当前位置:网站首页>Market segmentation of supermarket customers based on purchase behavior data (RFM model)
Market segmentation of supermarket customers based on purchase behavior data (RFM model)
2022-07-06 06:39:00 【Nothing (sybh)】
Catalog
subject :
Market segmentation of supermarket customers based on purchase behavior data : Existing supermarket customers' purchasing behavior RFM Data sets ( Data files :RFM data .txt), Please use various clustering algorithms to segment customer groups . Complete the following questions :
1) Analyze customers' purchasing behavior RFM Data set ,R\F\M What are the distribution characteristics of these three variables ?(10 branch )
2) Try to divide the purchase into 4 class , And analyze the buying behavior characteristics of each customer .(10 branch )
3) Evaluation model , And analyze and gather 4 Whether the class is appropriate .(30 branch )
One 、 What do you mean RFM?
RFM It is a model for clustering user quality , Corresponding to three indicators
R(Recency): The time interval of the user's last consumption , Measure whether users are likely to lose
F(Frequency)
: The cumulative consumption frequency of users in the recent period , Measure user stickiness
M(Money): The user's accumulated consumption amount in the recent period , Measure users' spending power and loyalty
Two 、 clustering
# Item 1 : E-commerce user quality RFM Clustering analysis
from sklearn.cluster import KMeans
from sklearn import metrics
import matplotlib.pyplot as plt
from sklearn import preprocessing
# Import and clean data
data = pd.read_table('RFM data .txt',sep=" ")
# data=pd.read_table("RFM data .txt",encoding="gbk",sep=" ")
# data.user_id = data.user_id.astype('str')
print(data.info())
print(data.describe())
X = data.values[:,1:]
# Data standardization (z_score)
Model = preprocessing.StandardScaler()
X = Model.fit_transform(X)
# iteration , Choose the right one K
ch_score = []
ss_score = []
inertia = []
for k in range(2,10):
clf = KMeans(n_clusters=k,max_iter=1000)
pred = clf.fit_predict(X)
ch = metrics.calinski_harabasz_score(X,pred)
ss = metrics.silhouette_score(X,pred)
ch_score.append(ch)
ss_score.append(ss)
inertia.append(clf.inertia_)
# Make a comparison
fig = plt.figure()
ax1 = fig.add_subplot(131)
plt.plot(list(range(2,10)),ch_score,label='ch',c='y')
plt.title('CH(calinski_harabaz_score)')
plt.legend()
ax2 = fig.add_subplot(132)
plt.plot(list(range(2,10)),ss_score,label='ss',c='b')
plt.title(' Profile factor ')
plt.legend()
ax3 = fig.add_subplot(133)
plt.plot(list(range(2,10)),inertia,label='inertia',c='g')
plt.title('inertia')
plt.legend()
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['font.serif'] = ['SimHei'] # Set the normal display of Chinese
plt.show()
This time, we use 3 The clustering quality is comprehensively determined by three indicators ,CH, Contour coefficient and inertia fraction , The first 1 And the 3 The bigger the better , The contour coefficient is closer to 1 The better , Taken together , Gather into 4 Class effect is better
CH The bigger the indicator , The better the clustering effect is
clf.inertia_ It is a clustering evaluation index , I often use this . Talk about his shortcomings : This evaluation parameter represents the sum of the distances from a certain point in the cluster to the cluster , Although this method shows the fineness of clustering when the evaluation parameters are the smallest , But in this case, the division will be too fine , And it does not consider maximizing the distance from the point outside the cluster , therefore , I recommend method 2 :
Use the contour coefficient method K Choice of value , Here it is , I need to explain the contour coefficient , And why the contour coefficient is selected as the standard of internal evaluation , The formula of contour coefficient is :S=(b-a)/max(a,b), among a Is the average distance between a single sample and all samples in the same cluster ,b Is the average of all samples from a single sample to different clusters .
# According to the best K value , Clustering results
model = KMeans(n_clusters=4,max_iter=1000)
model.fit_predict(X)
labels = pd.Series(model.labels_)
centers = pd.DataFrame(model.cluster_centers_)
result1 = pd.concat([centers,labels.value_counts().sort_index(ascending=True)],axis=1) # Put the cluster center and the number of clusters together
result1.columns = list(data.columns[1:]) + ['counts']
print(result1)
result = pd.concat([data,labels],axis=1) # Put the original data and clustering results together
result.columns = list(data.columns)+['label'] # Change column names
pd.options.display.max_columns = None # Set to show all columns
print(result.groupby(['label']).agg('mean')) # Calculate the mean value of each index in groups
# Map the clustering results
fig = plt.figure()
ax1= fig.add_subplot(131)
ax1.plot(list(range(1,5)),result1.R,c='y',label='R')
plt.title('R indicators ')
plt.legend()
ax2= fig.add_subplot(132)
ax2.plot(list(range(1,5)),result1.F,c='b',label='F')
plt.title('F indicators ')
plt.legend()
ax3= fig.add_subplot(133)
ax3.plot(list(range(1,5)),result1.M,c='g',label='M')
plt.title('M indicators ')
plt.legend()
plt.show()
3、 ... and 、 Result analysis
1. Losing customers : Long consumption cycle 、 Less consumption 、 Consumption capacity is almost zero
2. Active customers : The consumption cycle is short 、 More times of consumption 、 Strong consumption ability
3.VIP customer : The consumption cycle is general 、 Consumption times are average 、 Consumption capacity is very strong
4. Wandering customers : The consumption cycle is general 、 Less consumption 、 Average spending power
1. Losing customers : Long consumption cycle 、 Less consumption 、 Consumption capacity is almost zero
2.vip customer : The consumption cycle is short 、 More times of consumption 、 Consumption capacity is very strong
3. Ordinary customers : The consumption cycle is general 、 Consumption times are average 、 Average spending power
边栏推荐
- SSO流程分析
- Private cloud disk deployment
- Data security -- 13 -- data security lifecycle management
- Changes in the number of words in English papers translated into Chinese
- CS-证书指纹修改
- leetcode 24. Exchange the nodes in the linked list in pairs
- In English translation of papers, how to do a good translation?
- 金融德语翻译,北京专业的翻译公司
- [Tera term] black cat takes you to learn TTL script -- serial port automation skill in embedded development
- On the first day of clock in, click to open a surprise, and the switch statement is explained in detail
猜你喜欢
女生学软件测试难不难 入门门槛低,学起来还是比较简单的
[ 英语 ] 语法重塑 之 英语学习的核心框架 —— 英语兔学习笔记(1)
LeetCode 731. My schedule II
机器学习植物叶片识别
生物医学英文合同翻译,关于词汇翻译的特点
ECS accessKey key disclosure and utilization
论文摘要翻译,多语言纯人工翻译
MySQL is sorted alphabetically
详解SQL中Groupings Sets 语句的功能和底层实现逻辑
It is necessary to understand these characteristics in translating subtitles of film and television dramas
随机推荐
基於JEECG-BOOT的list頁面的地址欄參數傳遞
The internationalization of domestic games is inseparable from professional translation companies
云服务器 AccessKey 密钥泄露利用
国产游戏国际化离不开专业的翻译公司
Private cloud disk deployment
女生学软件测试难不难 入门门槛低,学起来还是比较简单的
(practice C language every day) reverse linked list II
一文读懂简单查询代价估算
Data type of MySQL
私人云盘部署
Introduction and underlying analysis of regular expressions
How to convert flv file to MP4 file? A simple solution
How effective is the Chinese-English translation of international economic and trade contracts
Luogu p2089 roast chicken
記一個基於JEECG-BOOT的比較複雜的增删改功能的實現
Py06 字典 映射 字典嵌套 键不存在测试 键排序
LeetCode每日一题(1997. First Day Where You Have Been in All the Rooms)
Making interactive page of "left tree and right table" based on jeecg-boot
Facebook AI & Oxford proposed a video transformer with "track attention" to perform SOTA in video action recognition tasks
查询字段个数