机器学习---xgboost与lightgbm效果比较（2）

- 背景

根据“kaggle案例--Instacart Market Basket Analysis（1） ”生成的数据，对xgboost与lightGBM进行效果比较。

- 效果比较

数据量： (847466, 20)xgboost训练时间41s, 精度0.27lightgbm 训练时间9s, 精度0.28可以发现， lightgbm训练速度确实比xgboost快很多，且精度损失不大。

- 测试代码

• （1）生成训练测试数据

import osimport datetimefrom datetime import datetimeimport pandas as pdfrom sklearn.cross_validation import train_test_split #2.7python可用os.chdir(r'd:\pywork\Instacart')data=pd.read_csv('data.txt')train = data.loc[data.eval_set == "train",:]train.drop(['eval_set', 'user_id', 'product_id', 'order_id'], axis=1, inplace=True)train.loc[:, 'reordered'] = train.reordered.fillna(0)X_train, X_val, y_train, y_val = train_test_split(train.drop('reordered', axis=1), train.reordered,                                    test_size=0.9, random_state=42)

• （2）xgboost训练与预测结果

import xgboost #traind_train = xgboost.DMatrix(X_train, y_train)xgb_params = {    "objective"         : "reg:logistic"    ,"eval_metric"      : "logloss"    ,"eta"              : 0.1    ,"max_depth"        : 6    ,"min_child_weight" :10    ,"gamma"            :0.70    ,"subsample"        :0.76    ,"colsample_bytree" :0.95    ,"alpha"            :2e-05    ,"lambda"           :10}watchlist= [(d_train, "train")]xgb_start=datetime.now()bst = xgboost.train(params=xgb_params, dtrain=d_train, num_boost_round=80, evals=watchlist, verbose_eval=10)xgb_end=datetime.now()print 'spendt time :'+ str((xgb_end-xgb_start).seconds)+'(s)'xgboost.plot_importance(bst)''' train result:[0]     train-logloss:0.625642[10]    train-logloss:0.335753[20]    train-logloss:0.269213[30]    train-logloss:0.252115[40]    train-logloss:0.247442[50]    train-logloss:0.245712[60]    train-logloss:0.244735[70]    train-logloss:0.243973[79]    train-logloss:0.243472spendt time :41(s)'''#predict 预测pre_data = xgboost.DMatrix(X_val, y_val)predict=bst.predict(pre_data)X_val['reorder']=y_valX_val['pre']=predictprint "test score is :"precision=float(len(X_val[(X_val['pre']>0.5) & (X_val['reorder']==1)]))/\          float(len(X_val[X_val['pre']>0.5]))   # predict score>0.5 and lables=1 nums divied predict score>0.5 numsrecall=float(len(X_val[(X_val['pre']>0.5) & (X_val['reorder']==1)]))/\          float(len(X_val[X_val['reorder']==1]))     # predict score>0.5 and lables=1 nums divied lables=1f1_score=2*(precision*recall)/(precision+recall)'f1_score: 0.27986198335189855'

• （3） lightgbm训练与预测结果

# trianimport numpy as npimport lightgbm as lgblabels = np.array(y_train, dtype=np.int8)d_train = lgb.Dataset(X_train,                      label=labels)  # , 'order_hour_of_day', 'dow'    params = {    'task': 'train',    'boosting_type': 'gbdt',    'objective': 'binary',    'metric': {'binary_logloss'},    'num_leaves': 96,    'max_depth': 10,    'feature_fraction': 0.9,    'bagging_fraction': 0.95,    'bagging_freq': 5}ROUNDS = 100watchlist=[d_train]lgb_start=datetime.now()bst = lgb.train(params=params, train_set=d_train, num_boost_round=ROUNDS,valid_sets=watchlist,verbose_eval=10)lgb_end=datetime.now()print 'spendt time :'+str((lgb_end-lgb_start).seconds)+'(s)'''' lgb train score:[10]    training's binary_logloss: 0.348039[20]    training's binary_logloss: 0.271007[30]    training's binary_logloss: 0.250972[40]    training's binary_logloss: 0.245258[50]    training's binary_logloss: 0.242898[60]    training's binary_logloss: 0.241338[70]    training's binary_logloss: 0.240099[80]    training's binary_logloss: 0.239047[90]    training's binary_logloss: 0.238009[100]   training's binary_logloss: 0.236996spendt time :9(s)'''#pridictX_val=X_val.drop(['reorder','pre'], axis=1)# xgboost使用时添加了pre,所以这里删除predict=bst.predict(X_val) #X_val 类型string/numpy array/scipy.sparseX_val['reorder']=y_valX_val['pre']=predictprint "test score is :"precision=float(len(X_val[(X_val['pre']>0.5) & (X_val['reorder']==1)]))/\          float(len(X_val[X_val['pre']>0.5]))   # predict score>0.5 and lables=1 nums divied predict score>0.5 numsrecall=float(len(X_val[(X_val['pre']>0.5) & (X_val['reorder']==1)]))/\          float(len(X_val[X_val['reorder']==1]))     # predict score>0.5 and lables=1 nums divied lables=1f1_score=2*(precision*recall)/(precision+recall)print 'F1 score is :'+str(f1_score)'F1 score is :0.28706938762'