一、前述
根据前文中架构,本文我们讨论线下部分构建训练集部分。因为我们离线部分模型的选择是逻辑回归,所以我们数据必须有x和y.
二、具体流程
1.从数据库中分离出我们需要的数据。
用户行为表(日志)
用户历史下载表
商品词表(商品的基本特征)
2.构建训练集中的关联特征
流程:
2.构建训练集中的基本特征
总结:注意特征名离散化因为如果特征不离散化会造成数据之间有关系。
三、具体构建过程
1、hive建表
真实的生产场景涉及到大概五十张表的字段,这里全部简化流程,直接给出最终的三张表:
应用词表:
CREATE EXTERNAL TABLE IF NOT EXISTS dim_rcm_hitop_id_list_ds
(
hitop_id STRING,
name STRING,
author STRING,
sversion STRING,
ischarge SMALLINT,
designer STRING,
font STRING,
icon_count INT,
stars DOUBLE,
price INT,
file_size INT,
comment_num INT,
screen STRING,
dlnum INT
)row format delimited fields terminated by '\t';
/**
*
*模拟app的商品词表
hitop_id STRING, 应用软件ID
name STRING, 名称
author STRING, 作者
sversion STRING, 版本号
ischarge SMALLINT, 收费软件
designer STRING, 设计者
font STRING, 字体
icon_count INT, 有几张配图
stars DOUBLE, 评价星级
price INT, 价格
file_size INT, 大小
comment_num INT, 评论数据
screen STRING, 分辨率
dlnum INT 下载数量
*/
用户历史下载表:
CREATE EXTERNAL TABLE IF NOT EXISTS dw_rcm_hitop_userapps_dm
(
device_id STRING,
devid_applist STRING,
device_name STRING,
pay_ability STRING
)row format delimited fields terminated by '\t';
正负例样本(用户当前行为即日志)表:
CREATE EXTERNAL TABLE IF NOT EXISTS dw_rcm_hitop_sample2learn_dm
(
label STRING,
device_id STRING,
hitop_id STRING,
screen STRING,
en_name STRING,
ch_name STRING,
author STRING,
sversion STRING,
mnc STRING,
event_local_time STRING,
interface STRING,
designer STRING,
is_safe INT,
icon_count INT,
update_time STRING,
stars DOUBLE,
comment_num INT,
font STRING,
price INT,
file_size INT,
ischarge SMALLINT,
dlnum INT
)row format delimited fields terminated by '\t';
2、load数据
分别往三张表load数据: 商品词表: load data local inpath '/opt/sxt/recommender/script/applist.txt' into table dim_rcm_hitop_id_list_ds; 用户历史下载表: load data local inpath '/opt/sxt/recommender/script/userdownload.txt' into table dw_rcm_hitop_userapps_dm; 正负例样本表: load data local inpath '/opt/sxt/recommender/script/sample.txt' into table dw_rcm_hitop_sample2learn_dm;
3、构建训练数据
3.1创建临时表
创建处理数据时所需要的临时表 CREATE TABLE IF NOT EXISTS tmp_dw_rcm_hitop_prepare2train_dm ( device_id STRING, label STRING, hitop_id STRING, screen STRING, ch_name STRING, author STRING, sversion STRING, mnc STRING, interface STRING, designer STRING, is_safe INT, icon_count INT, update_date STRING, stars DOUBLE, comment_num INT, font STRING, price INT, file_size INT, ischarge SMALLINT, dlnum INT, idlist STRING, device_name STRING, pay_ability STRING )row format delimited fields terminated by '\t'; 最终保存训练集的表 CREATE TABLE IF NOT EXISTS dw_rcm_hitop_prepare2train_dm ( label STRING, features STRING )row format delimited fields terminated by '\t';
3.2 训练数据预处理过程
首先将数据从正负例样本和用户历史下载表数据加载到临时表中
INSERT OVERWRITE TABLE tmp_dw_rcm_hitop_prepare2train_dm
SELECT
t2.device_id,
t2.label,
t2.hitop_id,
t2.screen,
t2.ch_name,
t2.author,
t2.sversion,
t2.mnc,
t2.interface,
t2.designer,
t2.is_safe,
t2.icon_count,
to_date(t2.update_time),
t2.stars,
t2.comment_num,
t2.font,
t2.price,
t2.file_size,
t2.ischarge,
t2.dlnum,
t1.devid_applist,
t1.device_name,
t1.pay_ability
FROM
(
SELECT
device_id,
devid_applist,
device_name,
pay_ability
FROM
dw_rcm_hitop_userapps_dm
) t1
RIGHT OUTER JOIN
(
SELECT
device_id,
label,
hitop_id,
screen,
ch_name,
author,
sversion,
IF (mnc IN ('00','01','02','03','04','05','06','07'), mnc,'x') AS mnc,
interface,
designer,
is_safe,
IF (icon_count <= 5,icon_count,6) AS icon_count,
update_time,
stars,
IF ( comment_num IS NULL,0,
IF ( comment_num <= 10,comment_num,11)) AS comment_num,
font,
price,
IF (file_size <= 2*1024*1024,2,
IF (file_size <= 4*1024*1024,4,
IF (file_size <= 6*1024*1024,6,
IF (file_size <= 8*1024*1024,8,
IF (file_size <= 10*1024*1024,10,
IF (file_size <= 12*1024*1024,12,
IF (file_size <= 14*1024*1024,14,
IF (file_size <= 16*1024*1024,16,
IF (file_size <= 18*1024*1024,18,
IF (file_size <= 20*1024*1024,20,21)))))))))) AS file_size,
ischarge,
IF (dlnum IS NULL,0,
IF (dlnum <= 50,50,
IF (dlnum <= 100,100,
IF (dlnum <= 500,500,
IF (dlnum <= 1000,1000,
IF (dlnum <= 5000,5000,
IF (dlnum <= 10000,10000,
IF (dlnum <= 20000,20000,20001)))))))) AS dlnum
FROM
dw_rcm_hitop_sample2learn_dm
) t2
ON (t1.device_id = t2.device_id);
选择右外关联的原因是因为以用户行为为基准。
这张表得到的数据就是关联特征中的数据,截图如下:
然后再利用python脚本处理格式 这里要先讲python脚本加载到hive中 ADD FILE /opt/sxt/recommender/script/dw_rcm_hitop_prepare2train_dm.py; 可以通过list files;查看是不是python文件加载到了hive
在hive中使用python脚本处理数据的原理: Hive会以输出流的形式将数据交给python脚本,python脚本以输入流的形式来接受数据,接受来数据以后,在python中就行一系列的数据处理,处理完毕后,又以输出流的形式交给Hive,交给了hive就说明了就处理后的数据成功保存到hive表中了。
INSERT OVERWRITE TABLE dw_rcm_hitop_prepare2train_dm
SELECT
TRANSFORM (t.*)
USING 'python dw_rcm_hitop_prepare2train_dm.py'
AS (label,features)
FROM
(
SELECT
label,
hitop_id,
screen,
ch_name,
author,
sversion,
mnc,
interface,
designer,
icon_count,
update_date,
stars,
comment_num,
font,
price,
file_size,
ischarge,
dlnum,
idlist,
device_name,
pay_ability
FROM
tmp_dw_rcm_hitop_prepare2train_dm
) t;
python处理流程:
#! /usr/bin/env python
# -*- coding: utf-8 -*-
# ----------------------------------------------------------------------------
# File Name: dw_rcm_hitop_prepare2train_dm.py
# Copyright(C)Huawei Technologies Co.,Ltd.1998-2014.All rights reserved.
# Describe:
# Input: tmp_dw_rcm_hitop_prepare2train_dm
# Output: dw_rcm_hitop_prepare2train_dm
import sys
import codecs
import random
import math
import time
import datetime
if __name__ == "__main__":
random.seed(time.time())
for l in sys.stdin:
d = l.strip().split('\t')
if len(d) != 21:
continue
# Extract data from the line
label = d.pop(0)
hitop_id = d.pop(0)
screen = d.pop(0)
ch_name = d.pop(0)
author = d.pop(0)
sversion = d.pop(0)
mnc = d.pop(0)
interface = d.pop(0)
designer = d.pop(0)
icon_count = d.pop(0)
update_date = d.pop(0)
stars = d.pop(0)
comment_num = d.pop(0)
font = d.pop(0)
price = d.pop(0)
file_size = d.pop(0)
ischarge = d.pop(0)
dlnum = d.pop(0)
hitopids = d.pop(0)
device_name = d.pop(0)
pay_ability = d.pop(0)
# Construct feature vector
features = []
features.append(("Item.id,%s" % hitop_id, 1))
features.append(("Item.screen,%s" % screen, 1))
features.append(("Item.name,%s" % ch_name, 1))
features.append(("All,0",1))
features.append(("Item.author,%s" % author, 1))
features.append(("Item.sversion,%s" % sversion, 1))
features.append(("Item.network,%s" % mnc, 1))
features.append(("Item.dgner,%s" % designer, 1))
features.append(("Item.icount,%s" % icon_count, 1))
features.append(("Item.stars,%s" % stars, 1))
features.append(("Item.comNum,%s" % comment_num,1))
features.append(("Item.font,%s" % font,1))
features.append(("Item.price,%s" % price,1))
features.append(("Item.fsize,%s" % file_size,1))
features.append(("Item.ischarge,%s" % ischarge,1))
features.append(("Item.downNum,%s" % dlnum,1))
####User.Item and User.Item*Item
idlist = hitopids[:-2].split(',')
idCT = 0;
for id in idlist:
features.append(("User.Item*Item,%s" % id +'*'+hitop_id, 1))
idCT += 1
if idCT >= 3: #取每一个用户的前3个下载历史进行关联,因为用户量比较多,所以这里最后结果覆盖还是比较全的。
break;
features.append(("User.phone*Item,%s" % device_name + '*' + hitop_id,1))#升维
features.append(("User.pay*Item.price,%s" % pay_ability + '*' + price,1))
# Output
output = "%s\t%s" % (label, ",".join([ "%s:%d" % (f, v) for f, v in features ]))#这里join相当于是把list中的数据进行拆分,然后添加上分号。
print output
经过上述处理之后的数据如图所示:
特征工程部分前期准别结束。