1.lightGBM安裝

在anaconda中輸入：pip install lightGBM即可

輸入import lightgbm as lgb做測試

2.lightGBM改進

1.直方圖差加速:直方圖演算法的基本思想是先把連續的浮點特徵值離散化成k個整數，同時構造一個寬度為k的直方圖。在遍歷資料的時候，根據離散化後的值作為索引在直方圖中累積統計量，當遍歷一次資料後，直方圖累積了需要的統計量，然後根據直方圖的離散值，遍歷尋找最優的分割點。記憶體消耗降低，計算上的代價也大幅降低2.leaf-wise:每次從當前所有葉子中，找到分裂增益最大的一個葉子，然後分裂，如此迴圈。因此同Level-wise相比，在分裂次數相同的情況下，Leaf-wise可以降低更多的誤差，得到更好的精度。可能會長出比較深的決策樹，產生過擬合。因此LightGBM在Leaf-wise之上增加了一個最大深度限制，在保證高效率的同時防止過擬合。

3.lightGBM使用

#讀模型資料
import pandas as pd
df=pd.read_csv('all_data.csv')

#剔除模型無關變數和完全線性相關的變數

"""取樣"""
df_1 = df[df['LABEL'] == 1]
df_0 = df[df['LABEL'] == 0]
df3 = df_0.sample(10000)

df_con= pd.concat([df_1, df3], ignore_index=True)

#剔除模型無關變數和完全線性相關的變數
select_var = ~df.columns.isin(["PERSONID","LABEL","CREATETIME_time"])

X = df.ix[:, select_var]
Y = df.LABEL.values
#X=(X-X.mean(axis=0))/X.std(axis=0)
X=X.fillna(0)

from sklearn.cross_validation import train_test_split
train_X, test_X, train_y, test_y = train_test_split(X, Y, train_size = 0.8, random_state = 123)



import lightgbm as lgb
lgb_train = lgb.Dataset(train_X, train_y, free_raw_data=False)
lgb_eval = lgb.Dataset(test_X, test_y, reference=lgb_train,free_raw_data=False)



param = {
'task': 'train',
'boosting_type': 'gbdt',
'objective': 'binary',
'metric': {'l2', 'auc'},
'num_leaves': 40,
'learning_rate': 0.01,
'feature_fraction': 0.8,
'bagging_fraction': 0.8,
'bagging_freq': 5,
'verbose': 0
}
param['is_unbalance']='true'
param['metric'] = 'auc'




bst=lgb.cv(param,lgb_train, num_boost_round=1000, nfold=6, early_stopping_rounds=100)

estimators = lgb.train(param,lgb_train,num_boost_round=len(bst['auc-mean']))
 
print('Start training...')


y_pred = estimators.predict(test_X, num_iteration=estimators.best_iteration)

from sklearn import metrics
print('The roc of prediction is:', metrics.roc_auc_score(test_y, y_pred) )

print('Feature names:', estimators.feature_name())
print('Feature importances:', list(estimators.feature_importance()))