E-commerce data mining based on machine learning | Data exploration

This article is reproduced from the WeChat public account "Youerhut", author Youerhut. To reprint this article, please contact the Youerhut public account.

Hello everyone, I am Peter~

I recently obtained a piece of IC electronic product e-commerce data, and will conduct data analysis and mining on 3 topics later:

1. The first stage: Statistical visualization analysis based on pandas, numpy, matplotlib, seaborn, plotly and other libraries
2. The second stage: User portrait analysis based on machine learning clustering algorithm and RFM model
3. The third stage: Brand, product and product category correlation mining based on association rule algorithm

This article is the first stage, the main content includes:

• Data Preparation Processing
• Data exploration EDA
• Multi-angle comparative analysis

##Import library

In [1]:

import pandas as pd
import numpy as np

import time
import os
from datetime import datetime
import matplotlib.pyplot as plt
import seaborn as sns
%matplotlib inline
#设置中文编码和负号的正常显示
plt.rcParams['font.sans-serif']=['SimHei']
plt.rcParams['axes.unicode_minus']=False

import plotly_express as px
import plotly.graph_objects as go

import missingno as ms

from sklearn.cluster import KMeans
from sklearn.preprocessing import MinMaxScaler

Basic data information

Reading data

df = pd.read_csv(
"ic_sale.csv",
 encoding="utf-8",# 指定编码
 cnotallow={"order_id":str,"product_id":str,"category_id":str,"user_id":str} # 指定字段类型
)
df.head()

The role of the converters parameter: multiple id fields in the data are all numbers. csv or excel files are treated as numbers (expressed in scientific notation); essentially they are "string" information without any size meaning. You need to specify the type when reading.

Basic information

After reading in, check the basic information of the data:

In [3]:

# 1、数据shape

df.shape

Out[3]:

(564169, 11)

In [4]:

# 2、数据字段类型

df.dtypes

Out[4]:

event_timeobject
order_idobject
product_idobject
category_id object
category_code object
brand object
pricefloat64
user_id object
ageint64
sex object
local object
dtype: object

In [5]:

Descriptive statistical information is for numeric fields:

# 3、数据描述统计信息

df.describe()

Out[5]:

##pricecountmean

	##age
564169.000000	564169.000000
##208.269324	##33.184388	std	304.559875	10.122088
min	0.000000	16.000000
25%	23.130000	24.000000
50%	87.940000	33.000000
75%	277.750000	42.000000
max	18328.680000	50.000000

In [6]:

# 4、总共多少个不同客户

df["user_id"].nunique()

Out[6]:

6908

In [7]:

# 5、总共多少个不同品牌

df["brand"].nunique()

Out[7]:

868

In [8]:

# 6、总共多少个订单

df["order_id"].nunique()

Out[8]:

234232

In [9]:

# 7、总共多少个产品

df["product_id"].nunique()

Out[9]:

3756

数据预处理

数据筛选

从描述统计信息中发现price字段的最小值是0，应该是没有成交的数据；我们选择price大于0的信息：

In [10]:

df = df[df["price"] > 0]

缺失值处理

缺失值情况

In [11]:

df.isnull().sum()

Out[11]:

event_time0
order_id0
product_id0
category_id 0
category_code129344
brand 27215
price 0
user_id 0
age 0
sex 0
local 0
dtype: int64

可以看到缺失值体现在字段：

• category_code：类别
• brand：品牌

In [12]:

ms.bar(df,color="blue")# 缺失值可视化

plt.show()

缺失值填充

In [13]:

df.fillna("missing",inplace=True)

In [14]:

df.isnull().sum()# 填充之后无缺失值

Out[14]:

event_time 0
order_id 0
product_id 0
category_id0
category_code0
brand0
price0
user_id0
age0
sex0
local0
dtype: int64

时间字段处理

字段类型转化

读进来的数据中时间字段是object类型，需要将其转成时间格式的类型

In [15]:

df["event_time"][:5] # 处理前

Out[15]:

02020-04-24 11:50:39 UTC
12020-04-24 11:50:39 UTC
22020-04-24 14:37:43 UTC
32020-04-24 14:37:43 UTC
42020-04-24 19:16:21 UTC
Name: event_time, dtype: object

In [16]:

# 去掉最后的UTC
df["event_time"] = df["event_time"].apply(lambda x: x[:19])

In [17]:

# 时间数据类型转化：字符类型---->指定时间格式

df['event_time'] = pd.to_datetime(df['event_time'], format="%Y-%m-%d %H:%M:%S")

字段衍生

In [18]:

# 提取多个时间相关字段

df['month']=df['event_time'].dt.month
df['day'] = df['event_time'].dt.day
df['dayofweek']=df['event_time'].dt.dayofweek
df['hour']=df['event_time'].dt.hour

In [19]:

df["event_time"][:5] # 处理后

Out[19]:

0 2020-04-24 11:50:39
1 2020-04-24 11:50:39
2 2020-04-24 14:37:43
3 2020-04-24 14:37:43
4 2020-04-24 19:16:21
Name: event_time, dtype: datetime64[ns]

可以看到字段类型已经发生了变化

整体趋势分析

分析1：每月成交金额多少？

In [20]:

amount_by_month = df.groupby("month")["price"].sum().reset_index()
amount_by_month

Out[20]:

	month	price
0	1	1953358.17
1	2	2267809.88
2	3	2897486.26
3	4	1704422.41
4	5	7768637.79
5	6	7691244.33
6	7	16354029.27
7	8	27982605.44
8	9	17152310.57
9	10	19765680.76
10	11	11961511.52

In [21]:

fig = px.scatter(amount_by_month,x="month",y="price",size="price",color="price")

fig.update_layout(height=500, width=1000, title_text="每月成交金额")

fig.show()

可以看到：

• 8月份是整个销售的顶峰
• 下半年的整体销售会好于下半年

分析2：月订单量如何变化？

In [22]:

order_by_month = df.groupby("month")["order_id"].nunique().reset_index()
order_by_month

Out[22]:

	month	order_id
0	1	10353
1	2	11461
2	3	12080
3	4	9001
4	5	30460
5	6	28978
6	7	57659
7	8	73897
8	9	345
9	10	14
10	11	6

In [23]:

fig = px.line(order_by_month,x="month",y="order_id")

fig.update_layout(height=500, width=1000, title_text="每月成交订单量")

fig.show()

关于订单量：

• 从1到8月份是一个逐渐上升的趋势；尤其是4到8月份；可能是五一假期或者暑假、开学季引起的
• 9、10月份订单量陡降：开学之后销量下降快

分析3：月消费人数/人次如何变化？

In [24]:

# nunique：对每个user_id进行去重：消费人数
# count：统计user_id 的次数；消费人次（存在一人多次购买）

people_by_month = df.groupby("month")["user_id"].agg(["nunique","count"]).reset_index()
people_by_month

Out[24]:

	month	nunique	count
0	1	1388	15575
1	2	1508	17990
2	3	1597	18687
3	4	1525	11867
4	5	3168	40332
5	6	3966	41355
6	7	5159	76415
7	8	6213	100006
8	9	5497	70496
9	10	4597	104075
10	11	3134	67332

In [25]:

fig = px.line(people_by_month,x="month",y="nunique")

fig.update_layout(height=500, width=1000, title_text="每月成交人数")

fig.show()

fig = px.line(people_by_month,x="month",y="count")

fig.update_layout(height=500, width=1000, title_text="每月成交人次")

fig.show()

分析4：每月订单价多少？

In [27]:

amount_by_month# 每月成交金额

Out[27]:

	month	price
0	1	1953358.17
1	2	2267809.88
2	3	2897486.26
3	4	1704422.41
4	5	7768637.79
5	6	7691244.33
6	7	16354029.27
7	8	27982605.44
8	9	17152310.57
9	10	19765680.76
10	11	11961511.52

In [28]:

order_by_month# 每月订单数

Out[28]:

	month	order_id
0	1	10353
1	2	11461
2	3	12080
3	4	9001
4	5	30460
5	6	28978
6	7	57659
7	8	73897
8	9	345
9	10	14
10	11	6

In [29]:

amount_by_userid = pd.merge(amount_by_month,order_by_month)

amount_by_userid

Out[29]:

	month	price	order_id
0	1	1953358.17	10353
1	2	2267809.88	11461
2	3	2897486.26	12080
3	4	1704422.41	9001
4	5	7768637.79	30460
5	6	7691244.33	28978
6	7	16354029.27	57659
7	8	27982605.44	73897
8	9	17152310.57	345
9	10	19765680.76	14
10	11	11961511.52	6

In [30]:

amount_by_userid["average"] = amount_by_userid["price"] / amount_by_userid["order_id"]

amount_by_userid

fig = px.line(amount_by_userid,x="month",y="average")

fig.update_layout(height=500, width=1000, title_text="每月客单价")

fig.show()

从上面的折线图可以看出来：

1. 1到8月份月订单量基本持平；可能是有很多批量的订单；通过量大带来利润：量的路线
2. 9到10月份：月单价急剧上升；订单量少，但是金额；可能存在大额消费的用户：质的路线

分析5：每个订单包含多少产品

In [32]:

product_by_order = df.groupby("order_id")["product_id"].count().reset_index().sort_values("product_id",ascending=False)

product_by_order.head(10)

Out[32]:

	order_id	product_id
234208	2388440981134640000	15021
234210	2388440981134660000	14891
234211	2388440981134670000	14845
234212	2388440981134680000	14765
234202	2388440981134580000	14587
234205	2388440981134610000	14571
234207	2388440981134630000	14443
234204	2388440981134600000	14416
234206	2388440981134620000	14414
234203	2388440981134590000	14194

In [33]:

fig = px.bar(product_by_order[:20],
 x="order_id",
 y="product_id",
 text="product_id"
)

fig.show()

一个订单下包含的产品数量是不同；上万的订单可能是小型的ic元器件产品。

不同省份对比

分析6：订单量、用户量和成交金额对比

不同省份下的订单量、用户量和成交金额对比

In [34]:

local = df.groupby("local").agg({"order_id":"nunique","user_id":"nunique","price":sum}).reset_index()
local.head()

Out[34]:

##6433736.854Guangdong51471

In [35]:

df1 = local.sort_values("order_id",ascending=True)# 订单量升序
df1

Out[35]:

	local	order_id	user_id	price
0	上海	39354	5680	19837942.20
1	北京	38118	5702	19137748.75
2	Sichuan	13396	3589	6770891.28
##3	Tianjin	13058	3497
		##6085	26013770.86

##6968674.412四川##6357286.879 Hunan1387934816983078.88##1##56804

local

order_id

user_id

price

6

浙江

12790

3485

6522657.59

8

湖北

12810

3488

5993820.57

3

天津

13058

3497

##6433736.85

10

Chongqing

13058

3496

6479488.14

7

Hainan

13076

3587

##13396

3589

6770891.28

##5

江苏

13575

3598

Beijing

38118

5702

##19137748.75

##0

上海

##39354

##19837942.20

Guangdong

51471

6085

##26013770.86

In [36]:

fig = px.pie(df1, names="local",labels="local",values="price")

fig.update_traces(
textpositinotallow="inside",
textinfo="percent+label"
)

fig.show()

无疑：广东省No.1

每个省份的订单量对比：

fig = px.bar(df1,x="order_id",y="local",orientatinotallow="h")

fig.show()

# 整体的可视化效果

fig = px.scatter_3d(local,
x="order_id",
y="user_id",
z="price",
color="order_id",
hover_name="local"
 )

fig.show()

通过3D散点图我们发现：广东省真的是一骑绝尘！

• 订单量多；订单金额也大：主打搞钱
• 除去北上广，湖南和江苏的用户群是最多的，有前景

分析7：不同省份的客户钟爱哪些品牌？

In [39]:

local_brand = df.groupby(["local","brand"]).size().to_frame().reset_index()

local_brand.columns = ["local","brand","number"]# 修改字段名

local_brand

# 根据local和number进行排序
local_brand.sort_values(["local","number"],ascending=[True,False],inplace=True,ignore_index=True)
local_brand = local_brand[local_brand["brand"] != "missing"]
# 每个local下面最受欢迎的前3个品牌
local_brand = local_brand.groupby("local").head(3)
local_brand

fig = px.bar(local_brand,
 x="brand",
 y="number",
 color="number",
 facet_col="local")

fig.update_layout(height=500,width=1000)

fig.show()

看来大家都很喜欢: samsung 、apple、ava

不同时间对比

分析8：下单时间对比

In [43]:

df.columns

Out[43]:

Index(['event_time', 'order_id', 'product_id', 'category_id', 'category_code',
 'brand', 'price', 'user_id', 'age', 'sex', 'local', 'month', 'day',
 'dayofweek', 'hour'],
dtype='object')

In [44]:

df2 = df.groupby("dayofweek")["order_id"].nunique().reset_index()
df2

Out[44]:

	dayofweek	order_id
0	0	35690
1	1	34256
2	2	31249
3	3	31555
4	4	33010
5	5	34772
6	6	33922

In [45]:

plt.figure(figsize=(12,7))

df2["order_id"].plot.bar()

plt.xticks(range(7),['周一','周二','周三','周四','周五','周六','周日'],rotatinotallow=0)
plt.xlabel('星期')
plt.ylabel('订单量')
plt.title('订单数随星期变化')

plt.show()

分析9：每小时订单量

In [46]:

df3 = df.groupby("hour")["order_id"].nunique().reset_index()
df3.head(10)

Out[46]:

	hour	order_id
0	0	2865
1	1	2711
2	2	3981
3	3	6968
4	4	12176
5	5	16411
6	6	18667
7	7	20034
8	8	20261
9	9	20507

In [47]:

plt.figure(figsize=(14,8))
df3["order_id"].plot()

plt.xlabel('小时')
plt.ylabel('订单数量')
plt.title('订单随小时数变化')

plt.grid()
plt.show()

用户都喜欢在上午8、9、10点下单；可能是刚开始上班工作，大家更积极

不同用户消费行为分析

分析10：消费次数和消费金额

In [48]:

df4 = df.groupby("user_id").agg({"order_id":"nunique", "price":sum})

fig = px.scatter(df4,
x="order_id",
y="price",
color="price",
size="price")

fig.show()

• 同时存在低频高额和高频高额用户

分析11：用户消费周期

In [50]:

# 用户消费周期

# shift函数：移动一个单位

purchase_time=df.groupby('user_id').apply(lambda x: x['event_time'] - x['event_time'].shift()).dt.days
purchase_time

Out[50]:

user_id
151591562543995000096014NaN
1515915625440030000374760 NaN
 48492735.0
1515915625440050000463812 NaN
 473430 1.0
 ... 
1515915625514880000564132 0.0
 564143 0.0
 564164 0.0
1515915625514890000564158 NaN
 564165 0.0
Name: event_time, Length: 564130, dtype: float64

In [51]:

purchase_time[purchase_time>0].describe()

Out[51]:

count120629.000000
mean 35.494500
std 663.803583
min 1.000000
25% 2.000000
50% 4.000000
75%12.000000
max 18466.000000
Name: event_time, dtype: float64

说明：

1. 至少消费两次的用户的消费周期是4天
2. 有75%的客户消费周期在12天

分析12：用户复购行为

In [52]:

pivoted_counts = df.pivot_table(index='user_id',
 columns='month',
 values='order_id',
 aggfunc='nunique').fillna(0)

pivoted_counts

Out[52]:

pivoted_counts_map.sum() / pivoted_counts_map.count()

# 结果
month
1 0.406340
2 0.439655
3 0.474640
4 0.700328
5 0.829861
6 0.792990
7 0.891452
8 0.920328
9 0.781153
100.609963
110.419592
dtype: float64

(pivoted_counts_map.sum()/pivoted_counts_map.count()).plot(figsize=(12,6))

plt.xticks(range(11),columns_month)

plt.title('复购率')
plt.show()

• 复购的高峰期在4、6、9月份
• 10月份开始，销售开始冷淡；复购急降

The above is the detailed content of E-commerce data mining based on machine learning | Data exploration. For more information, please follow other related articles on the PHP Chinese website!