Datawhale干货

作者：阿水，北京航空航天大学，Datawhale成员

LightGBM是基于XGBoost的一款可以快速并行的树模型框架，内部集成了多种集成学习思路，在代码实现上对XGBoost的节点划分进行了改进，内存占用更低训练速度更快。

LightGBM官网：https://lightgbm.readthedocs.io/en/latest/

参数介绍：https://lightgbm.readthedocs.io/en/latest/Parameters.html

本文内容如下，原始代码获取方式见文末。

• 1 安装方法

• 2 调用方法

• • 2.1 定义数据集

  • 2.2 模型训练

  • 2.3 模型保存与加载

  • 2.4 查看特征重要性

  • 2.5 继续训练

  • 2.6 动态调整模型超参数

  • 2.7 自定义损失函数

• 3 调参方法

• • 人工调参

  • 网格搜索

  • 贝叶斯优化

1 安装方法

LightGBM的安装非常简单，在Linux下很方便的就可以开启GPU训练。可以优先选用从pip安装，如果失败再从源码安装。

• 安装方法：从源码安装

git clone --recursive https://github.com/microsoft/LightGBM ;
cd LightGBM
mkdir build ; cd build
cmake ..# 开启MPI通信机制，训练更快
# cmake -DUSE_MPI=ON ..# GPU版本，训练更快
# cmake -DUSE_GPU=1 ..
make -j4

• 安装方法：pip安装

# 默认版本
pip install lightgbm# MPI版本
pip install lightgbm --install-option=--mpi# GPU版本
pip install lightgbm --install-option=--gpu

2 调用方法

在Python语言中LightGBM提供了两种调用方式，分为为原生的API和Scikit-learn API，两种方式都可以完成训练和验证。当然原生的API更加灵活，看个人习惯来进行选择。

2.1 定义数据集

df_train = pd.read_csv('https://cdn.coggle.club/LightGBM/examples/binary_classification/binary.train', header=None, sep='\t')
df_test = pd.read_csv('https://cdn.coggle.club/LightGBM/examples/binary_classification/binary.test', header=None, sep='\t')
W_train = pd.read_csv('https://cdn.coggle.club/LightGBM/examples/binary_classification/binary.train.weight', header=None)[0]
W_test = pd.read_csv('https://cdn.coggle.club/LightGBM/examples/binary_classification/binary.test.weight', header=None)[0]y_train = df_train[0]
y_test = df_test[0]
X_train = df_train.drop(0, axis=1)
X_test = df_test.drop(0, axis=1)
num_train, num_feature = X_train.shape# create dataset for lightgbm
# if you want to re-use data, remember to set free_raw_data=Falselgb_train = lgb.Dataset(X_train, y_train,weight=W_train, free_raw_data=False)lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train,weight=W_test, free_raw_data=False)

2.2 模型训练

params = {'boosting_type': 'gbdt','objective': 'binary','metric': 'binary_logloss','num_leaves': 31,'learning_rate': 0.05,'feature_fraction': 0.9,'bagging_fraction': 0.8,'bagging_freq': 5,'verbose': 0
}# generate feature names
feature_name = ['feature_' + str(col) for col in range(num_feature)]
gbm = lgb.train(params,lgb_train,num_boost_round=10,valid_sets=lgb_train,  # eval training datafeature_name=feature_name,categorical_feature=[21])

2.3 模型保存与加载

# save model to file
gbm.save_model('model.txt')print('Dumping model to JSON...')
model_json = gbm.dump_model()with open('model.json', 'w+') as f:json.dump(model_json, f, indent=4)

2.4 查看特征重要性

# feature names
print('Feature names:', gbm.feature_name())# feature importances
print('Feature importances:', list(gbm.feature_importance()))

2.5 继续训练

# continue training
# init_model accepts:
# 1. model file name
# 2. Booster()
gbm = lgb.train(params,lgb_train,num_boost_round=10,init_model='model.txt',valid_sets=lgb_eval)
print('Finished 10 - 20 rounds with model file...')

2.6 动态调整模型超参数

# decay learning rates
# learning_rates accepts:
# 1. list/tuple with length = num_boost_round
# 2. function(curr_iter)
gbm = lgb.train(params,lgb_train,num_boost_round=10,init_model=gbm,learning_rates=lambda iter: 0.05 * (0.99 ** iter),valid_sets=lgb_eval)
print('Finished 20 - 30 rounds with decay learning rates...')# change other parameters during training
gbm = lgb.train(params,lgb_train,num_boost_round=10,init_model=gbm,valid_sets=lgb_eval,callbacks=[lgb.reset_parameter(bagging_fraction=[0.7] * 5 + [0.6] * 5)])
print('Finished 30 - 40 rounds with changing bagging_fraction...')

2.7 自定义损失函数

# self-defined objective function
# f(preds: array, train_data: Dataset) -> grad: array, hess: array
# log likelihood loss
def loglikelihood(preds, train_data):labels = train_data.get_label()preds = 1. / (1. + np.exp(-preds))grad = preds - labelshess = preds * (1. - preds)return grad, hess# self-defined eval metric
# f(preds: array, train_data: Dataset) -> name: string, eval_result: float, is_higher_better: bool
# binary error
# NOTE: when you do customized loss function, the default prediction value is margin
# This may make built-in evalution metric calculate wrong results
# For example, we are doing log likelihood loss, the prediction is score before logistic transformation
# Keep this in mind when you use the customization
def binary_error(preds, train_data):labels = train_data.get_label()preds = 1. / (1. + np.exp(-preds))return 'error', np.mean(labels != (preds > 0.5)), Falsegbm = lgb.train(params,lgb_train,num_boost_round=10,init_model=gbm,fobj=loglikelihood,feval=binary_error,valid_sets=lgb_eval)
print('Finished 40 - 50 rounds with self-defined objective function and eval metric...')

2.8 调参方法

人工调参

For Faster Speed

• Use bagging by setting bagging_fraction and bagging_freq

• Use feature sub-sampling by setting feature_fraction

• Use small max_bin

• Use save_binary to speed up data loading in future learning

• Use parallel learning, refer to Parallel Learning Guide <./Parallel-Learning-Guide.rst>__

For Better Accuracy

• Use large max_bin (may be slower)

• Use small learning_rate with large num_iterations

• Use large num_leaves (may cause over-fitting)

• Use bigger training data

• Try dart

Deal with Over-fitting

• Use small max_bin

• Use small num_leaves

• Use min_data_in_leaf and min_sum_hessian_in_leaf

• Use bagging by set bagging_fraction and bagging_freq

• Use feature sub-sampling by set feature_fraction

• Use bigger training data

• Try lambda_l1, lambda_l2 and min_gain_to_split for regularization

• Try max_depth to avoid growing deep tree

• Try extra_trees

• Try increasing path_smooth

网格搜索

lg = lgb.LGBMClassifier(silent=False)
param_dist = {"max_depth": [4,5, 7],"learning_rate" : [0.01,0.05,0.1],"num_leaves": [300,900,1200],"n_estimators": [50, 100, 150]}grid_search = GridSearchCV(lg, n_jobs=-1, param_grid=param_dist, cv = 5, scoring="roc_auc", verbose=5)
grid_search.fit(train,y_train)
grid_search.best_estimator_, grid_search.best_score_

贝叶斯优化

import warnings
import time
warnings.filterwarnings("ignore")
from bayes_opt import BayesianOptimization
def lgb_eval(max_depth, learning_rate, num_leaves, n_estimators):params = {"metric" : 'auc'}params['max_depth'] = int(max(max_depth, 1))params['learning_rate'] = np.clip(0, 1, learning_rate)params['num_leaves'] = int(max(num_leaves, 1))params['n_estimators'] = int(max(n_estimators, 1))cv_result = lgb.cv(params, d_train, nfold=5, seed=0, verbose_eval =200,stratified=False)return 1.0 * np.array(cv_result['auc-mean']).max()lgbBO = BayesianOptimization(lgb_eval, {'max_depth': (4, 8),'learning_rate': (0.05, 0.2),'num_leaves' : (20,1500),'n_estimators': (5, 200)}, random_state=0)lgbBO.maximize(init_points=5, n_iter=50,acq='ei')
print(lgbBO.max)

本文代码，可以在后台回复【lgb】，下载本文的代码Notebook！

“干货学习，点赞三连↓