RFM模型是市场营销和CRM客户管理中经常用到的探索性分析方法,透过模型深入挖掘客户行为背后的价值规律,进而更好地利用数据价值推动业务发展和客户管理。
RFM是三种客户行为的英文缩写:
R:Recency —— 客户最近一次交易时间的间隔。R值越大,表示客户交易距今越久,反之则越近; F:Frequency—— 客户在最近一段时间内交易的次数。F值越大,表示客户交易越频繁,反之则不够活跃; M:Monetary —— 客户在最近一段时间内交易的金额。M值越大,表示客户价值越高,反之则越低。
一般通过对RFM三个原始指标进行分箱操作(分位数法),获得三个指标各自的若干个水平因子(需要注意因子水平大小的对应的实际意义)。
R_S:基于最近一次交易日期计算得分,距离当前日期越近,则得分越高,否则得分越低; F_S:基于交易频率计算得分,交易频率越高,则得分越高,否则得分越低; M_S:基于交易金额得分,交易金额越高,则得分越高,反之得分越低。
同时为了对每个客户进行综合评价,也可将以上三个得分进行加权计算(权重规则可由专家制定或者营销人员自行根据业务决定,这里统一采用100:10:1)。
RFM = 100R_S + 10F_S + 1*M_S
RFM核心便是构建在R、F、M三个指标得分构成的立方体组合内,形成一个非常直观的客户价值矩阵。
最终通过对R_S、F_S、M_S三指标的得分组合,形成八种客户价值类型,营销人员可以通过以上组合形成的客户类群,针对性的进行活动营销,进而提升客户价值和营收水平。
通过RFM分析识别优质客户,可以据此制定个性化沟通与营销服务,可以为营销决策提供更好地支持。
以下是利用R语言构建RFM模型的简要步骤:
1、数据准备:
## !/user/bin/env RStudio 1.1.423
## -*- coding: utf-8 -*-
## RFM Model
#* 最近一次消费(Recency)
#* 消费频率(Frenquency)
#* 消费金额(Monetary)
setwd('D:/R/File/')
library('magrittr')
library('dplyr')
library('scales')
library('ggplot2')
library("easyGgplot2")
library("Hmisc")
library('foreign')
library('lubridate')
mydata <- spss.get("trade.sav",datevars = '交易日期',reencode = 'GBK')
names(mydata) <- c('OrderID','UserID','PayDate','PayAmount')
start_time <- as.POSIXct("2017/01/01", format="%Y/%m/%d") %>% as.numeric()
end_time <- as.POSIXct("2017/12/31", format="%Y/%m/%d") %>% as.numeric()
set.seed(233333)
mydata$PayDate <- runif(nrow(mydata),start_time,end_time) %>% as.POSIXct(origin="1970-01-01") %>% as.Date()
mydata$interval <- difftime(max(mydata$PayDate),mydata$PayDate ,units="days") %>% round() %>% as.numeric()
salesRFM <- mydata %>% group_by(UserID) %>%
summarise(
Monetary = sum(PayAmount),
Frequency = n(),
Recency = min(interval)
)
2、计算得分
#分箱得分
salesRFM <- mutate(
salesRFM,
rankR = 6- cut(salesRFM$Recency,breaks = quantile(salesRFM$Recency, probs = seq(0, 1, 0.2),names = FALSE),include.lowest = TRUE,labels=F),
rankF = cut(salesRFM$Frequency ,breaks = quantile(salesRFM$Frequency, probs = seq(0, 1, 0.2),names = FALSE),include.lowest = TRUE,labels=F),
rankM = cut(salesRFM$Monetary ,breaks = quantile(salesRFM$Monetary, probs = seq(0, 1, 0.2),names = FALSE),include.lowest = TRUE,labels=F),
rankRMF = 100*rankR + 10*rankF + 1*rankM
)
#标准化得分(也是一种计算得分的方法)
salesRFM <- mutate(
salesRFM,
rankR1 = 1 - rescale(salesRFM$Recency,to = c(0,1)),
rankF1 = rescale(salesRFM$Frequency,to = c(0,1)),
rankM1 = rescale(salesRFM$Monetary,to = c(0,1)),
rankRMF1 = 0.5*rankR + 0.3*rankF + 0.2*rankM
)
3、客户分类:
#对R\F\M分类:
salesRFM <- within(salesRFM,{
R_S = ifelse(rankR > mean(rankR),2,1)
F_S = ifelse(rankF > mean(rankF),2,1)
M_S = ifelse(rankM > mean(rankM),2,1)
})
#客户类型归类:
salesRFM <- within(salesRFM,{
Custom = NA
Custom[R_S == 2 & F_S == 2 & M_S == 2] = '高价值客户'
Custom[R_S == 1 & F_S == 2 & M_S == 2] = '重点保持客户'
Custom[R_S == 2 & F_S == 1 & M_S == 2] = '重点发展客户'
Custom[R_S == 1 & F_S == 1 & M_S == 2] = '重点挽留客户'
Custom[R_S == 2 & F_S == 2 & M_S == 1] = '重点保护客户'
Custom[R_S == 1 & F_S == 2 & M_S == 1] = '一般保护客户'
Custom[R_S == 2 & F_S == 1 & M_S == 1] = '一般发展客户'
Custom[R_S == 1 & F_S == 1 & M_S == 1] = '潜在客户'
})
4、分析结果可视化:
4.1 查看RFM分箱后客户分布状况:
#RFM分箱计数
ggplot(salesRFM,aes(rankF)) +
geom_bar()+
facet_grid(rankM~rankR) +
theme_gray()
4.2 RFM热力图:
#RFM heatmap
heatmap_data <- salesRFM %>% group_by(rankF,rankR) %>% dplyr::summarize(M_mean = mean(Monetary))
ggplot(heatmap_data,aes(rankF,rankR,fill =M_mean )) +
geom_tile() +
scale_fill_distiller(palette = 'RdYlGn',direction = 1)
4.3 RFM直方图:
#RFM直方图
p1 <- ggplot(salesRFM,aes(Recency)) +
geom_histogram(bins = 10,fill = '#362D4C')
p2 <- ggplot(salesRFM,aes(Frequency)) +
geom_histogram(bins = 10,fill = '#362D4C')
p3 <- ggplot(salesRFM,aes(Monetary)) +
geom_histogram(bins = 10,fill = '#362D4C')
ggplot2.multiplot(p1,p2,p3, cols=3)
4.4 RFM两两交叉散点图:
#RFM 两两交叉散点图
p1 <- ggplot(salesRFM,aes(Monetary,Recency)) +
geom_point(shape = 21,fill = '#362D4C' ,colour = 'white',size = 2)
p2 <- ggplot(salesRFM,aes(Monetary,Frequency)) +
geom_point(shape = 21,fill = '#362D4C' ,colour = 'white',size = 2)
p3 <- ggplot(salesRFM,aes(Frequency,Recency)) +
geom_point(shape = 21,fill = '#362D4C' ,colour = 'white',size = 2)
ggplot2.multiplot(p1,p2,p3, cols=1)
5 数据结果导出
#导出结果数据
write.csv(salesRFM,'salesRFM.csv')
Python:
1、数据准备
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import time
import numpy as np
import pandas as pd
import savReaderWriter as spss
import os
from datetime import datetime,timedelta
np.random.seed(233333)
os.chdir('D:/R/File')
pd.set_option('display.float_format', lambda x: '%.3f' % x)
with spss.SavReader('trade.sav',returnHeader = True ,ioUtf8=True,rawMode = True,ioLocale='chinese') as reader:
mydata = pd.DataFrame(list(reader)[1:],columns = list(reader)[0])
mydata['交易日期'] = mydata['交易日期'].map(lambda x: reader.spss2strDate(x,"%Y-%m-%d", None))
mydata.rename(columns={'订单ID':'OrderID','客户ID':'UserID','交易日期':'PayDate','交易金额':'PayAmount'},inplace=True)
start_time = int(time.mktime(time.strptime('2017/01/01', '%Y/%m/%d')))
end_time = int(time.mktime(time.strptime('2017/12/31', '%Y/%m/%d')))
mydata['PayDate'] = pd.Series(np.random.randint(start_time,end_time,len(mydata))).map(lambda x: time.strftime("%Y-%m-%d", time.localtime(x)))
mydata['interval'] = [(datetime.now() - pd.to_datetime(i,format ='%Y %m %d')).days for i in mydata['PayDate']]
mydata = mydata.astype({'OrderID':'int64','UserID':'int64','PayAmount':'int64'})
print('---------#######-----------')
print(mydata.head())
print('---------#######-----------')
print(mydata.tail())
print('…………………………………………………………………………')
print(mydata.dtypes)
print('---------#######------------')
2、得分计算:
#按照用户ID聚合交易频次、交易总额及首次购买时间
mydata.set_index('UserID', inplace=True)
salesRFM = mydata.groupby(level = 0).agg({
'PayAmount': np.sum,
'PayDate': 'count',
'interval': np.min
})
# make the column names more meaningful
salesRFM.rename(columns={
'PayAmount': 'Monetary',
'PayDate': 'Frequency',
'interval':'Recency'
}, inplace=True)
salesRFM.head()
#均值划分
salesRFM = salesRFM.assign(
rankR = pd.qcut(salesRFM['Recency'], q = [0, .2, .4, .6,.8,1.] , labels = [5,4,3,2,1]),
rankF = pd.qcut(salesRFM['Frequency'],q = [0, .2, .4, .6,.8,1.] , labels = [1,2,3,4,5]),
rankM = pd.qcut(salesRFM['Monetary'] ,q = [0, .2, .4, .6,.8,1.] , labels = [1,2,3,4,5])
)
salesRFM['rankRMF'] = 100*salesRFM['rankR'] + 10*salesRFM['rankF'] + 1*salesRFM['rankM']
#特征缩放——0-1标准化
from sklearn import preprocessing
min_max_scaler = preprocessing.MinMaxScaler()
salesRFM1 = min_max_scaler.fit_transform(salesRFM.loc[:,['Recency','Frequency','Monetary']].values)
salesRFM = salesRFM.assign(
rankR1 = 1 - salesRFM1[:,0],
rankF1 = salesRFM1[:,1],
rankM1 = salesRFM1[:,2]
)
salesRFM['rankRFM1'] = 0.5*salesRFM['rankR1'] + 0.3*salesRFM['rankF1'] + 0.2*salesRFM['rankM1']
3、客户分类:
#对R\F\M分类:
salesRFM = salesRFM.astype({'rankR':'int64','rankF':'int64','rankM':'int64'})
salesRFM = salesRFM.assign(
R_S = salesRFM['rankR'].map(lambda x: 2 if x > salesRFM['rankR'].mean() else 1),
F_S = salesRFM['rankF'].map(lambda x: 2 if x > salesRFM['rankF'].mean() else 1),
M_S = salesRFM['rankM'].map(lambda x: 2 if x > salesRFM['rankM'].mean() else 1)
)
#客户类型归类:
salesRFM['Custom'] = np.NaN
salesRFM.loc[(salesRFM['R_S'] == 2) & (salesRFM['F_S'] == 2) & (salesRFM['M_S'] == 2),'Custom'] = '高价值客户'
salesRFM.loc[(salesRFM['R_S'] == 1) & (salesRFM['F_S'] == 2) & (salesRFM['M_S'] == 2),'Custom'] = '重点保持客户'
salesRFM.loc[(salesRFM['R_S'] == 2) & (salesRFM['F_S'] == 1) & (salesRFM['M_S'] == 2),'Custom'] = '重点发展客户'
salesRFM.loc[(salesRFM['R_S'] == 1) & (salesRFM['F_S'] == 1) & (salesRFM['M_S'] == 2),'Custom'] = '重点挽留客户'
salesRFM.loc[(salesRFM['R_S'] == 2) & (salesRFM['F_S'] == 2) & (salesRFM['M_S'] == 1),'Custom'] = '重点保护客户'
salesRFM.loc[(salesRFM['R_S'] == 1) & (salesRFM['F_S'] == 2) & (salesRFM['M_S'] == 1),'Custom'] = '一般保护客户'
salesRFM.loc[(salesRFM['R_S'] == 2) & (salesRFM['F_S'] == 1) & (salesRFM['M_S'] == 1),'Custom'] = '一般发展客户'
salesRFM.loc[(salesRFM['R_S'] == 1) & (salesRFM['F_S'] == 1) & (salesRFM['M_S'] == 1),'Custom'] = '潜在客户'
RFM模型仅仅是一个前期的探索性分析,可以利用RFM模型输出的指标结果还可以进行其他分类以及降维模型的构建,深入探索客户数据价值,挖掘潜在营销点。
数据文件及code可以点击下面的GitHub链接获取:
https://github.com/ljtyduyu/DataWarehouse/tree/master/Model