lightgbm使用leaf_wise tree生长策略，leaf_wise_tree的优点是收敛速度快，缺点是容易过拟合。
# lightgbm关键参数
# lightgbm调参方法cv
1 # -*- coding: utf-8 -*-
2 """
3 # 作者：wanglei5205
4 # 邮箱：wanglei5205@126.com
5 # 博客：http://cnblogs.com/wanglei5205
6 # github：http://github.com/wanglei5205
7 """
8 ### 导入模块
9 import numpy as np
10 import pandas as pd
11 import lightgbm as lgb
12 from sklearn import metrics
13
14 ### 载入数据
15 print('载入数据')
16 dataset1 = pd.read_csv('G:/ML/ML_match/IJCAI/data3.22/3.22ICJAI/data/7_train_data1.csv')
17 dataset2 = pd.read_csv('G:/ML/ML_match/IJCAI/data3.22/3.22ICJAI/data/7_train_data2.csv')
18 dataset3 = pd.read_csv('G:/ML/ML_match/IJCAI/data3.22/3.22ICJAI/data/7_train_data3.csv')
19 dataset4 = pd.read_csv('G:/ML/ML_match/IJCAI/data3.22/3.22ICJAI/data/7_train_data4.csv')
20 dataset5 = pd.read_csv('G:/ML/ML_match/IJCAI/data3.22/3.22ICJAI/data/7_train_data5.csv')
21
22 print('数据去重')
23 dataset1.drop_duplicates(inplace=True)
24 dataset2.drop_duplicates(inplace=True)
25 dataset3.drop_duplicates(inplace=True)
26 dataset4.drop_duplicates(inplace=True)
27 dataset5.drop_duplicates(inplace=True)
28
29 print('数据合并')
30 trains = pd.concat([dataset1,dataset2],axis=0)
31 trains = pd.concat([trains,dataset3],axis=0)
32 trains = pd.concat([trains,dataset4],axis=0)
33
34 online_test = dataset5
35
36 ### 数据拆分(训练集+验证集+测试集)
37 print('数据拆分')
38 from sklearn.model_selection import train_test_split
39 train_xy,offline_test = train_test_split(trains,test_size = 0.2,random_state=21)
40 train,val = train_test_split(train_xy,test_size = 0.2,random_state=21)
41
42 # 训练集
43 y_train = train.is_trade # 训练集标签
44 X_train = train.drop(['instance_id','is_trade'],axis=1) # 训练集特征矩阵
45
46 # 验证集
47 y_val = val.is_trade # 验证集标签
48 X_val = val.drop(['instance_id','is_trade'],axis=1) # 验证集特征矩阵
49
50 # 测试集
51 offline_test_X = offline_test.drop(['instance_id','is_trade'],axis=1) # 线下测试特征矩阵
52 online_test_X = online_test.drop(['instance_id'],axis=1) # 线上测试特征矩阵
53
54 ### 数据转换
55 print('数据转换')
56 lgb_train = lgb.Dataset(X_train, y_train, free_raw_data=False)
57 lgb_eval = lgb.Dataset(X_val, y_val, reference=lgb_train,free_raw_data=False)
58
59 ### 设置初始参数--不含交叉验证参数
60 print('设置参数')
61 params = {
62 'boosting_type': 'gbdt',
63 'objective': 'binary',
64 'metric': 'binary_logloss',
65 }
66
67 ### 交叉验证(调参)
68 print('交叉验证')
69 min_merror = float('Inf')
70 best_params = {}
71
72 # 准确率
73 print("调参1：提高准确率")
74 for num_leaves in range(20,200,5):
75 for max_depth in range(3,8,1):
76 params['num_leaves'] = num_leaves
77 params['max_depth'] = max_depth
78
79 cv_results = lgb.cv(
80 params,
81 lgb_train,
82 seed=2018,
83 nfold=3,
84 metrics=['binary_error'],
85 early_stopping_rounds=10,
86 verbose_eval=True
87 )
88
89 mean_merror = pd.Series(cv_results['binary_error-mean']).min()
90 boost_rounds = pd.Series(cv_results['binary_error-mean']).argmin()
91
92 if mean_merror < min_merror:
93 min_merror = mean_merror
94 best_params['num_leaves'] = num_leaves
95 best_params['max_depth'] = max_depth
96
97 params['num_leaves'] = best_params['num_leaves']
98 params['max_depth'] = best_params['max_depth']
99
100 # 过拟合
101 print("调参2：降低过拟合")
102 for max_bin in range(1,255,5):
103 for min_data_in_leaf in range(10,200,5):
104 params['max_bin'] = max_bin
105 params['min_data_in_leaf'] = min_data_in_leaf
106
107 cv_results = lgb.cv(
108 params,
109 lgb_train,
110 seed=42,
111 nfold=3,
112 metrics=['binary_error'],
113 early_stopping_rounds=3,
114 verbose_eval=True
115 )
116
117 mean_merror = pd.Series(cv_results['binary_error-mean']).min()
118 boost_rounds = pd.Series(cv_results['binary_error-mean']).argmin()
119
120 if mean_merror < min_merror:
121 min_merror = mean_merror
122 best_params['max_bin']= max_bin
123 best_params['min_data_in_leaf'] = min_data_in_leaf
124
125 params['min_data_in_leaf'] = best_params['min_data_in_leaf']
126 params['max_bin'] = best_params['max_bin']
127
128 print("调参3：降低过拟合")
129 for feature_fraction in [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]:
130 for bagging_fraction in [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]:
131 for bagging_freq in range(0,50,5):
132 params['feature_fraction'] = feature_fraction
133 params['bagging_fraction'] = bagging_fraction
134 params['bagging_freq'] = bagging_freq
135
136 cv_results = lgb.cv(
137 params,
138 lgb_train,
139 seed=42,
140 nfold=3,
141 metrics=['binary_error'],
142 early_stopping_rounds=3,
143 verbose_eval=True
144 )
145
146 mean_merror = pd.Series(cv_results['binary_error-mean']).min()
147 boost_rounds = pd.Series(cv_results['binary_error-mean']).argmin()
148
149 if mean_merror < min_merror:
150 min_merror = mean_merror
151 best_params['feature_fraction'] = feature_fraction
152 best_params['bagging_fraction'] = bagging_fraction
153 best_params['bagging_freq'] = bagging_freq
154
155 params['feature_fraction'] = best_params['feature_fraction']
156 params['bagging_fraction'] = best_params['bagging_fraction']
157 params['bagging_freq'] = best_params['bagging_freq']
158
159 print("调参4：降低过拟合")
160 for lambda_l1 in [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]:
161 for lambda_l2 in [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]:
162 for min_split_gain in [0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0]:
163 params['lambda_l1'] = lambda_l1
164 params['lambda_l2'] = lambda_l2
165 params['min_split_gain'] = min_split_gain
166
167 cv_results = lgb.cv(
168 params,
169 lgb_train,
170 seed=42,
171 nfold=3,
172 metrics=['binary_error'],
173 early_stopping_rounds=3,
174 verbose_eval=True
175 )
176
177 mean_merror = pd.Series(cv_results['binary_error-mean']).min()
178 boost_rounds = pd.Series(cv_results['binary_error-mean']).argmin()
179
180 if mean_merror < min_merror:
181 min_merror = mean_merror
182 best_params['lambda_l1'] = lambda_l1
183 best_params['lambda_l2'] = lambda_l2
184 best_params['min_split_gain'] = min_split_gain
185
186 params['lambda_l1'] = best_params['lambda_l1']
187 params['lambda_l2'] = best_params['lambda_l2']
188 params['min_split_gain'] = best_params['min_split_gain']
189
190
191 print(best_params)
192
193 ### 训练
194 params['learning_rate']=0.01
195 lgb.train(
196 params, # 参数字典
197 lgb_train, # 训练集
198 valid_sets=lgb_eval, # 验证集
199 num_boost_round=2000, # 迭代次数
200 early_stopping_rounds=50 # 早停次数
201 )
202
203 ### 线下预测
204 print ("线下预测")
205 preds_offline = lgb.predict(offline_test_X, num_iteration=lgb.best_iteration) # 输出概率
206 offline=offline_test[['instance_id','is_trade']]
207 offline['preds']=preds_offline
208 offline.is_trade = offline['is_trade'].astype(np.float64)
209 print('log_loss', metrics.log_loss(offline.is_trade, offline.preds))
210
211 ### 线上预测
212 print("线上预测")
213 preds_online = lgb.predict(online_test_X, num_iteration=lgb.best_iteration) # 输出概率
214 online=online_test[['instance_id']]
215 online['preds']=preds_online
216 online.rename(columns={'preds':'predicted_score'},inplace=True) # 更改列名
217 online.to_csv("./data/20180405.txt",index=None,sep=' ') # 保存结果
218
219 ### 保存模型
220 from sklearn.externals import joblib
221 joblib.dump(lgb,'lgb.pkl')
222
223 ### 特征选择
224 df = pd.DataFrame(X_train.columns.tolist(), columns=['feature'])
225 df['importance']=list(lgb.feature_importance()) # 特征分数
226 df = df.sort_values(by='importance',ascending=False) # 特征排序
227 df.to_csv("./data/feature_score_20180331.csv",index=None,encoding='gbk') # 保存分数
来源：oschina
链接：https://my.oschina.net/u/2624214/blog/1922841