用户行为分析

1、数据集，某宝数据。

下载数据后，利用kettle将文本文件导入mysql数据库。数据time从 2014-11-18 00到2014-12-18 23。包含如下字段：

 item_id（产品ID），behavior_type（包含点击、收藏、加购物车、购买四种行为，分别用数字1、2、3、4表示），user_geohash（地理位置），item_category（品类ID），time（发生行为的时间）。

 2、目标

• 整体用户购物情况
• 用户行为转化漏斗
• 购买率高和购买率为 0 的人群有什么特征
• 基于时间维度了解用户的行为习惯
• 基于RFM模型的用户分析

3、数据清洗处理

1）缺失值

user_geohash列含大量NULL值，后续不对该字段处理

2）数据一致化处理

time字段含年月日和小时，将小时单独分解为1列。time列之前新建1列time的副本date 存年月日，time列存时间。方法用到replace和substring_index(按特点分隔符，分隔后索要截取字符串)

　　alter table exc1 add date varchar(20) not null after item_category;

　　update exc1 set date = time

　　update exc1 set date = replace(date,date,substring_index(date,' ',1))

　　update exc1 set time = replace(time, time, substring_index(time, ' ', -1 ))

3）behavior_type列4个值（点击1、收藏2、加购物车3、购买4）转换为‘pv’,'fav','cart','buy'

　　update exc1 set  behavior_type = replace(behavior_type ,1,'pv')；

　　update exc1 set  behavior_type = replace(behavior_type ,2,'fav')

　　update exc1 set  behavior_type = replace(behavior_type ,3,'cart')

　　update exc1 set  behavior_type = replace(behavior_type ,4,'buy')　

4) 查看表结构，发现date类型不是date型，将其转换为date型

 alter table exc1 modify date date;

4、构建模型及分析

1）整体用户购物情况

pv总访问量

select  count(behavior_type) as '总访问量'  from exc1 group by behavior_type having behavior_type = 'pv';

　　

日均访问量

select date, count(behavior_type) as '日均访问量' from exc1 where behavior_type = 'pv' group by date order by date ;

　　

UV(用户总数)

select count(distinct user_id) '用户总数' from exc1;

　　

 有购买行为的用户数量

 select count(distinct user_id) '购买用户数量' from exc1 where behavior_type = 'buy';

　　

 用户的购物情况

create view user_behavior as 
select user_id, count(behavior_type),
sum(case when behavior_type = 'pv' then 1 else 0 end)  as '点击次数',
sum(case when behavior_type = 'fav' then 1 else 0 end)  as '收藏次数',
sum(case when behavior_type = 'cart' then 1 else 0 end)  as '加购数',
sum(case when behavior_type = 'buy' then 1 else 0 end)  as '购买次数'
from exc1 
group by user_id
order by count(behavior_type) desc;

　　

　　

  

 复购率

select concat(round(sum(case when 购买次数>=2 then 1 else 0 end)/sum(case when 购买次数>0 then 1 else 0 end)*100), '%') as '复购率'
from user_behavior

2）用户购买行为漏斗

用户购物行为总计

select sum(点击次数) '总点击数', sum(收藏次数) '收藏总数',sum(加购数) '加购物车总数', sum(购买次数) '购买总数'  from user_behavior;

　　

用户购买行为转换率

select CONCAT( round(sum(点击次数)*100/sum(点击次数),2),'%') as 'pv',
CONCAT( round((sum(收藏次数)+sum(加购数))*100/sum(点击次数),2),'%') as 'pv_to_favcart',
CONCAT( round(sum(购买次数)*100/sum(点击次数),2),'%') as 'pv_to_buy'
from user_behavior;

　

制作动态漏斗图：

import pandas as pd 
import pyecharts as pec 

dict = {'pv':100, 'pv_to_favcart':5.07,'pv_to_buy':1.04} 
user = pd.DataFrame(data = dict,index= range(1))

attr = ['点击','收藏或加购物车','购买']
value = [user.ix[:,'pv'],user.ix[:,'pv_to_favcart'],user.ix[:,'pv_to_buy']]
funnel = pec.Funnel('用户行为漏斗图',width=800,height = 600,title_pos = 'left')
funnel.add(name = '用户行为情况' ,
		   attr = attr, 
		   value = value,
		   is_label_show = True,
		   label_formatter ='{b}{c}%',
		   label_pos = 'outside',
		   is_legend_show = True
		   )

funnel.render()

　　

 

 可以看出用户点击后收藏和加购物车的概率在5.04%左右，真正购买的只有1.04%，购买转化率比较低，说明后期还有很大的空间。

3）购买率高和购买率为0的人群特征分析

购买率高的人群特征（购买率降序排序）：

select user_id, 点击次数,收藏次数,加购数,购买次数,
round(购买次数/点击次数*100,2) as 购买率
from user_behavior
group by user_id
order by 购买率 desc ;

　　

　　

按购买次数进行排序（购买次数降序排序）

select user_id, 点击次数,收藏次数,加购数,购买次数,
concat(round(购买次数/点击次数*100,2),'%') as 购买率
from user_behavior
group by user_id
order by 购买次数 desc;

　　

 可以发现购买率高点击次数并不高，一部分人点击2次就购买了，这部分人没有收藏也没有加入购物车，属于有目的的购物群体，缺啥买啥，一般属于理智型购物群体。

购买率为0的人群：

select user_id, 点击次数,收藏次数,加购数,购买次数,
round(购买次数/点击次数*100,2) as 购买率
from user_behavior
group by user_id
order by 购买率 asc

　　

点击次数多，加购物车或收藏较多考虑有可能是为商家活动做准备。 

点击次数多，购买率低或为0的客户为克制型客户，此类客户爱比较，思虑多，自制性比较强。或者说不会支付。

4）基于时间维度了解用户行为习惯

一天中用户活跃时段分布

select time, count(behavior_type),
sum(case when behavior_type = 'pv' then 1 else 0 end) as '点击次数',
sum(case when behavior_type = 'fav' then 1 else 0 end) as '收藏次数',
sum(case when behavior_type = 'cart' then 1 else 0 end) as '加购数',
sum(case when behavior_type = 'buy' then 1 else 0 end) as '购买次数'
from exc1
group by time 
order by time 

　　

mysql数据导出为csv表：

select time, count(behavior_type),
sum(case when behavior_type = 'pv' then 1 else 0 end) as '点击次数',
sum(case when behavior_type = 'fav' then 1 else 0 end) as '收藏次数',
sum(case when behavior_type = 'cart' then 1 else 0 end) as '加购数',
sum(case when behavior_type = 'buy' then 1 else 0 end) as '购买次数'
from exc1
group by time 
order by time
into outfile 'E:/Pro/users.csv' fields terminated by ','enclosed by '"'lines terminated by '
';

  绘制双轴图：

import pandas as pd 
import pyecharts as pec 

names = ['time','count(behavior_type)','点击次数','收藏次数','加购数','购买次数']
data = pd.read_csv('E:/Pro/users.csv',header=None,names = names)
attr = ["{}时".format(i) for i in range(24)]


bar = pec.Bar('用户行为-时间图',width =1500,height= 600)
bar.add('用户行为总数量',attr, data.iloc[:,1],is_label_show=False)
bar.add('点击总量',attr, data.iloc[:,2])

line = pec.Line()
for i in range(3,len(names)):
	line.add(names[i],attr,data.iloc[:,i])

overlap = pec.Overlap()
overlap.add(bar)
overlap.add(line, yaxis_index=1, is_add_yaxis = True)
overlap.show_config()
overlap.render('./times.html')

　　

 由图形可以看出每天0-5点用户活跃度保存降低状态，4-5点至最低值，6-10点用户活跃度快速上升，10-18点活跃度保存平稳状态，17-23点活跃度持续上升，21-22时到达最高峰，23时有所下降。与用户的运营行为可以参与用户活跃时段进行。

一周中用户活跃分布

select date_format(date,'%W') as weeks, count(behavior_type),
sum(case when behavior_type = 'pv' then 1 else 0 end) as '点击次数',
sum(case when behavior_type = 'fav' then 1 else 0 end) as '收藏次数',
sum(case when behavior_type = 'cart' then 1 else 0 end) as '加购数',
sum(case when behavior_type = 'buy' then 1 else 0 end) as '购买次数'
from exc1
where date between '2014-11-23' and '2014-12-13'
group by weeks
order by weeks 

　　

 同样绘制双轴图：

 用户活跃度到周五上升比较明显，周五到周末小幅度下降，到周日又有慢慢的回升趋势。星期4到星期5用户活跃度激增，考虑可能是双12带来的影响。

 5) 基于RFM模型找出有价值的客户

R-Recency: 最近一次购买时间
F-Frequency: 消费频率
M-Money：消费金额

数据集没有消费金额，对最近一次购买时间和消费频率进行评分

• 针对R-Recency评分（间隔天数越少客户价值越大，间隔天数升序排序）

select a.* , 
(@r:=@r+1) as recent_rank from (
select user_id, datediff('2014-12-19', max(date)) as recent from exc1 
where behavior_type = 'buy'
group by user_id
order by recent
)a,(select @r:=0)b ;

　

•  针对消费频率（F-Frequency）（购买频率越大，客户价值越大）

select a.* , 
(@r2:=@r2+1) as freq_rank from (
select user_id, count(behavior_type) as frequency from exc1 
where behavior_type = 'buy'
group by user_id
order by frequency desc 
)a,(select @r2:=0)b 

　　

•  联合合并2者，并加入评分关系

select m.user_id , n.frequency , recent_rank, freq_rank,
concat(
case when recent_rank <= (4330)/4 then '4'
when recent_rank > (4330)/4 and recent_rank <= (4330)/2 then '3'
when recent_rank > (4330)/2 and recent_rank <= (4330)/4*3 then '2'
else '1' end ,
case when freq_rank <= (4330)/4 then '4'
when freq_rank > (4330)/4 and freq_rank <= (4330)/2 then '3'
when freq_rank > (4330)/2 and freq_rank <= (4330)/4*3 then '2'
else '1' end
) as user_value
from (
select a.*,(@r1:=@r1+1) as recent_rank from (
select user_id, datediff('2014-12-19',max(date)) as recent 
from exc1  
where behavior_type = 'buy' 
group by user_id  order by recent
) a, (select @r1:=0 )as b) m ,
( select a.* , (@r2:=@r2+1) as freq_rank from (
select user_id, count(behavior_type) as frequency 
from exc1
where behavior_type = 'buy'
group by user_id order by frequency desc
)a ,(select @r2:=0) as b) as n
where m.user_id = n.user_id;

　　

 通过得分可以看出user_value 为‘41’的用户为关注频次比较高，购买能力不足的用户，可以选择适当促销打折或其他捆绑销售来促进客户下单

而user_value 为‘14’的用户这类关注度忠诚度不高，购买力强的客户需要关注其购物习性做精准化营销。