【Python学习系列二十四】scikit-learn库逻辑回归实现唯品会用户购买行为预测

1、背景：http://www.datafountain.cn/#/competitions/260/intro

                  DataFountain上的唯品会用户购买行为预测比赛题目，笔者用逻辑回归实现，分数是0.48比较弱，代码这里参考。

2、通过比赛提取的特征如下：

   

特征类别	特征名	特征说明	训练说明
基本特征	u_id	用户唯一标识
	spu_id	商品唯一标识
	brand_id	商品所属的品牌标识
	cat_id	商品所属的品类标识
人的特征	u_buy_num	购买次数
	u_click_num	点击次数
	u_buy_date	购买天数
	u_click_date	点击天数
	u_num_ratio	购买点击次数比，购买次数/点击次数
	u_date_ratio	购买点击天数比，购买天数/点击天数
	u_buy_freq	购买频率，购买次数/90天
	u_click_freq	点击频率，购买次数/90天
商品的特征	spu_buy_num	购买次数
	spu_click_num	点击次数
	spu_buy_date	购买天数
	spu_click_date	点击天数
	spu_num_ratio	购买点击次数比，购买次数/点击次数
	spu_date_ratio	购买点击天数比，购买天数/点击天数
	spu_buy_freq	购买频率，购买次数/90天
	spu_click_freq	点击频率，购买次数/90天
品牌的特征	brand_buy_num	购买次数
	brand_click_num	点击次数
	brand_buy_date	购买天数
	brand_click_date	点击天数
	brand_num_ratio	购买点击次数比，购买次数/点击次数
	brand_date_ratio	购买点击天数比，购买天数/点击天数
	brand_buy_freq	购买频率，购买次数/90天
	brand_click_freq	点击频率，购买次数/90天
品类的特征	cat_buy_num	购买次数
	cat_click_num	点击次数
	cat_buy_date	购买天数
	cat_click_date	点击天数
	cat_num_ratio	购买点击次数比，购买次数/点击次数
	cat_date_ratio	购买点击天数比，购买天数/点击天数
	cat_buy_freq	购买频率，购买次数/90天
	cat_click_freq	点击频率，购买次数/90天
标记	action_type	该用户是否会在当日购买此商品（0否，1是）

3、逻辑回归参考代码如下：

# -*- coding: utf-8 -*-

import pandas as pd
import time
from sklearn import preprocessing
from sklearn.linear_model import LogisticRegression  
from sklearn import metrics  

def main():
    #第一步：加载训练集和测试集
    #加载带标记数据
    label_ds=pd.read_csv(r"train_features_0714.txt",sep='\t',encoding='utf8',
                         names=["u_id","u_buy_num","u_click_num","u_buy_date","u_click_date","u_num_ratio","u_date_ratio","u_buy_freq","u_click_freq","u_last_date",\
                                "spu_id","spu_buy_num","spu_click_num","spu_buy_date","spu_click_date","spu_num_ratio","spu_date_ratio","spu_buy_freq","spu_click_freq","spu_last_date",\
                                "brand_id","brand_buy_num","brand_click_num","brand_buy_date","brand_click_date","brand_num_ratio","brand_date_ratio","brand_buy_freq","brand_click_freq","brand_last_date",\
                                "cat_id","cat_buy_num","cat_click_num","cat_buy_date","cat_click_date","cat_num_ratio","cat_date_ratio","cat_buy_freq","cat_click_freq","cat_last_date",\
                                "action_type"]) 
    #人特征
    label_ds["u_id"] = label_ds["u_id"].astype("int")
    label_ds["u_buy_num"] = label_ds["u_buy_num"].astype("int")
    label_ds["u_click_num"] = label_ds["u_click_num"].astype("int")
    label_ds["u_buy_date"] = label_ds["u_buy_date"].astype("int")
    label_ds["u_click_date"] = label_ds["u_click_date"].astype("int")
    label_ds["u_num_ratio"] = label_ds["u_num_ratio"].astype("float")
    label_ds["u_date_ratio"] = label_ds["u_date_ratio"].astype("float")
    label_ds["u_buy_freq"] = label_ds["u_buy_freq"].astype("float")
    label_ds["u_click_freq"] = label_ds["u_click_freq"].astype("float")
    label_ds["u_last_date"] = label_ds["u_last_date"].astype("int")
    #商品特征
    label_ds["spu_id"] = label_ds["spu_id"].astype("int")
    label_ds["spu_buy_num"] = label_ds["spu_buy_num"].astype("int")
    label_ds["spu_click_num"] = label_ds["spu_click_num"].astype("int")
    label_ds["spu_buy_date"] = label_ds["spu_buy_date"].astype("int")
    label_ds["spu_click_date"] = label_ds["spu_click_date"].astype("int")
    label_ds["spu_num_ratio"] = label_ds["spu_num_ratio"].astype("float")
    label_ds["spu_date_ratio"] = label_ds["spu_date_ratio"].astype("float")
    label_ds["spu_buy_freq"] = label_ds["spu_buy_freq"].astype("float")
    label_ds["spu_click_freq"] = label_ds["spu_click_freq"].astype("float")
    label_ds["spu_last_date"] = label_ds["spu_last_date"].astype("int")
    #品牌特征
    label_ds["brand_id"] = label_ds["brand_id"].astype("int")
    label_ds["brand_buy_num"] = label_ds["brand_buy_num"].astype("int")
    label_ds["brand_click_num"] = label_ds["brand_click_num"].astype("int")
    label_ds["brand_buy_date"] = label_ds["brand_buy_date"].astype("int")
    label_ds["brand_click_date"] = label_ds["brand_click_date"].astype("int")
    label_ds["brand_num_ratio"] = label_ds["brand_num_ratio"].astype("float")
    label_ds["brand_date_ratio"] = label_ds["brand_date_ratio"].astype("float")
    label_ds["brand_buy_freq"] = label_ds["brand_buy_freq"].astype("float")
    label_ds["brand_click_freq"] = label_ds["brand_click_freq"].astype("float")
    label_ds["brand_last_date"] = label_ds["brand_last_date"].astype("int")
    #品类特征
    label_ds["cat_id"] = label_ds["cat_id"].astype("int")
    label_ds["cat_buy_num"] = label_ds["cat_buy_num"].astype("int")
    label_ds["cat_click_num"] = label_ds["cat_click_num"].astype("int")
    label_ds["cat_buy_date"] = label_ds["cat_buy_date"].astype("int")
    label_ds["cat_click_date"] = label_ds["cat_click_date"].astype("int")
    label_ds["cat_num_ratio"] = label_ds["cat_num_ratio"].astype("float")
    label_ds["cat_date_ratio"] = label_ds["cat_date_ratio"].astype("float")
    label_ds["cat_buy_freq"] = label_ds["cat_buy_freq"].astype("float")
    label_ds["cat_click_freq"] = label_ds["cat_click_freq"].astype("float")
    label_ds["cat_last_date"] = label_ds["cat_last_date"].astype("int")
    #标记
    label_ds["action_type"] = label_ds["action_type"].astype("int")
    print "训练集，有", label_ds.shape[0], "行", label_ds.shape[1], "列" 
    #加载未标记数据
    unlabel_ds=pd.read_csv(r"test_features_0714.txt",sep='\t',encoding='utf8',
                         names=["id","uid","spu_id","brand_id","cat_id",\
                                "u_buy_num","u_click_num","u_buy_date","u_click_date","u_num_ratio","u_date_ratio","u_buy_freq","u_click_freq","u_last_date",\
                                "spu_buy_num","spu_click_num","spu_buy_date","spu_click_date","spu_num_ratio","spu_date_ratio","spu_buy_freq","spu_click_freq","spu_last_date",\
                                "brand_buy_num","brand_click_num","brand_buy_date","brand_click_date","brand_num_ratio","brand_date_ratio","brand_buy_freq","brand_click_freq","brand_last_date",\
                                "cat_buy_num","cat_click_num","cat_buy_date","cat_click_date","cat_num_ratio","cat_date_ratio","cat_buy_freq","cat_click_freq","cat_last_date",]) 
    #人特征
    unlabel_ds["id"] = unlabel_ds["id"].astype("int")
    unlabel_ds["u_id"] = unlabel_ds["u_id"].astype("int")
    unlabel_ds["u_buy_num"] = unlabel_ds["u_buy_num"].astype("int")#391万
    unlabel_ds["u_click_num"] = unlabel_ds["u_click_num"].astype("int")
    unlabel_ds["u_buy_date"] = unlabel_ds["u_buy_date"].astype("int")
    unlabel_ds["u_click_date"] = unlabel_ds["u_click_date"].astype("int")
    unlabel_ds["u_num_ratio"] = unlabel_ds["u_num_ratio"].astype("float")
    unlabel_ds["u_date_ratio"] = unlabel_ds["u_date_ratio"].astype("float")
    unlabel_ds["u_buy_freq"] = unlabel_ds["u_buy_freq"].astype("float")
    unlabel_ds["u_click_freq"] = unlabel_ds["u_click_freq"].astype("float")
    unlabel_ds["u_last_date"] = unlabel_ds["u_last_date"].astype("int")
    #商品特征
    unlabel_ds["spu_id"] = unlabel_ds["spu_id"].astype("int")
    unlabel_ds["spu_buy_num"] = unlabel_ds["spu_buy_num"].astype("int")
    unlabel_ds["spu_click_num"] = unlabel_ds["spu_click_num"].astype("int")
    unlabel_ds["spu_buy_date"] = unlabel_ds["spu_buy_date"].astype("int")
    unlabel_ds["spu_click_date"] = unlabel_ds["spu_click_date"].astype("int")
    unlabel_ds["spu_num_ratio"] = unlabel_ds["spu_num_ratio"].astype("float")#241万
    unlabel_ds["spu_date_ratio"] = unlabel_ds["spu_date_ratio"].astype("float")
    unlabel_ds["spu_buy_freq"] = unlabel_ds["spu_buy_freq"].astype("float")
    unlabel_ds["spu_click_freq"] = unlabel_ds["spu_click_freq"].astype("float")
    unlabel_ds["spu_last_date"] = unlabel_ds["spu_last_date"].astype("int")
    #品牌特征
    unlabel_ds["brand_id"] = unlabel_ds["brand_id"].astype("int")
    unlabel_ds["brand_buy_num"] = unlabel_ds["brand_buy_num"].astype("int")
    unlabel_ds["brand_click_num"] = unlabel_ds["brand_click_num"].astype("int")
    unlabel_ds["brand_buy_date"] = unlabel_ds["brand_buy_date"].astype("int")
    unlabel_ds["brand_click_date"] = unlabel_ds["brand_click_date"].astype("int")
    unlabel_ds["brand_num_ratio"] = unlabel_ds["brand_num_ratio"].astype("float")
    unlabel_ds["brand_date_ratio"] = unlabel_ds["brand_date_ratio"].astype("float")
    unlabel_ds["brand_buy_freq"] = unlabel_ds["brand_buy_freq"].astype("float")
    unlabel_ds["brand_click_freq"] = unlabel_ds["brand_click_freq"].astype("float")
    unlabel_ds["brand_last_date"] = unlabel_ds["brand_last_date"].astype("int")
    #品类特征
    unlabel_ds["cat_id"] = unlabel_ds["cat_id"].astype("int")
    unlabel_ds["cat_buy_num"] = unlabel_ds["cat_buy_num"].astype("int")
    unlabel_ds["cat_click_num"] = unlabel_ds["cat_click_num"].astype("int")
    unlabel_ds["cat_buy_date"] = unlabel_ds["cat_buy_date"].astype("int")
    unlabel_ds["cat_click_date"] = unlabel_ds["cat_click_date"].astype("int")
    unlabel_ds["cat_num_ratio"] = unlabel_ds["cat_num_ratio"].astype("float")
    unlabel_ds["cat_date_ratio"] = unlabel_ds["cat_date_ratio"].astype("float")
    unlabel_ds["cat_buy_freq"] = unlabel_ds["cat_buy_freq"].astype("float")
    unlabel_ds["cat_click_freq"] = unlabel_ds["cat_click_freq"].astype("float")
    unlabel_ds["cat_last_date"] = unlabel_ds["cat_last_date"].astype("int")
    print "测试集，有", unlabel_ds.shape[0], "行", unlabel_ds.shape[1], "列" 
        
    #模型训练
    ds_0=label_ds[label_ds['action_type']==0]#标记为0的样本
    ds_0_train=ds_0.sample(frac=0.01)#抽0.01出来训练
    ds_1=label_ds[label_ds['action_type']==1]#标记为1的样本
    ds_train=ds_1.append(ds_0_train)
    label_X=ds_train[['u_num_ratio','spu_num_ratio','brand_num_ratio','cat_num_ratio']]
    label_X_scale=preprocessing.scale(label_X)#归一化
    label_y = ds_train['action_type']#类别  ds=label_ds[label_ds['action_type']==0]
    model =LogisticRegression()#ensemble.GradientBoostingClassifier()
    model.fit(label_X_scale, label_y)   
    #第五步：模型验证和选择
    test_df=ds_train.sample(frac=0.2)#抽0.2验证
    test_X=test_df[['u_num_ratio','spu_num_ratio','brand_num_ratio','cat_num_ratio']]
    test_X_scale=preprocessing.scale(test_X)#归一化
    test_y=test_df['action_type']#类别
    predicted = model.predict(test_X_scale)  
    f1_score = metrics.f1_score(test_y, predicted) #模型评估  
    print f1_score
    #第六步：模型预测
    unlabe_X = unlabel_ds[['u_num_ratio','spu_num_ratio','brand_num_ratio','cat_num_ratio']]
    unlabe_X_scale=preprocessing.scale(unlabe_X)#归一化
    unlabel_y=model.predict_proba(unlabe_X_scale)[:,1]#预测返回概率值，通过概率值阈值选择正例样本 
    out_y=pd.DataFrame(unlabel_y,columns=['prob']) #返回判定正例的比例 
    out_y["prob"]=out_y["prob"].apply(lambda x: '{0:.3f}'.format(x))
    out_1=out_y[out_y["prob"]>'0.5'] #看大于0.5的个数
    print out_1.shape
    out_y['prob'].value_counts() #看值分布
    out_y.to_csv('fangjs/outvip.txt',index=False,header=None)#输出预测数据 
    
#执行
if __name__ == '__main__':  
    start = time.clock()  
    main()
    end = time.clock()  
    print('finish all in %s' % str(end - start))