型训练整合到Pipeline中from sklearn.pipeline import Pipeline # 构建XGBoost完整流程 xgb_pipe Pipeline([ (feature_selection, SelectFromModel(XGBClassifier())), (classification, XGBClassifier()) ]) # 设置网格搜索参数 param_grid { feature_selection__threshold: [mean, median], classification__max_depth: [3, 5], classification__learning_rate: [0.1, 0.2] } # 执行网格搜索 grid_search GridSearchCV(xgb_pipe, param_grid, cv5, n_jobs-1) grid_search.fit(X_train, y_train)4.4 特征筛选对模型性能的影响我们通过实验验证特征筛选的效果original_score cross_val_score(xgb_clf, X_train, y_train, cv5).mean() selected_score cross_val_score(xgb_clf, X_train_selected, y_train, cv5).mean() print(f原始特征准确率: {original_score:.4f}) print(f筛选后特征准确率: {selected_score:.4f}) print(f特征数量减少: {100*(1-X_train_selected.shape[1]/X_train.shape[1]):.1f}%)典型结果可能显示特征数量减少30-50%模型准确率基本持平或略有提升训练速度显著提高5. 完整项目实战案例5.1 房价预测回归问题以波士顿房价数据集为例展示完整流程from sklearn.datasets import fetch_california_housing from sklearn.metrics import mean_squared_error # 加载数据 housing fetch_california_housing() X, y housing.data, housing.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # 构建回归模型 xgb_reg XGBRegressor(objectivereg:squarederror) lgb_reg LGBMRegressor(objectiveregression) # 参数网格 param_grid { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1], n_estimators: [100, 200] } # 网格搜索 xgb_grid GridSearchCV(xgb_reg, param_grid, cv5, scoringneg_mean_squared_error) xgb_grid.fit(X_train, y_train) lgb_grid GridSearchCV(lgb_reg, param_grid, cv5, scoringneg_mean_squared_error) lgb_grid.fit(X_train, y_train) # 评估结果 xgb_pred xgb_grid.predict(X_test) lgb_pred lgb_grid.predict(X_test) print(fXGBoost RMSE: {np.sqrt(mean_squared_error(y_test, xgb_pred)):.4f}) print(fLightGBM RMSE: {np.sqrt(mean_squared_error(y_test, lgb_pred)):.4f})5.2 分类任务完整流程以手写数字识别为例from sklearn.datasets import load_digits digits load_digits() X, y digits.data, digits.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # 构建Pipeline pipeline Pipeline([ (feature_selection, SelectFromModel(LGBMClassifier())), (classification, XGBClassifier()) ]) # 训练与评估 pipeline.fit(X_train, y_train) print(f测试集准确率: {pipeline.score(X_test, y_test):.4f}) # 可视化特征重要性 selected_features pipeline.named_steps[feature_selection].get_support() print(f原始特征数: {X.shape[1]}, 筛选后特征数: {sum(selected_features)})6. 性能优化与高级技巧6.1 内存与计算效率优化对于大型数据集这些技巧可以显著提升效率# XGBoost内存优化配置 xgb_fast XGBClassifier( tree_methodhist, # 使用直方图算法 predictorcpu_predictor, # CPU预测 enable_categoricalTrue, # 自动处理类别特征 n_jobs-1 # 使用所有CPU核心 ) # LightGBM内存优化配置 lgb_fast LGBMClassifier( boosting_typegoss, # 使用GOSS算法 max_bin255, # 减少直方图分箱数 n_jobs-1 )6.2 类别不平衡处理对于不平衡数据集可以调整这些参数# 计算类别权重 from sklearn.utils.class_weight import compute_class_weight classes np.unique(y_train) weights compute_class_weight(balanced, classesclasses, yy_train) class_weights dict(zip(classes, weights)) # 应用权重 xgb_balanced XGBClassifier(scale_pos_weightclass_weights[1]/class_weights[0]) lgb_balanced LGBMClassifier(class_weightclass_weights)6.3 自定义评估指标两种框架都支持自定义评估函数from sklearn.metrics import f1_score def custom_f1(y_true, y_pred): return f1, f1_score(y_true, y_pred, averagemacro), True # XGBoost使用自定义指标 xgb_clf.fit(X_train, y_train, eval_metriccustom_f1, eval_set[(X_test, y_test)])7. 模型部署与生产化建议7.1 模型持久化与加载保存和加载训练好的模型import joblib # 保存XGBoost模型 joblib.dump(xgb_grid.best_estimator_, xgb_model.pkl) # 保存LightGBM模型 joblib.dump(lgb_grid.best_estimator_, lgb_model.pkl) # 加载模型 xgb_loaded joblib.load(xgb_model.pkl) lgb_loaded joblib.load(lgb_model.pkl)7.2 转换为ONNX格式对于生产环境可以转换为ONNX格式from onnxmltools.convert import convert_xgboost from onnxruntime import InferenceSession # 转换XGBoost模型 onnx_model convert_xgboost(xgb_loaded, initial_types[(input, FloatTensorType([None, X.shape[1]]))]) # 保存ONNX模型 with open(model.onnx, wb) as f: f.write(onnx_model.SerializeToString()) # 加载运行 session InferenceSession(model.onnx) inputs {input: X_test.astype(np.float32)} outputs session.run(None, inputs)7.3 模型监控与更新建立模型性能监控机制from datetime import datetime def log_performance(model, X, y, model_name): preds model.predict(X) accuracy (preds y).mean() with open(performance.log, a) as f: f.write(f{datetime.now()},{model_name},{accuracy:.4f}\n) # 记录性能 log_performance(xgb_loaded, X_test, y_test, xgb_prod) log_performance(lgb_loaded, X_test, y_test, lgb_prod)在实际项目中XGBoost和LightGBM的Sklearn接口可以完美融入现有的机器学习工作流。从我的经验来看当特征维度较高时LightGBM通常训练速度更快而在中小型数据集上XGBoost有时能获得略好的性能。关键是根据具体问题选择合适的工具并充分利用交叉验证和特征筛选来构建稳健的模型。
实战对比：用Sklearn接口搞定XGBoost和LightGBM的交叉验证与特征筛选（附完整代码）
实战对比用Sklearn接口搞定XGBoost和LightGBM的交叉验证与特征筛选附完整代码在机器学习项目实践中XGBoost和LightGBM作为当前最强大的梯度提升框架已经成为解决结构化数据问题的标配工具。但对于习惯Scikit-learn工作流的开发者来说这两个框架的原生API总显得有些格格不入——直到它们提供了完整的Sklearn兼容接口。本文将带您用完全Sklearn风格的代码实现从交叉验证、网格搜索到特征筛选的全流程实战让您在不改变现有工作习惯的前提下充分发挥这两个框架的性能优势。1. 环境配置与基础用法对比1.1 安装与初始化设置两种框架的安装都非常简单但需要注意版本兼容性pip install xgboost1.7.0 lightgbm3.3.5 scikit-learn1.2.2初始化模型时XGBoost和LightGBM的Sklearn接口几乎可以无缝替换# XGBoost分类器 from xgboost import XGBClassifier xgb_clf XGBClassifier(n_estimators100, learning_rate0.1, max_depth3) # LightGBM分类器 from lightgbm import LGBMClassifier lgb_clf LGBMClassifier(n_estimators100, learning_rate0.1, max_depth3)关键参数对照表参数功能XGBoost参数名LightGBM参数名典型值范围树的数量n_estimatorsn_estimators50-1000学习率learning_ratelearning_rate0.01-0.3最大深度max_depthmax_depth3-10特征采样比例colsample_bytreefeature_fraction0.6-1.0样本采样比例subsamplebagging_fraction0.6-1.01.2 数据准备与基础训练我们以乳腺癌数据集为例展示基础训练流程from sklearn.datasets import load_breast_cancer from sklearn.model_selection import train_test_split data load_breast_cancer() X, y data.data, data.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # XGBoost训练 xgb_clf.fit(X_train, y_train) print(fXGBoost测试集准确率: {xgb_clf.score(X_test, y_test):.4f}) # LightGBM训练 lgb_clf.fit(X_train, y_train) print(fLightGBM测试集准确率: {lgb_clf.score(X_test, y_test):.4f})注意LightGBM默认会自动处理类别型特征而XGBoost需要手动进行独热编码等处理2. 交叉验证实战对比2.1 使用cross_val_score的基础验证最直接的方式是使用Sklearn的cross_val_scorefrom sklearn.model_selection import cross_val_score # 5折交叉验证 xgb_scores cross_val_score(xgb_clf, X, y, cv5, scoringaccuracy) print(fXGBoost交叉验证平均准确率: {xgb_scores.mean():.4f} (±{xgb_scores.std():.4f})) lgb_scores cross_val_score(lgb_clf, X, y, cv5, scoringaccuracy) print(fLightGBM交叉验证平均准确率: {lgb_scores.mean():.4f} (±{lgb_scores.std():.4f}))2.2 自定义交叉验证策略对于需要更复杂验证策略的场景可以使用KFold自定义from sklearn.model_selection import KFold import numpy as np kf KFold(n_splits5, shuffleTrue, random_state42) xgb_scores, lgb_scores [], [] for train_idx, val_idx in kf.split(X): X_train, X_val X[train_idx], X[val_idx] y_train, y_val y[train_idx], y[val_idx] xgb_clf.fit(X_train, y_train) xgb_scores.append(xgb_clf.score(X_val, y_val)) lgb_clf.fit(X_train, y_train) lgb_scores.append(lgb_clf.score(X_val, y_val)) print(fXGBoost自定义交叉验证结果: {np.mean(xgb_scores):.4f}) print(fLightGBM自定义交叉验证结果: {np.mean(lgb_scores):.4f})2.3 早停法(early stopping)实现两种框架都支持早停机制但实现方式略有差异# XGBoost早停 xgb_clf.fit(X_train, y_train, eval_set[(X_test, y_test)], early_stopping_rounds10, verboseTrue) # LightGBM早停 lgb_clf.fit(X_train, y_train, eval_set[(X_test, y_test)], early_stopping_rounds10, eval_metricaccuracy, verbose10)提示早停轮数(early_stopping_rounds)通常设为10-50太小可能导致提前终止太大则失去早停意义3. 网格搜索超参数优化3.1 基础网格搜索实现使用GridSearchCV进行参数调优from sklearn.model_selection import GridSearchCV # XGBoost参数网格 xgb_params { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1, 0.2], n_estimators: [100, 200], subsample: [0.6, 0.8, 1.0] } # LightGBM参数网格 lgb_params { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1, 0.2], n_estimators: [100, 200], feature_fraction: [0.6, 0.8, 1.0] } # 并行网格搜索 xgb_grid GridSearchCV(xgb_clf, xgb_params, cv5, n_jobs-1, verbose1) xgb_grid.fit(X_train, y_train) lgb_grid GridSearchCV(lgb_clf, lgb_params, cv5, n_jobs-1, verbose1) lgb_grid.fit(X_train, y_train) print(fXGBoost最佳参数: {xgb_grid.best_params_}) print(fLightGBM最佳参数: {lgb_grid.best_params_})3.2 随机搜索与贝叶斯优化对于大型参数空间推荐使用更高效的搜索方法from sklearn.model_selection import RandomizedSearchCV from scipy.stats import uniform, randint # XGBoost随机搜索 xgb_random RandomizedSearchCV( xgb_clf, { max_depth: randint(3, 10), learning_rate: uniform(0.01, 0.3), n_estimators: randint(100, 500), subsample: uniform(0.6, 0.4) }, n_iter20, cv5, n_jobs-1 ) xgb_random.fit(X_train, y_train)对于追求极致性能的场景可以尝试贝叶斯优化库如optunaimport optuna def objective(trial): params { max_depth: trial.suggest_int(max_depth, 3, 10), learning_rate: trial.suggest_float(learning_rate, 0.01, 0.3), n_estimators: trial.suggest_int(n_estimators, 100, 500), subsample: trial.suggest_float(subsample, 0.6, 1.0) } model XGBClassifier(**params) return cross_val_score(model, X_train, y_train, cv5, scoringaccuracy).mean() study optuna.create_study(directionmaximize) study.optimize(objective, n_trials50)4. 特征工程与筛选实战4.1 基于重要性的特征筛选两种框架都提供了特征重要性评估import matplotlib.pyplot as plt # XGBoost特征重要性 plt.figure(figsize(10, 6)) xgb.plot_importance(xgb_grid.best_estimator_, height0.8) plt.title(XGBoost特征重要性) plt.show() # LightGBM特征重要性 plt.figure(figsize(10, 6)) lgb.plot_importance(lgb_grid.best_estimator_, height0.8) plt.title(LightGBM特征重要性) plt.show()4.2 使用SelectFromModel自动化筛选Sklearn的SelectFromModel可以基于重要性阈值自动筛选特征from sklearn.feature_selection import SelectFromModel # XGBoost特征选择 selector SelectFromModel(xgb_grid.best_estimator_, thresholdmedian) X_train_selected selector.fit_transform(X_train, y_train) print(f原始特征数: {X_train.shape[1]}, 筛选后特征数: {X_train_selected.shape[1]}) # 获取被选中的特征索引 selected_features selector.get_support(indicesTrue) print(重要特征:, [data.feature_names[i] for i in selected_features])4.3 特征筛选与模型训练的Pipeline将特征筛选和模型训练整合到Pipeline中from sklearn.pipeline import Pipeline # 构建XGBoost完整流程 xgb_pipe Pipeline([ (feature_selection, SelectFromModel(XGBClassifier())), (classification, XGBClassifier()) ]) # 设置网格搜索参数 param_grid { feature_selection__threshold: [mean, median], classification__max_depth: [3, 5], classification__learning_rate: [0.1, 0.2] } # 执行网格搜索 grid_search GridSearchCV(xgb_pipe, param_grid, cv5, n_jobs-1) grid_search.fit(X_train, y_train)4.4 特征筛选对模型性能的影响我们通过实验验证特征筛选的效果original_score cross_val_score(xgb_clf, X_train, y_train, cv5).mean() selected_score cross_val_score(xgb_clf, X_train_selected, y_train, cv5).mean() print(f原始特征准确率: {original_score:.4f}) print(f筛选后特征准确率: {selected_score:.4f}) print(f特征数量减少: {100*(1-X_train_selected.shape[1]/X_train.shape[1]):.1f}%)典型结果可能显示特征数量减少30-50%模型准确率基本持平或略有提升训练速度显著提高5. 完整项目实战案例5.1 房价预测回归问题以波士顿房价数据集为例展示完整流程from sklearn.datasets import fetch_california_housing from sklearn.metrics import mean_squared_error # 加载数据 housing fetch_california_housing() X, y housing.data, housing.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # 构建回归模型 xgb_reg XGBRegressor(objectivereg:squarederror) lgb_reg LGBMRegressor(objectiveregression) # 参数网格 param_grid { max_depth: [3, 5, 7], learning_rate: [0.01, 0.1], n_estimators: [100, 200] } # 网格搜索 xgb_grid GridSearchCV(xgb_reg, param_grid, cv5, scoringneg_mean_squared_error) xgb_grid.fit(X_train, y_train) lgb_grid GridSearchCV(lgb_reg, param_grid, cv5, scoringneg_mean_squared_error) lgb_grid.fit(X_train, y_train) # 评估结果 xgb_pred xgb_grid.predict(X_test) lgb_pred lgb_grid.predict(X_test) print(fXGBoost RMSE: {np.sqrt(mean_squared_error(y_test, xgb_pred)):.4f}) print(fLightGBM RMSE: {np.sqrt(mean_squared_error(y_test, lgb_pred)):.4f})5.2 分类任务完整流程以手写数字识别为例from sklearn.datasets import load_digits digits load_digits() X, y digits.data, digits.target X_train, X_test, y_train, y_test train_test_split(X, y, test_size0.2, random_state42) # 构建Pipeline pipeline Pipeline([ (feature_selection, SelectFromModel(LGBMClassifier())), (classification, XGBClassifier()) ]) # 训练与评估 pipeline.fit(X_train, y_train) print(f测试集准确率: {pipeline.score(X_test, y_test):.4f}) # 可视化特征重要性 selected_features pipeline.named_steps[feature_selection].get_support() print(f原始特征数: {X.shape[1]}, 筛选后特征数: {sum(selected_features)})6. 性能优化与高级技巧6.1 内存与计算效率优化对于大型数据集这些技巧可以显著提升效率# XGBoost内存优化配置 xgb_fast XGBClassifier( tree_methodhist, # 使用直方图算法 predictorcpu_predictor, # CPU预测 enable_categoricalTrue, # 自动处理类别特征 n_jobs-1 # 使用所有CPU核心 ) # LightGBM内存优化配置 lgb_fast LGBMClassifier( boosting_typegoss, # 使用GOSS算法 max_bin255, # 减少直方图分箱数 n_jobs-1 )6.2 类别不平衡处理对于不平衡数据集可以调整这些参数# 计算类别权重 from sklearn.utils.class_weight import compute_class_weight classes np.unique(y_train) weights compute_class_weight(balanced, classesclasses, yy_train) class_weights dict(zip(classes, weights)) # 应用权重 xgb_balanced XGBClassifier(scale_pos_weightclass_weights[1]/class_weights[0]) lgb_balanced LGBMClassifier(class_weightclass_weights)6.3 自定义评估指标两种框架都支持自定义评估函数from sklearn.metrics import f1_score def custom_f1(y_true, y_pred): return f1, f1_score(y_true, y_pred, averagemacro), True # XGBoost使用自定义指标 xgb_clf.fit(X_train, y_train, eval_metriccustom_f1, eval_set[(X_test, y_test)])7. 模型部署与生产化建议7.1 模型持久化与加载保存和加载训练好的模型import joblib # 保存XGBoost模型 joblib.dump(xgb_grid.best_estimator_, xgb_model.pkl) # 保存LightGBM模型 joblib.dump(lgb_grid.best_estimator_, lgb_model.pkl) # 加载模型 xgb_loaded joblib.load(xgb_model.pkl) lgb_loaded joblib.load(lgb_model.pkl)7.2 转换为ONNX格式对于生产环境可以转换为ONNX格式from onnxmltools.convert import convert_xgboost from onnxruntime import InferenceSession # 转换XGBoost模型 onnx_model convert_xgboost(xgb_loaded, initial_types[(input, FloatTensorType([None, X.shape[1]]))]) # 保存ONNX模型 with open(model.onnx, wb) as f: f.write(onnx_model.SerializeToString()) # 加载运行 session InferenceSession(model.onnx) inputs {input: X_test.astype(np.float32)} outputs session.run(None, inputs)7.3 模型监控与更新建立模型性能监控机制from datetime import datetime def log_performance(model, X, y, model_name): preds model.predict(X) accuracy (preds y).mean() with open(performance.log, a) as f: f.write(f{datetime.now()},{model_name},{accuracy:.4f}\n) # 记录性能 log_performance(xgb_loaded, X_test, y_test, xgb_prod) log_performance(lgb_loaded, X_test, y_test, lgb_prod)在实际项目中XGBoost和LightGBM的Sklearn接口可以完美融入现有的机器学习工作流。从我的经验来看当特征维度较高时LightGBM通常训练速度更快而在中小型数据集上XGBoost有时能获得略好的性能。关键是根据具体问题选择合适的工具并充分利用交叉验证和特征筛选来构建稳健的模型。
