一、使用步骤
0.首先展示最后的结果
参数含义
learning_rate 一般设置在0.05-0.1之间
n_estimators 100-1000 boosting的迭代次数
min_split_gain 0 执行节点分裂的最小增益 不建议去调整
min_child_sample 一个叶子上的最小数据量,默认设置为20。根据数据量来确定，当数据量比较大时，应提升这个数值，
让叶子节点的数据分布相对稳定。
min_child_weight 一个叶子上的最小hessian和。默认设置为0.001，一般设置为1。
需要算法调节的参数
max_depth 树模型的最大深度。防止过拟合的最重要的参数，一般限制为3~5之间
num_leaves 一棵树上的叶子节点个数。默认设置为31，和max_depth配合来空值树的形状，一般设置为(0, 2^max_depth - 1]的一个数值。
subsample 默认设置为1，一般设置为0。8~1.0之间，防止过拟合。
colsample_bytree 一般设置为0.8~1.0之间，防止过拟合
reg_alpha L1正则化参数
reg_lambda L2正则化参数 较大的数值会让各个特征对模型的影响力趋于均匀
def roc_auc_plot(clf,x_train,y_train,x_test, y_test):
train_auc = roc_auc_score(y_train,clf.predict_proba(x_train)[:,1])
train_fpr, train_tpr, _ = roc_curve(y_train,clf.predict_proba(x_train)[:,1])
train_ks = abs(train_fpr-train_tpr).max()
print('train_ks = ', train_ks)
print('train_auc = ', train_auc)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
from matplotlib import pyplot as plt
plt.plot(train_fpr,train_tpr,label = 'train_roc')
plt.plot(test_fpr,test_tpr,label = 'test_roc')
plt.plot([0,1],[0,1],'k--', c='r')
plt.xlabel('False positive rate')
plt.ylabel('True positive rate')
plt.title('ROC Curve')
plt.legend(loc = 'best')
plt.show()
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=16,
colsample_bytree=0.8,
subsample=0.8,
reg_alpha=0.7777777777777778,
reg_lambda=0.3,
objective='binary')
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
roc_auc_plot(clf,x_train,y_train,x_test, y_test)
1.读入清洗后的数据集
代码如下：
import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.metrics import roc_auc_score,roc_curve,auc
import lightgbm as lgb
from multiprocessing import cpu_count
from sklearn.model_selection import KFold
import time
from sklearn.model_selection import GridSearchCV, StratifiedKFold, cross_val_score
import numpy as np
df_Master_clean = pd.read_csv(r'F:\教师培训\ppd7\df_Master_clean.csv', encoding='gb18030')
c=['Idx','target','ListingInfo','sample_status']
x = df_Master_clean[df_Master_clean['target'].notnull()].drop(columns=c)
y = df_Master_clean[df_Master_clean['target'].notnull()]['target']
x_train,x_test, y_train, y_test = train_test_split(x,y,random_state=2,test_size=0.2)
2.参数调整
代码如下（示例）：
#随机设置一组参数为基础
lgb_model1 = lgb.LGBMClassifier(n_estimators=20,
boosting_type='gbdt',
learning_rate=0.01,
min_child_samples=100,
min_child_weight=0.003,
max_depth=4,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
import matplotlib.pyplot as plt
通过观察auc大小调参，找到最佳n_estimators
axisx = range(100,1001,10)
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=i,
boosting_type='gbdt',
learning_rate=0.01,
min_child_samples=100,
min_child_weight=0.003,
max_depth=4,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.xticks(axisx)
plt.legend()
plt.show()
import matplotlib.pyplot as plt
通过观察auc大小调参，找到最佳learning_rate,learning_rate=0.04, max_auc=0.7720493433228038
axisx = np.linspace(0.01,0.05,5)
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=i,
min_child_samples=100,
min_child_weight=0.003,
max_depth=4,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('learning_rate={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参，找到最佳min_child_samples min_child_samples=68, max_auc=0.7754420865945606
axisx = np.linspace(0.01,0.05,5)
axisx = range(65,75,1)
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=i,
min_child_weight=0.003,
max_depth=4,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('min_child_samples={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参，min_child_weight对auc面积没有影响
axisx = np.linspace(0.001,0.005,10)
axisx = range(65,75,1)
axisx = [0.001, 0.005]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=i,
max_depth=4,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('min_child_weight={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参max_depth, max_depth=4, max_auc=0.7754420865945606
axisx = np.linspace(0.001,0.005,10)
axisx = range(65,75,1)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=i,
num_leaves=100,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('max_depth={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参num_leaves, num_leaves=16, max_auc=0.7754420865945606
axisx = np.linspace(0.001,0.005,10)
axisx = range(10,20,1)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=i,
colsample_bytree=0.7,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('num_leaves={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参colsample_bytree, colsample_bytree=0.8, max_auc=0.7770645958950596
axisx = np.linspace(0.8,0.85,5)
axisx = range(10,101,10)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=16,
colsample_bytree=i,
subsample=0.6,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
max_features = 140,)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('colsample_bytree={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参subsample, 此参数没有对auc产生影响
axisx = np.linspace(0.7,0.8,10)
axisx = range(10,101,10)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=16,
colsample_bytree=0.8,
subsample=i,
reg_alpha=0.03,
reg_lambda=0.3,
objective='binary',
)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('subsample={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参reg_alpha, reg_alpha=0.7777777777777778, max_auc=0.7773465038084242
axisx = np.linspace(0.7,0.8,10)
axisx = range(0,1,10)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=16,
colsample_bytree=0.8,
subsample=0.8,
reg_alpha=i,
reg_lambda=0.3,
objective='binary',
)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('reg_alpha={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
通过观察auc大小调参reg_lambda, reg_lambda=0.30000000000000004, max_auc=0.7773465038084242
axisx = np.linspace(0.28,0.32,5)
axisx = range(0,1,10)
axisx = [3, 4, 5, 6, 7]
test_aucs = []
for i in axisx:
lgb_model = lgb.LGBMClassifier(n_estimators=800,
boosting_type='gbdt',
learning_rate=0.04,
min_child_samples=68,
min_child_weight=0.01,
max_depth=4,
num_leaves=16,
colsample_bytree=0.8,
subsample=0.8,
reg_alpha=0.7777777777777778,
reg_lambda=i,
objective='binary',
)
clf = lgb_model.fit(x_train, y_train,
eval_set=[(x_train, y_train),(x_test,y_test)],
eval_metric='auc',early_stopping_rounds=100)
test_auc = roc_auc_score(y_test,clf.predict_proba(x_test)[:,1])
test_aucs.append(test_auc)
print('reg_lambda={}, max_auc={}'.format(axisx[test_aucs.index(max(test_aucs))], max(test_aucs)))
test_fpr, test_tpr, _ = roc_curve(y_test,clf.predict_proba(x_test)[:,1])
test_ks = abs(test_fpr-test_tpr).max()
print('test_ks = ', test_ks)
print('test_auc = ', test_auc)
plt.figure(figsize=(20,5))
plt.plot(axisx, test_aucs,c="red",label="LGBC")
plt.ylabel('auc')
plt.xticks(axisx)
plt.legend()
plt.show()
总结
我们观察整个调参过程，有的参数并没有对auc产生影响，有的效果很明显，总的来说调参后的效果要好些。
Original: https://blog.csdn.net/weixin_43827767/article/details/120586336
Author: sunnuan01
Title: 数据挖掘竞赛lightgbm通过求最大auc调参
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/699682/
转载文章受原作者版权保护。转载请注明原作者出处！