# 参考了 魔镜杯风控算法实战（三）调参篇
https://zhuanlan.zhihu.com/p/104376738
但是原文章中一些参数已经在最新的lightgbm 中被废弃了，auc-mean改名为valid auc-mean！所以，代码也需要做一些细微的调整才可以运行起来，所以我写出这篇经验供大家参考。
1）先调参：
num_boost_round（n_estimators）
使用lgb自带的cv来计算，先限定learning_rate为0.1，这样调整别的参数的时候可以迭代得更快一些。
import time
import lightgbm as lgb
from lightgbm import log_evaluation, early_stopping
start = time.time()
lgb_train = lgb.Dataset(X_train,y_train)
lgb_test = lgb.Dataset(X_test,y_test,reference=lgb_train)
base_parmas={'boosting_type':'gbdt',
'learning_rate':0.1,
'num_leaves':31,
'max_depth':-1,
'bagging_fraction':0.7,
'feature_fraction':0.7,
'lambda_l1':0,
'lambda_l2':0,
'min_data_in_leaf':20,
'min_sum_hessian_inleaf':0.001,
'metric':'auc'}
callbacks = [log_evaluation(period=100), early_stopping(stopping_rounds=30)]
cv_result = lgb.cv(train_set=lgb_train,
num_boost_round=1000,
nfold=5,
stratified=True,
shuffle=True,
params=base_parmas,
metrics='auc',
callbacks=callbacks,
seed=0)
end = time.time()
print('迭代次数:{}'.format(len(cv_result['valid auc-mean'])))
print('交叉验证的AUC:{}'.format(max(cv_result['valid auc-mean'])))
print('运行时间为{}秒'.format(round(end-start,0)))得到：
迭代次数:444 交叉验证的AUC:0.7941171177211479 运行时间为3.0秒
注意我改动了cv_result['valid auc-mean']) auc-mean已经更名为了'valid auc-mean
2）然后：
下面配合GridSearchCV时必须使用sklearn接口的lightgbm。
- num_leaves
这里选择先粗调再细调的策略，偷懒的话也可以直接细调，一步到位，但是代码会跑的久一些。
start = time.time()
params1={'num_leaves':list(range(50,150,5))}
model_lgb1=lgb.LGBMClassifier(
learning_rate=0.1,
n_estimators=444,
max_depth=-1,
bagging_fraction=0.7,
feature_fraction=0.7,
lambda_l1=0,
lambda_l2=0,
min_data_in_leaf=20,
min_sum_hessian_inleaf=0.001)
grid_search1=GridSearchCV(estimator=model_lgb1,cv=5,param_grid=params1,n_jobs=-1,scoring='roc_auc')
grid_search1.fit(X_train,y_train)
end = time.time()
print('最优参数为:{}'.format(grid_search1.best_params_))
print('最优分数为:{}'.format(grid_search1.best_score_))
print('运行时间为{}秒'.format(round(end-start,0)))最优参数为:{'num_leaves': 55}
最优分数为:0.7928556904381894
运行时间为38.0秒
start = time.time()
params2={'num_leaves':list(range(30,60,2))}
model_lgb2=lgb.LGBMClassifier(
learning_rate=0.1,
n_estimators=444,
max_depth=-1,
bagging_fraction=0.7,
feature_fraction=0.7,
lambda_l1=0,
lambda_l2=0,
min_data_in_leaf=20,
min_sum_hessian_inleaf=0.001)
grid_search2=GridSearchCV(estimator=model_lgb2,cv=5,param_grid=params2,n_jobs=-1,scoring='roc_auc')
grid_search2.fit(X_train,y_train)
end = time.time()
print('最优参数为:{}'.format(grid_search2.best_params_))
print('最优分数为:{}'.format(grid_search2.best_score_))
print('运行时间为{}秒'.format(round(end-start,0)))最优参数为:{'num_leaves': 38}
最优分数为:0.7930652393271933
运行时间为16.0秒
3）调优：bagging_fraction，feature_fraction
start = time.time()
params3={'bagging_fraction':[i/10 for i in range(4,11,1)],'feature_fraction':[i/10 for i in range(4,11,1)]}
model_lgb3=lgb.LGBMClassifier(
learning_rate=0.1,
n_estimators=444,
num_leaves=38,
max_depth=-1,
lambda_l1=0,
lambda_l2=0,
min_data_in_leaf=20,
min_sum_hessian_inleaf=0.001)
grid_search3=GridSearchCV(estimator=model_lgb3,cv=5,param_grid=params3,n_jobs=-1,scoring='roc_auc')
grid_search3.fit(X_train,y_train)
end = time.time()
print('最优参数为:{}'.format(grid_search3.best_params_))
print('最优分数为:{}'.format(grid_search3.best_score_))
print('运行时间为{}秒'.format(round(end-start,0)))最优参数为:{'bagging_fraction': 0.4, 'feature_fraction': 0.4}
最优分数为:0.7930652393271933
运行时间为55.0秒
4）调优：lambda_l1，lambda_l2
start = time.time()
params4={'lambda_l1':[0,0.001,0.01,0.03,0.08,0.3,0.5],'lambda_l2':[0,0.001,0.01,0.03,0.08,0.3,0.5]}
model_lgb4=lgb.LGBMClassifier(
learning_rate=0.1,
n_estimators=444,
num_leaves=38,
max_depth=-1,
bagging_fraction=0.4,
feature_fraction=0.4,
min_data_in_leaf=20,
min_sum_hessian_inleaf=0.001)
grid_search4=GridSearchCV(estimator=model_lgb4,cv=5,param_grid=params4,n_jobs=-1,scoring='roc_auc')
grid_search4.fit(X_train,y_train)
end = time.time()
print('最优参数为:{}'.format(grid_search4.best_params_))
print('最优分数为:{}'.format(grid_search4.best_score_))
print('运行时间为{}秒'.format(round(end-start,0)))最优参数为:{'lambda_l1': 0.3, 'lambda_l2': 0.08}
最优分数为:0.793414505029456
运行时间为47.0秒
请注意原文还需要调优min_data_in_leaf,min_sum_hessian_inleaf，但实际上这两个参数不需要调优，已经和新版的lightgbm冲突，不用再调了
5）上面几个确定好了以后，我们设置一个比较小的learning_rate 0.005，来确定最终的num_boost_round
start = time.time()
final_parmas={'boosting_type':'gbdt',
'learning_rate':0.005,
'num_leaves':38,
'max_depth':-1,
'bagging_fraction':0.4,
'feature_fraction':0.4,
'lambda_l1':0.3,
'lambda_l2':0.08,
'metric':'auc'}
callbacks = [log_evaluation(period=100), early_stopping(stopping_rounds=30)]
cv_result = lgb.cv(train_set=lgb_train,
num_boost_round=444,
nfold=5,
stratified=True,
shuffle=True,
params=final_parmas,
metrics='auc',
callbacks=callbacks,
seed=0)
end = time.time()
print('最大的迭代次数:{}'.format(len(cv_result['valid auc-mean'])))
print('交叉验证的AUC:{}'.format(max(cv_result['valid auc-mean'])))
print('运行时间为{}秒'.format(round(end-start,0)))最大的迭代次数:444 交叉验证的AUC:0.7607449893610381 运行时间为2.0秒
6）最后进行最后建模：最优参数已经全部计算好了，我们把它代入模型并直接将AUC可视化，看一下优化以后的结果如何
from sklearn import metrics
final_model = lgb.LGBMClassifier(learning_rate=0.005,
n_estimators=500,
num_leaves=38,
max_depth=-1,
bagging_fraction=0.4,
feature_fraction=0.4,
lambda_l1=0.3,
lambda_l2=0.08,
)
final_model.fit(X_train,y_train)
final_pre= final_model.predict_proba(X_test)[:,1]
fpr, tpr, threshold = metrics.roc_curve(y_test,final_pre)
roc_auc = metrics.auc(fpr, tpr)
plt.figure(figsize=(6,6))
plt.title('Validation ROC')
plt.plot(fpr, tpr, 'b', label = 'Val AUC = %0.3f' % roc_auc)
plt.legend(loc = 'lower right')
plt.plot([0, 1], [0, 1],'r--')
plt.xlim([0, 1])
plt.ylim([0, 1])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()[LightGBM] [Warning] feature_fraction is set=0.4, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.4 [LightGBM] [Warning] lambda_l1 is set=0.3, reg_alpha=0.0 will be ignored. Current value: lambda_l1=0.3 [LightGBM] [Warning] lambda_l2 is set=0.08, reg_lambda=0.0 will be ignored. Current value: lambda_l2=0.08 [LightGBM] [Warning] bagging_fraction is set=0.4, subsample=1.0 will be ignored. Current value: bagging_fraction=0.4 [LightGBM] [Warning] feature_fraction is set=0.4, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.4 [LightGBM] [Warning] lambda_l1 is set=0.3, reg_alpha=0.0 will be ignored. Current value: lambda_l1=0.3 [LightGBM] [Warning] lambda_l2 is set=0.08, reg_lambda=0.0 will be ignored. Current value: lambda_l2=0.08 [LightGBM] [Warning] bagging_fraction is set=0.4, subsample=1.0 will be ignored. Current value: bagging_fraction=0.4 [LightGBM] [Info] Number of positive: 2715, number of negative: 2676 [LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000100 seconds. You can set `force_col_wise=true` to remove the overhead. [LightGBM] [Info] Total Bins 254 [LightGBM] [Info] Number of data points in the train set: 5391, number of used features: 1 [LightGBM] [Info] [binary:BoostFromScore]: pavg=0.503617 -> initscore=0.014469 [LightGBM] [Info] Start training from score 0.014469 [LightGBM] [Warning] feature_fraction is set=0.4, colsample_bytree=1.0 will be ignored. Current value: feature_fraction=0.4 [LightGBM] [Warning] lambda_l1 is set=0.3, reg_alpha=0.0 will be ignored. Current value: lambda_l1=0.3 [LightGBM] [Warning] lambda_l2 is set=0.08, reg_lambda=0.0 will be ignored. Current value: lambda_l2=0.08 [LightGBM] [Warning] bagging_fraction is set=0.4, subsample=1.0 will be ignored. Current value: bagging_fraction=0.4
7）评估结果 混淆矩阵输出：
# 如果需要更详细的分类报告（包括精确度、召回率、F1分数等）
from sklearn.metrics import classification_report
y_pred=final_model.predict(X_test)
print("Classification Report:")
print(classification_report(y_test, y_pred))Classification Report: precision recall f1-score support
0 0.69 0.67 0.68 1175
1 0.67 0.69 0.68 1136
accuracy 0.68 2311整个流程就完成了，说下最终感受，这一圈代码搞完算上查资料花了接近一小时，但是模型效果并没有比默认参数提高多少。
我贴出默认lightGBM参数的建模：
# 不网格搜索 lgbm
import numpy as np
from sklearn.model_selection import train_test_split
from imblearn.over_sampling import RandomOverSampler
from lightgbm import LGBMClassifier
# 假设 data_total 是您的原始数据DataFrame
# data_total['ABS'] 是特征列，data_total['财务舞弊'] 是目标列
# 转换数据类型为numpy数组
X = np.array(data_total['ABS']).reshape(-1, 1)
y = np.array(data_total['财务舞弊'])
# 上采样
ros = RandomOverSampler(sampling_strategy='auto', random_state=0)
X_resampled, y_resampled = ros.fit_resample(X, y)
# 划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(X_resampled, y_resampled, test_size=0.3, random_state=25)
# 初始化LightGBM分类器，这里使用默认参数
clf = LGBMClassifier(random_state=0)
# 训练模型
clf.fit(X_train, y_train)
# 在测试集上评估模型
# 注意：LGBMClassifier的score方法默认返回准确率，对于不平衡数据集可能不是最佳选择
test_score = clf.score(X_test, y_test)
print("Test score (accuracy) with default parameters:", test_score)
# 如果您想使用其他指标（如ROC AUC）来评估测试集，可以使用sklearn的metrics模块
from sklearn.metrics import roc_auc_score
y_pred_proba = clf.predict_proba(X_test)[:, 1] # 获取正类的预测概率
test_auc = roc_auc_score(y_test, y_pred_proba)
print("Test AUC with default parameters:", test_auc)
扫一扫
1079
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
