基于Python的Xgboost模型实现

# -*- coding: utf-8 -*-
"""
Created on Mon Nov 28 13:22:08 2016
@author: yy
"""

## part1: 模型训练
import pandas as pd
import numpy as np
import os
from datetime import datetime
import xgboost as xgb

traindata=pd.read_csv("/home/op1/yy/traindata.csv")
traindata["add_date"]= pd.to_datetime(traindata["orderdate"]) #转化为日期格式

#时间过滤
traindata_a=traindata[(traindata["add_date"]<'2016-10-31')] #训练集
testdata_a=traindata[(traindata["add_date"]>='2016-10-31') & (traindata["add_date"]<='2016-11-06') & (traindata["rank1"]<=500)] #预测集
                     
#特征列表
colnames=[ 
#'masterhotel'
'order_cii_notcancelcii'        
,'city'             
,'order_cii_ahead_1day'
,'order_cii_ahead_3days_avg'
,'order_cii_ahead_7days_avg'
,'order_cii_30days_avg'
,'order_cii_ahead_sameoneweek'
,'order_cii_ahead_sametwoweeks_avg'
,'star'                
,'goldstar'            
,'level'      
,'ratingservice'       
,'novoters'            
,'week_day'       
,'working_day'
,'cii_ahead_sameoneweek'
,'cii_ahead_sametwoweeks_avg'
,'cii_ahead_samethreeweeks_avg'
,'cii_ahead_samefourweeks_avg'
,'simple_estimate_constant'
,'cii_ahead_1day_avg'
,'cii_ahead_3days_avg'
,'cii_ahead_7days_avg'
,'order_ahead_lt_1days'
,'order_ahead_lt_2days'
,'order_ahead_lt_3days'
,'order_ahead_lt_7days'
,'order_ahead_lt_14days'
,'order_alldays'     
,'click_ahead_1day'    
,'click_ahead_2days'   
,'click_ahead_3days'   
,'click_ahead_7days'   
,'click_ahead_14days'  
,'browse_0day_uv'      
,'browse_1day_uv'      
,'browse_2day_uv'      
,'browse_3day_uv'      
,'browse_4day_uv'      
,'browse_5day_uv'      
,'browse_6day_uv'      
,'browse_7_14day_uv'   
,'browse_14daymore_uv' 
,'order_cii_14days_avg'               
,'order_cii_21days_avg' 
,'order_cii_ahead_samethreeweeks_avg' 
,'order_cii_ahead_samefourweeks_avg']

#dtrain = xgb.DMatrix( data, label=label)
#dtrain = xgb.DMatrix(data, label=label)
#训练集构建

label=traindata_a[colnames[0]]
dtrain = xgb.DMatrix(traindata_a[colnames[1:]], label=traindata_a[colnames[0]])

num_round=800
params = {} 
params["objective"] = "reg:linear" 
params["eta"] = 0.1
params["max_depth"] = 4 
params["eval_metric"] = "rmse"
params["silent"] = 0
plst = list(params.items()) #Using 5000 rows for early stopping. 

bst = xgb.train( plst, dtrain, num_round)

#预测集构建
dtest=xgb.DMatrix(testdata_a[colnames[1:]])

#结果预测
y_bar=bst.predict(dtest)

#y.shape
##  评价指标
#testdata_a[colnames[0,]]
#testdata_a.columns
actual_values=testdata_a[["masterhotel","add_date","order_cii_notcancelcii","rank1"]]
actual_values["y_bar"]=y_bar
actual_values["mae"]=abs(actual_values["y_bar"]-actual_values["order_cii_notcancelcii"])

top100=actual_values[actual_values["rank1"]<=100]
mae100=top100.groupby("add_date").mean()
mae500=actual_values.groupby("add_date").mean()

## part2: 存储模型，方便后续直接调用
bst.save_model('xgb.model') # 用于存储训练出的模型

#模型初始化
bst=xgb.Booster() #注意：名字要保持一致，否则报错！
bst.load_model('/home/op1/yuanmin/xgb.model')

y_bar=bst.predict(dtest)

## 参数重要性

pd.Series(bst.get_fscore()).sort_values(ascending=False)

pd.Series(bst.get_fscore()).sort_values(ascending=False)/sum(pd.Series(bst.get_fscore()).sort_values(ascending=False)) #重要性归一化

附两篇参考文章：

1、https://www.analyticsvidhya.com/blog/2016/03/complete-guide-parameter-tuning-xgboost-with-codes-python/

2、http://www.cnblogs.com/haobang008/p/5909207.html