航空公司客户价值分析-KMeans聚类

航空公司客户价值分析-KMeans聚类

import numpy as np
import pandas as pd

data = pd.read_csv(r'***\chapter7\demo\data\air_data.csv', encoding='utf-8')
data.head()

	MEMBER_NO	FFP_DATE	FIRST_FLIGHT_DATE	GENDER	FFP_TIER	WORK_CITY	WORK_PROVINCE	WORK_COUNTRY	AGE	LOAD_TIME	...	ADD_Point_SUM	Eli_Add_Point_Sum	L1Y_ELi_Add_Points	Points_Sum	L1Y_Points_Sum	Ration_L1Y_Flight_Count	Ration_P1Y_Flight_Count	Ration_P1Y_BPS	Ration_L1Y_BPS	Point_NotFlight
0	54993	2006/11/02	2008/12/24	男	6	.	北京	CN	31.0	2014/03/31	...	39992	114452	111100	619760	370211	0.509524	0.490476	0.487221	0.512777	50
1	28065	2007/02/19	2007/08/03	男	6	NaN	北京	CN	42.0	2014/03/31	...	12000	53288	53288	415768	238410	0.514286	0.485714	0.489289	0.510708	33
2	55106	2007/02/01	2007/08/30	男	6	.	北京	CN	40.0	2014/03/31	...	15491	55202	51711	406361	233798	0.518519	0.481481	0.481467	0.518530	26
3	21189	2008/08/22	2008/08/23	男	5	Los Angeles	CA	US	64.0	2014/03/31	...	0	34890	34890	372204	186100	0.434783	0.565217	0.551722	0.448275	12
4	39546	2009/04/10	2009/04/15	男	6	贵阳	贵州	CN	48.0	2014/03/31	...	22704	64969	64969	338813	210365	0.532895	0.467105	0.469054	0.530943	39

5 rows × 44 columns

一、数据探索

explore = data.describe(percentiles = [], include = 'all').T
# percentils参数是指定计算多少的分位数，可不加该参数，此处是计算50%分位数，转置为了阅读方便
# describe（）函数自动计算的的字段有count(非空值数)，unique(唯一值数)，top(频数最高者)，freq(最高频数)，mean（平均数），std, min, 50%, max
explore.head()

	count	unique	top	freq	mean	std	min	50%	max
MEMBER_NO	62988	NaN	NaN	NaN	31494.5	18183.2	1	31494.5	62988
FFP_DATE	62988	3068	2011/01/13	184	NaN	NaN	NaN	NaN	NaN
FIRST_FLIGHT_DATE	62988	3406	2013/02/16	96	NaN	NaN	NaN	NaN	NaN
GENDER	62985	2	男	48134	NaN	NaN	NaN	NaN	NaN
FFP_TIER	62988	NaN	NaN	NaN	4.10216	0.373856	4	4	6

explore['null'] = len(data)-explore['count']  #生成一列空值数
explore.head()

	count	unique	top	freq	mean	std	min	50%	max	null
MEMBER_NO	62988	NaN	NaN	NaN	31494.5	18183.2	1	31494.5	62988	0
FFP_DATE	62988	3068	2011/01/13	184	NaN	NaN	NaN	NaN	NaN	0
FIRST_FLIGHT_DATE	62988	3406	2013/02/16	96	NaN	NaN	NaN	NaN	NaN	0
GENDER	62985	2	男	48134	NaN	NaN	NaN	NaN	NaN	3
FFP_TIER	62988	NaN	NaN	NaN	4.10216	0.373856	4	4	6	0

#导入属性信息表
shuxing = pd.read_excel(r'***\chapter7\demo\data\客户信息属性说明.xls')
shuxing.head()

	属性中文	属性英文	备注
0	会员卡号	MEMBER_NO	NaN
1	入会时间	FFP_DATE	办理会员卡的开始的时间
2	第一次飞行日期	FIRST_FLIGHT_DATE	NaN
3	性别	GENDER	NaN
4	会员卡级别	FFP_TIER	NaN

#合并表
explore1 = pd.merge(explore, shuxing, left_index=True, right_on='属性英文' ,how ='left' )
explore1

	count	unique	top	freq	mean	std	min	50%	max	null	属性中文	属性英文	备注
0	62988	NaN	NaN	NaN	31494.5	18183.2	1	31494.5	62988	0	会员卡号	MEMBER_NO	NaN
1	62988	3068	2011/01/13	184	NaN	NaN	NaN	NaN	NaN	0	入会时间	FFP_DATE	办理会员卡的开始的时间
2	62988	3406	2013/02/16	96	NaN	NaN	NaN	NaN	NaN	0	第一次飞行日期	FIRST_FLIGHT_DATE	NaN
3	62985	2	男	48134	NaN	NaN	NaN	NaN	NaN	3	性别	GENDER	NaN
4	62988	NaN	NaN	NaN	4.10216	0.373856	4	4	6	0	会员卡级别	FFP_TIER	NaN
5	60719	3310	广州	9385	NaN	NaN	NaN	NaN	NaN	2269	工作地城市	WORK_CITY	NaN
6	59740	1185	广东	17507	NaN	NaN	NaN	NaN	NaN	3248	工作地所在省份	WORK_PROVINCE	NaN
7	62962	118	CN	57748	NaN	NaN	NaN	NaN	NaN	26	工作地所在国家	WORK_COUNTRY	NaN
43	62568	NaN	NaN	NaN	42.4763	9.88591	6	41	110	420	NaN	AGE	NaN
9	62988	1	2014/03/31	62988	NaN	NaN	NaN	NaN	NaN	0	观测窗口的结束时间	LOAD_TIME	选取样本的时间宽度，距离现在最近的时间。
10	62988	NaN	NaN	NaN	11.8394	14.0495	2	7	213	0	飞行次数	FLIGHT_COUNT	频数
11	62988	NaN	NaN	NaN	10925.1	16339.5	0	5700	505308	0	观测窗口总基本积分	BP_SUM	航空公里的里程就相当于积分，积累一定分数可以兑换奖品和免费里程。
12	62988	NaN	NaN	NaN	0	0	0	0	0	0	第一年精英资格积分	EP_SUM_YR_1	NaN
13	62988	NaN	NaN	NaN	265.69	1645.7	0	0	74460	0	第二年精英资格积分	EP_SUM_YR_2	NaN
14	62437	NaN	NaN	NaN	5355.38	8109.45	0	2800	239560	551	第一年总票价	SUM_YR_1	NaN
15	62850	NaN	NaN	NaN	5604.03	8703.36	0	2773	234188	138	第二年总票价	SUM_YR_2	NaN
16	62988	NaN	NaN	NaN	17123.9	20960.8	368	9994	580717	0	观测窗口总飞行公里数	SEG_KM_SUM	NaN
17	62988	NaN	NaN	NaN	12777.2	17578.6	0	6978.26	558440	0	观测窗口总加权飞行公里数（Σ舱位折扣×航段距离）	WEIGHTED_SEG_KM	NaN
18	62988	731	2014/03/31	959	NaN	NaN	NaN	NaN	NaN	0	末次飞行日期	LAST_FLIGHT_DATE	最后一次飞行时间
19	62988	NaN	NaN	NaN	1.54215	1.787	0.25	0.875	26.625	0	观测窗口季度平均飞行次数	AVG_FLIGHT_COUNT	NaN
20	62988	NaN	NaN	NaN	1421.44	2083.12	0	752.375	63163.5	0	观测窗口季度平均基本积分累积	AVG_BP_SUM	NaN
21	62988	NaN	NaN	NaN	120.145	159.573	0	50	729	0	观察窗口内第一次乘机时间至MAX（观察窗口始端，入会时间）时长	BEGIN_TO_FIRST	NaN
22	62988	NaN	NaN	NaN	176.12	183.822	1	108	731	0	最后一次乘机时间至观察窗口末端时长	LAST_TO_END	NaN
23	62988	NaN	NaN	NaN	67.7498	77.5179	0	44.6667	728	0	平均乘机时间间隔	AVG_INTERVAL	NaN
24	62988	NaN	NaN	NaN	166.034	123.397	0	143	728	0	观察窗口内最大乘机间隔	MAX_INTERVAL	NaN
25	62988	NaN	NaN	NaN	540.317	3956.08	0	0	600000	0	观测窗口中第1年其他积分（合作伙伴、促销、外航转入等）	ADD_POINTS_SUM_YR_1	NaN
26	62988	NaN	NaN	NaN	814.689	5121.8	0	0	728282	0	观测窗口中第2年其他积分（合作伙伴、促销、外航转入等）	ADD_POINTS_SUM_YR_2	NaN
27	62988	NaN	NaN	NaN	0.319775	1.136	0	0	46	0	积分兑换次数	EXCHANGE_COUNT	NaN
28	62988	NaN	NaN	NaN	0.721558	0.185427	0	0.711856	1.5	0	平均折扣率	avg_discount	NaN
29	62988	NaN	NaN	NaN	5.76626	7.21092	0	3	118	0	第1年乘机次数	P1Y_Flight_Count	NaN
30	62988	NaN	NaN	NaN	6.07316	8.17513	0	3	111	0	第2年乘机次数	L1Y_Flight_Count	NaN
31	62988	NaN	NaN	NaN	5366.72	8537.77	0	2692	246197	0	第1年里程积分	P1Y_BP_SUM	NaN
32	62988	NaN	NaN	NaN	5558.36	9351.96	0	2547	259111	0	第2年里程积分	L1Y_BP_SUM	NaN
33	62988	NaN	NaN	NaN	265.69	1645.7	0	0	74460	0	观测窗口总精英积分	EP_SUM	NaN
34	62988	NaN	NaN	NaN	1355.01	7868.48	0	0	984938	0	观测窗口中其他积分（合作伙伴、促销、外航转入等）	ADD_Point_SUM	NaN
35	62988	NaN	NaN	NaN	1620.7	8294.4	0	0	984938	0	非乘机积分总和	Eli_Add_Point_Sum	NaN
36	62988	NaN	NaN	NaN	1080.38	5639.86	0	0	728282	0	第2年非乘机积分总和	L1Y_ELi_Add_Points	NaN
37	62988	NaN	NaN	NaN	12545.8	20507.8	0	6328.5	985572	0	总累计积分	Points_Sum	NaN
38	62988	NaN	NaN	NaN	6638.74	12601.8	0	2860.5	728282	0	第2年观测窗口总累计积分	L1Y_Points_Sum	NaN
39	62988	NaN	NaN	NaN	0.486419	0.319105	0	0.5	1	0	第2年的乘机次数比率	Ration_L1Y_Flight_Count	NaN
40	62988	NaN	NaN	NaN	0.513581	0.319105	0	0.5	1	0	第1年的乘机次数比率	Ration_P1Y_Flight_Count	NaN
41	62988	NaN	NaN	NaN	0.522293	0.339632	0	0.514252	0.999989	0	第1年里程积分占最近两年积分比例	Ration_P1Y_BPS	NaN
42	62988	NaN	NaN	NaN	0.468422	0.338956	0	0.476747	0.999993	0	第2年里程积分占最近两年积分比例	Ration_L1Y_BPS	NaN
43	62988	NaN	NaN	NaN	2.72815	7.36416	0	0	140	0	非乘机的积分变动次数	Point_NotFlight	NaN

二、数据预处理

1、数据清洗

丢弃票价为空的记录

丢弃票价为0、平均折扣率不为0、总飞行公里数大于0的记录

# （1）保留票价非空的数据 即'SUM_YR_1'不为空并且'SUM_YR_2'不为空
data_1 = data[(data['SUM_YR_1'].notnull())&(data['SUM_YR_2'].notnull())]

len(data_1) #剩下62299条数据

62299

# （2）保留票价非零的，或者，平均折扣率与总飞行公里数同时为0的
# 即'SUM_YR_1'不为0，或者'SUM_YR_2'不为0，或者平均折扣率与总飞行公里数同时为0 （即'SUM_YR_1'、'SUM_YR_2'都为0，同时平均折扣率与总飞行公里数同时为0，也保留）
index1 = data_1['SUM_YR_1']!=0
index2 = data_1['SUM_YR_2']!=0
index3 = (data_1['SEG_KM_SUM']==0)&(data_1['avg_discount']==0)
data_2 = data_1[(index1) | (index2) | (index3)]   #或 的关系

len(data_2) #剩下62044条数据

62044

2、属性指标选取

根据航空公司客户价值LRFMC模型，选取与LRFMC指标相关的6个数据：

L客户关系长度：LOAD_TIME - FFP_DATE 其中LOAD_TIME为观测窗口的结束时间，FFP_DATE为入会时间（办理会员卡的开始的时间）

R消费时间间隔：LAST_TO_END 最后一次乘机时间至观察窗口末端时长

F客户在观测窗口内乘坐公司飞机的次数：FLIGHT_COUNT 飞行次数 单位：次

M客户在观测窗口内乘坐公司飞机的次数：SEG_KM_SUM 观测窗口总飞行公里数 单位：公里

C客户在观测窗口内乘坐舱位所对应的的折扣系数的平均值：avg_discount 平均折扣率

data_3 = data_2[['LOAD_TIME','FFP_DATE','LAST_TO_END','FLIGHT_COUNT','SEG_KM_SUM','avg_discount']].copy()
data_3.head()
#数据中Load_time和FFP_Date是字符串格式

	LOAD_TIME	FFP_DATE	LAST_TO_END	FLIGHT_COUNT	SEG_KM_SUM	avg_discount
0	2014/03/31	2006/11/02	1	210	580717	0.961639
1	2014/03/31	2007/02/19	7	140	293678	1.252314
2	2014/03/31	2007/02/01	11	135	283712	1.254676
3	2014/03/31	2008/08/22	97	23	281336	1.090870
4	2014/03/31	2009/04/10	5	152	309928	0.970658

data_3['LOAD_TIME'] = pd.to_datetime(data_3['LOAD_TIME'])
data_3['FFP_DATE'] = pd.to_datetime(data_3['FFP_DATE'])
data_3['L'] = data_3['LOAD_TIME'] - data_3['FFP_DATE']
data_3['R'] = data_3['LAST_TO_END']
data_3['F'] = data_3['FLIGHT_COUNT']
data_3['M'] = data_3['SEG_KM_SUM']
data_3['C'] = data_3['avg_discount']
data_3.head()

	LOAD_TIME	FFP_DATE	LAST_TO_END	FLIGHT_COUNT	SEG_KM_SUM	avg_discount	L	R	F	M	C
0	2014-03-31	2006-11-02	1	210	580717	0.961639	2706 days	1	210	580717	0.961639
1	2014-03-31	2007-02-19	7	140	293678	1.252314	2597 days	7	140	293678	1.252314
2	2014-03-31	2007-02-01	11	135	283712	1.254676	2615 days	11	135	283712	1.254676
3	2014-03-31	2008-08-22	97	23	281336	1.090870	2047 days	97	23	281336	1.090870
4	2014-03-31	2009-04-10	5	152	309928	0.970658	1816 days	5	152	309928	0.970658

为消除量纲影响，为后续聚类做准备，进行标准化

def data_Znorm(df, *cols):
    df_n = df.copy()
    for col in cols:
        u = df_n[col].mean()
        std = df_n[col].std()
        df_n['Z'+col] = (df_n[col] - u) / std
    return(df_n)

data_Z = data_Znorm(data_3, 'L','R','F','M','C')
data_Z.head()

	LOAD_TIME	FFP_DATE	LAST_TO_END	FLIGHT_COUNT	SEG_KM_SUM	avg_discount	L	R	F	M	C	ZL	ZR	ZF	ZM	ZC
0	2014-03-31	2006-11-02	1	210	580717	0.961639	2706 days	1	210	580717	0.961639	1.435707	-0.944948	14.034016	26.761154	1.295540
1	2014-03-31	2007-02-19	7	140	293678	1.252314	2597 days	7	140	293678	1.252314	1.307152	-0.911894	9.073213	13.126864	2.868176
2	2014-03-31	2007-02-01	11	135	283712	1.254676	2615 days	11	135	283712	1.254676	1.328381	-0.889859	8.718869	12.653481	2.880950
3	2014-03-31	2008-08-22	97	23	281336	1.090870	2047 days	97	23	281336	1.090870	0.658476	-0.416098	0.781585	12.540622	1.994714
4	2014-03-31	2009-04-10	5	152	309928	0.970658	1816 days	5	152	309928	0.970658	0.386032	-0.922912	9.923636	13.898736	1.344335

三、模型构建

1、客户聚类

采用K-Means聚类算法对客户数据尽心给客户分群，聚成5类（5类是按照经验判断出来的）

K-Means需要事先确定聚成几类，层次聚类不用，参考https://blog.csdn.net/LuohenYJ/article/details/79981269

函数参考https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html#sklearn.cluster.KMeans

data_K = data_Z[['ZL','ZR','ZF','ZM','ZC']]

from sklearn.cluster import KMeans

kmodel = KMeans(n_clusters=5)
kmodel.fit(data_K)

KMeans(algorithm='auto', copy_x=True, init='k-means++', max_iter=300,
    n_clusters=5, n_init=10, n_jobs=1, precompute_distances='auto',
    random_state=None, tol=0.0001, verbose=0)

print(kmodel.cluster_centers_) #查看聚类中心
print(kmodel.labels_) 

[[-0.31367829  1.68625847 -0.57401599 -0.53682019 -0.1733261 ]
 [-0.70020646 -0.41488827 -0.16114258 -0.16095751 -0.25513154]
 [ 0.48332845 -0.79938326  2.4832016   2.42472391  0.30863003]
 [ 1.16066672 -0.37722119 -0.08691852 -0.09484404 -0.1559046 ]
 [ 0.05184279 -0.00266813 -0.22680311 -0.23125407  2.19134701]]
[2 2 2 ... 1 0 0]

#生成聚类结果表，参考https://www.cnblogs.com/itdyb/p/5691958.html
r1=pd.Series(kmodel.labels_).value_counts()  #得到每一类的聚类个数
r2=pd.DataFrame(kmodel.cluster_centers_) #得到聚类中心
print(r1)
print(r2)

1    24659
3    15740
0    12125
2     5336
4     4184
dtype: int64
          0         1         2         3         4
0 -0.313678  1.686258 -0.574016 -0.536820 -0.173326
1 -0.700206 -0.414888 -0.161143 -0.160958 -0.255132
2  0.483328 -0.799383  2.483202  2.424724  0.308630
3  1.160667 -0.377221 -0.086919 -0.094844 -0.155905
4  0.051843 -0.002668 -0.226803 -0.231254  2.191347

r=pd.concat([r1,r2],axis=1)
r.columns=['聚类个数'] + list(data_K.columns)
print(r)
# r.to_excel(resoutfile,index=False)

    聚类个数        ZL        ZR        ZF        ZM        ZC
0  12125 -0.313678  1.686258 -0.574016 -0.536820 -0.173326
1  24659 -0.700206 -0.414888 -0.161143 -0.160958 -0.255132
2   5336  0.483328 -0.799383  2.483202  2.424724  0.308630
3  15740  1.160667 -0.377221 -0.086919 -0.094844 -0.155905
4   4184  0.051843 -0.002668 -0.226803 -0.231254  2.191347

#画雷达图
import matplotlib.pyplot as plt
labels = np.array(['ZL','ZR','ZF','ZM','ZC'])

angles = np.linspace(0, 2*np.pi, 5, endpoint=False) # 分割圆周长
angles = np.concatenate((angles, [angles[0]])) # 闭合   

for i in range(4):
    data_i = r2.iloc[i]
    data_i.rename('i')
    data_i = np.concatenate((data_i, [data_0[i]])) # 闭合
    plt.polar(angles, data_i, 'o-', linewidth=1) #做极坐标系
    plt.fill(angles, data_i, alpha=0.25)# 填充

plt.thetagrids(angles * 180/np.pi, labels) # 设置网格、标签

(,
 )