本章节的目的是【明确目标用户群】，以更好的服务现有用户。

【知识点】

1.作图

显示中文

plt.rcParams['font.sans-serif'] = ['SimHei'] # 步骤一（替换sans-serif字体） plt.rcParams['axes.unicode_minus'] = False # 步骤二（解决坐标轴负数的负号显示问题）

2.数据库操作

sqlalchemy 引擎

engine = create_engine('mysql+pymysql://root:123456@localhost:3306/datascience')

3.批量读取文件

os.wolk()、os.path.join()用法

for root, dirs, files in os.walk(path): for file in files:`rfile = os.path.join(root,file)if rfile.split('.')[-1] == 'tsv':rdf = pd.read_csv(rfile, sep='\t')df = df.append(rdf)

4.groupby()以及agg() 的联合使用，应对不同列使用不同的函数

按月统计

affc = {'payment':'sum', 'log_date':'count'} dfm = df.groupby(['log_month', 'user_id']).agg(affc).reset_index()

修改列明

renam = {'log_date':'access_days'} dfm.rename(columns=renam, inplace=True)

5.KMeans 聚类的使用

单列的聚类（需要将单列应用 reshape（-1,1）格式化为1列）

from sklearn.cluster import KMeans a47 = action['A47'].reshape(-1, 1) kms = KMeans(n_clusters=3).fit(a47)

聚类的标签 labels_ 属性

cluster = kms.labels_

将标签添加至源数据中，运用groupby（）查看分组情况

action['cluster'] = cluster action.groupby(['cluster'])['user_id'].count()

可视化分组

snsdf = action[['user_id','A47','cluster']].sort_values(by='A47',ascending=False)
plt.figure(figsize=(8,5))
snsdf1 = snsdf.reset_index()
snsdf1[snsdf1['cluster']==2]['A47'].plot(color='r',label='2:重度用户')
snsdf1[snsdf1['cluster']==1]['A47'].plot(color='g',label='1:中度用户')
snsdf1[snsdf1['cluster']==0]['A47'].plot(color='b',label='0:轻度用户')
plt.legend()
plt.xlabel('用户分布')
plt.ylabel('排行榜得分')

6.主成分分析

数据预处理
- 提取要进行主成分分析的列
  paction = acc.iloc[:,3:(len(acc.columns)-1)]
- 删掉0值较多的列
  cc = paction[paction==0].count(axis=0)/len(paction) cc.plot() dd = cc[cc<.9] #删掉该列中90%以上都是0值的列 paction = paction[dd.index] paction.head()

删掉相关性较强的列

# 数据概览
corp = paction.corr()
sns.heatmap(corp)
mask = np.array(corp)
mask[np.tril_indices_from(mask)] = False        # 画下三角heatmap的方法
sns.heatmap(corp,mask=mask)# 通过下三角矩阵的方式，删掉相关性较强的数据列
coll = corp.columns
corp = pd.DataFrame(np.tril(corp, -1))         # 应用 np.tril(m, -1) 函数获取下三角，上三角数据全部置为0
corp.columns = coll
pac2 = paction.loc[:,(corp.abs()<.8).all()]      # 任何一个数都小于 0.8 的数据 all() 函数
pac2.head()

进行主成分分析

from sklearn.decomposition import PCA
pca = PCA()
pca.fit(pac2)redio = pca.explained_variance_ratio_          # pca.explained_variance_ratio_ 是PCA降维后的矩阵课解释性比率
print(redio) 
print(pca.singular_values_)                # singular_values_ 是奇异值矩阵

主成分的课解释性曲线

recu = redio.cumsum()                     # 应用 cumsum() 函数进行逐数据累加
plt.plot(recu)

获取降维后的数据以进行下一步

pca.set_params(n_components=10)              # 设置 维度 为 10 
pac3 = pd.DataFrame(pca.fit_transform(pac2))     # 使用fit_transform()函数训练并获得降维后的数据
pac3.head()

- 继续应用 KMENAS 进行聚类，得到所有用户的分类，然后再平均每个分类的每个行为的所有用户的值
- 继续应用相关性删除相关性强的列，获得最后主要观察指标

对主要观察指标进行雷达图展示

# 首先，对数据进行标准化处理
from sklearn.preprocessing import scale
ccccc = pd.DataFrame(scale(cccc))
ccccc.columns = cccc.columns# 画图
plt.figure(figsize=(8,8))                  
N = ccccc.shape[1]                      # 极坐标的分割分数
angles = np.linspace(0, 2*np.pi, N, endpoint=False)    # 设置雷达图的角度，用于平分切开一个圆面      
angles = np.concatenate((angles,[angles[0]]))   # 使雷达图一圈封闭起来
for i in range(len(ccccc)):values = ccccc.loc[i,:]              # 构造数据values = np.concatenate((values,[values[0]]))     # 为了使雷达图一圈封闭起来plt.polar(angles, values, 'o-', linewidth=2)      # 绘制
plt.legend(ccccc.index, loc='lower right')
plt.thetagrids(angles * 180/np.pi, labels=list(ccccc.columns))    # 添加极坐标的标签
plt.title('重要指标雷达图呈现')

一、库导入以及matplotlib显示中文

import pandas as pd
import numpy as np
import pymysql
from sqlalchemy import create_engine
import matplotlib.pyplot as plt
import seaborn as sns
import missingno as msno
import osplt.rcParams['font.sans-serif'] = ['SimHei'] # 步骤一（替换sans-serif字体）
plt.rcParams['axes.unicode_minus'] = False   # 步骤二（解决坐标轴负数的负号显示问题）
%matplotlib inline

数据库引擎

engine = create_engine('mysql+pymysql://root:123456@localhost:3306/datascience')

二、批量读取文件

def read_files(path):df = pd.DataFrame()for root, dirs, files in os.walk(path):for file in files:rfile = os.path.join(root,file)if rfile.split('.')[-1] == 'tsv':rdf = pd.read_csv(rfile, sep='\t')df = df.append(rdf)return df

action_path  = 'data/sample-data/section8/daily/action/'
dau_path = 'data/sample-data/section8/daily/dau/'
dpu_path = 'data/sample-data/section8/daily/dpu/'action = read_files(action_path)
dau = read_files(dau_path)
dpu = read_files(dpu_path)

查看数据完整性以及头部信息

print(action.isnull().sum().sum())
print(action.shape)
# print(action.info())
action.head()

0
(2653, 57)

	log_date	app_name	user_id	A2	A5	A6	A7	...	A45	A46	A47	A48	A49	A50	A51	A54
0	2013-10-31	game-01	654133	0	0	0	0	...	0	0	380	25655	0	0	0	46
1	2013-10-31	game-01	425530	0	10	1	233	...	19	20	180543	347	36	22	4	71
2	2013-10-31	game-01	709596	0	0	0	0	...	0	0	416	24817	0	0	0	2
3	2013-10-31	game-01	525047	2	9	0	0	...	22	22	35200	6412	21	0	0	109
4	2013-10-31	game-01	796908	0	0	0	0	...	29	29	388	25444	1	0	0	64

5 rows × 57 columns

print(dau.isnull().sum().sum())
print(dau.shape)
print(dau.info())
dau.head()

0
(509754, 3)
<class 'pandas.core.frame.DataFrame'>
Int64Index: 509754 entries, 0 to 2410
Data columns (total 3 columns):
log_date    509754 non-null object
app_name    509754 non-null object
user_id     509754 non-null int64
dtypes: int64(1), object(2)
memory usage: 15.6+ MB
None

	log_date	app_name	user_id
0	2013-05-01	game-01	608801
1	2013-05-01	game-01	712453
2	2013-05-01	game-01	776853
3	2013-05-01	game-01	823486
4	2013-05-01	game-01	113600

print(dpu.isnull().sum().sum())
print(dpu.shape)
print(dpu.info())
dpu.head()

0
(3532, 4)
<class 'pandas.core.frame.DataFrame'>
Int64Index: 3532 entries, 0 to 7
Data columns (total 4 columns):
log_date    3532 non-null object
app_name    3532 non-null object
user_id     3532 non-null int64
payment     3532 non-null int64
dtypes: int64(2), object(2)
memory usage: 138.0+ KB
None

	log_date	app_name	user_id	payment
0	2013-05-01	game-01	804005	571
1	2013-05-01	game-01	793537	81
2	2013-05-01	game-01	317717	81
3	2013-05-01	game-01	317717	81
4	2013-05-01	game-01	426525	324

# 写入数据库# action.to_sql('s8_action', engine, index=False)
# dau.to_sql('s8_dau', engine, index=False)
# dpu.to_sql('s8_dpu', engine, index=False)

三、数据预处理

1.合并 DAU DPU

df = pd.merge(dau, dpu[['log_date','user_id','payment']], how='left', on=['user_id','log_date'])
df.head()

	log_date	app_name	user_id	payment
0	2013-05-01	game-01	608801	NaN
1	2013-05-01	game-01	712453	NaN
2	2013-05-01	game-01	776853	NaN
3	2013-05-01	game-01	823486	NaN
4	2013-05-01	game-01	113600	NaN

# 将无消费记录的消费额设为 0 
print(df.payment.isnull().sum())
df['payment'].fillna(0, inplace=True)
print(df.payment.isnull().sum())

507151
0

# 添加消费额标志位
df['is_pay'] = df['payment'].apply( lambda x: 1 if x>0 else 0 )
df.head()

	log_date	app_name	user_id
0	2013-05-01	game-01	608801
1	2013-05-01	game-01	712453
2	2013-05-01	game-01	776853
3	2013-05-01	game-01	823486
4	2013-05-01	game-01	113600

2.按月统计

# 增加月份列
df['log_month'] = df['log_date'].apply(lambda x: x[0:7])
df.head()

	log_date	app_name	user_id	log_month
0	2013-05-01	game-01	608801	2013-05
1	2013-05-01	game-01	712453	2013-05
2	2013-05-01	game-01	776853	2013-05
3	2013-05-01	game-01	823486	2013-05
4	2013-05-01	game-01	113600	2013-05

巧妙运用 groupby 以及 agg 函数，统计出用户按月份的消费情况和登陆次数

# 按月统计
affc = {'payment':'sum', 'log_date':'count'}
dfm = df.groupby(['log_month', 'user_id']).agg(affc).reset_index()
# 修改列明
renam = {'log_date':'access_days'}
dfm.rename(columns=renam, inplace=True)
dfm.head()

	log_month	user_id	access_days
0	2013-05	65	1
1	2013-05	115	1
2	2013-05	194	1
3	2013-05	426	4
4	2013-05	539	1

4.使用 Kmeans 进行分类，得到排名靠前的用户，即重度用户/中度用户/轻度用户

A47 列即是排行榜得分，从分布图上看出，大部分用户得分很低，符合幂律曲线

# 
action['A47'].hist(bins=50, figsize=(6,4))

<matplotlib.axes._subplots.AxesSubplot at 0x1c21d894240>

png

sns.distplot(action['A47'],bins=50,kde=True)

<matplotlib.axes._subplots.AxesSubplot at 0x1c21af07a58>

png

对 A47 列进行聚类，分为3类

from sklearn.cluster import KMeansa47 = action['A47'].reshape(-1, 1)kms = KMeans(n_clusters=3).fit(a47)

D:\ProgramData\Anaconda3\lib\site-packages\ipykernel_launcher.py:3: FutureWarning: reshape is deprecated and will raise in a subsequent release. Please use .values.reshape(...) insteadThis is separate from the ipykernel package so we can avoid doing imports until

cluster = kms.labels_
kms.cluster_centers_

array([[  9359.84787792],[ 69386.11297071],[185857.17948718]])

action['cluster'] = cluster
action.head()

	log_date	app_name	user_id	A2	A5	A6	A7	...	A46	A47	A48	A49	A50	A51	A54	cluster
0	2013-10-31	game-01	654133	0	0	0	0	...	0	380	25655	0	0	0	46	0
1	2013-10-31	game-01	425530	0	10	1	233	...	20	180543	347	36	22	4	71	2
2	2013-10-31	game-01	709596	0	0	0	0	...	0	416	24817	0	0	0	2	0
3	2013-10-31	game-01	525047	2	9	0	0	...	22	35200	6412	21	0	0	109	0
4	2013-10-31	game-01	796908	0	0	0	0	...	29	388	25444	1	0	0	64	0

5 rows × 58 columns

action.groupby(['cluster'])['user_id'].count()

cluster
0    2096
1     479
2      78
Name: user_id, dtype: int64

图上显示，通过聚类分解后用户分为3个类， 0 表示轻度用户，排行榜得分最少； 1 表示中度用户，排行版得分居中； 2 表示重度用户，排行版得分较高，而且用户数量较少，符合实际情况。

snsdf = action[['user_id','A47','cluster']].sort_values(by='A47',ascending=False)
snsdf['user'] = range(len(snsdf))
sns.scatterplot(x='user',y='A47',hue='cluster',data=snsdf, palette='rainbow', alpha=.2)

<matplotlib.axes._subplots.AxesSubplot at 0x1c21b9bf898>

png

snsdf = action[['user_id','A47','cluster']].sort_values(by='A47',ascending=False)
snsdf['user'] = range(len(snsdf))plt.figure(figsize=(8,5))
snsdf1 = snsdf.reset_index()
snsdf1[snsdf1['cluster']==2]['A47'].plot(color='r',label='2:重度用户')
snsdf1[snsdf1['cluster']==1]['A47'].plot(color='g',label='1:中度用户')
snsdf1[snsdf1['cluster']==0]['A47'].plot(color='b',label='0:轻度用户')
plt.legend()
plt.xlabel('用户分布')
plt.ylabel('排行榜得分')

Text(0,0.5,'排行榜得分')

png

限定排名靠前的用户，即得分较高的重度和中度用户，以便接下来进行分析

acc = action[action['cluster']>=1]
acc.head()

	log_date	app_name	user_id	A5	A6	A7	...	A46	A47	A48	A49	A50	A51	A54	cluster
1	2013-10-31	game-01	425530	10	1	233	...	20	180543	347	36	22	4	71	2
5	2013-10-31	game-01	776120	9	0	0	...	38	142214	684	37	15	0	312	2
7	2013-10-31	game-01	276197	7	0	58	...	15	54602	4226	15	0	8	95	1
8	2013-10-31	game-01	221572	1	0	0	...	24	39891	5792	4	0	0	21	1
9	2013-10-31	game-01	692433	6	0	0	...	28	50706	4549	16	8	0	154	1

5 rows × 58 columns

5.主成分分析

获取关键的参数

paction = acc.iloc[:,3:(len(acc.columns)-1)]
paction.index=acc.user_id
paction.head()

	A1	A2	A3	A4	A5	A6	A7	A8	A9	A10	...	A45	A46	A47	A48	A49	A50	A51	A52	A53	A54
user_id
425530	0	0	0	0	10	1	233	58.25	288	230	...	19	20	180543	347	36	22	4	0	0.0	71
776120	0	0	0	0	9	0	0	0.00	325	195	...	19	38	142214	684	37	15	0	0	0.0	312
276197	0	0	0	0	7	0	58	7.25	150	100	...	15	15	54602	4226	15	0	8	0	0.0	95
221572	0	0	0	0	1	0	0	0.00	40	14	...	24	24	39891	5792	4	0	0	0	0.0	21
692433	0	0	0	0	6	0	0	0.00	102	95	...	15	28	50706	4549	16	8	0	0	0.0	154

5 rows × 54 columns

1.删掉 0 值比较多的列

cc = paction[paction==0].count(axis=0)/len(paction)
print(cc.head())
cc.plot()

A1    1.000000
A2    0.926391
A3    1.000000
A4    0.994614
A5    0.055655
dtype: float64<matplotlib.axes._subplots.AxesSubplot at 0x1c21bbb1470>

png

# cc[cc>.8]

dd = cc[cc<.95]

paction = paction[dd.index]
paction.head()

	A2	A5	A6	A7	A8	A9	A10	A11	A12	A13	...	A45	A46	A47	A48	A49	A50	A51	A52	A53	A54
user_id
425530	0	10	1	233	58.25	288	230	19	2	19	...	19	20	180543	347	36	22	4	0	0.0	71
776120	0	9	0	0	0.00	325	195	38	8	19	...	19	38	142214	684	37	15	0	0	0.0	312
276197	0	7	0	58	7.25	150	100	15	3	11	...	15	15	54602	4226	15	0	8	0	0.0	95
221572	0	1	0	0	0.00	40	14	0	0	3	...	24	24	39891	5792	4	0	0	0	0.0	21
692433	0	6	0	0	0.00	102	95	0	0	2	...	15	28	50706	4549	16	8	0	0	0.0	154

5 rows × 32 columns

2.删掉相关性较强的列

corp = paction.corr()
plt.figure(figsize=(15,8))
sns.heatmap(corp)

<matplotlib.axes._subplots.AxesSubplot at 0x1c21bc094a8>

png

画下三角heatmap，使用到的函数

mask = np.array(corp)
mask[np.tril_indices_from(mask)] = False
fig,ax = plt.subplots()
fig.set_size_inches(15,8)
sns.heatmap(corp,mask=mask)

<matplotlib.axes._subplots.AxesSubplot at 0x1c21bc09400>

png

获取矩阵的下三角，如果要获取上三角的话， np.tril(m, 1)

coll = corp.columns
corp = pd.DataFrame(np.tril(corp, -1))
corp.columns = coll
corp.head()

	A2	A5	A6	A7	...
0	0.000000	0.000000	0.000000	0.000000	...
1	0.069744	0.000000	0.000000	0.000000	...
2	0.076185	0.178833	0.000000	0.000000	...
3	0.158735	0.219395	0.371360	0.000000	...
4	0.167200	0.186124	0.242025	0.803161	...

5 rows × 32 columns

pac2 = paction.loc[:,(corp.abs()<.7).all()]      # 任何一个数都小于0.7 的数据
pac2.head()

	A2	A11	A12	A13	A20	A23	A24	A43	A44	A46	A48	A49	A50	A51	A53	A54
user_id
425530	0	19	2	19	0	0	0.5	23	0.92174	20	347	36	22	4	0.0	71
776120	0	38	8	19	0	0	0.0	20	0.90256	38	684	37	15	0	0.0	312
276197	0	15	3	11	0	0	0.0	10	0.92000	15	4226	15	0	8	0.0	95
221572	0	0	0	3	0	0	0.0	2	0.85714	24	5792	4	0	0	0.0	21
692433	0	0	0	2	0	0	0.0	11	0.73684	28	4549	16	8	0	0.0	154

进行主成分分析

from sklearn.decomposition import PCA

pca = PCA()
pca.fit(pac2)

PCA(copy=True, iterated_power='auto', n_components=None, random_state=None,svd_solver='auto', tol=0.0, whiten=False)

redio = pca.explained_variance_ratio_
print(redio) 
print(pca.singular_values_)

[9.97843804e-01 1.92024564e-03 1.20120771e-04 5.57014208e-052.67905481e-05 1.54533752e-05 9.31262940e-06 4.38846214e-063.02317261e-06 8.36725295e-07 1.31874979e-07 9.78197162e-083.86464536e-08 2.94647596e-08 1.82272465e-08 7.54580333e-09]
[3.96183910e+04 1.73797668e+03 4.34684952e+02 2.96004755e+022.05284590e+02 1.55911168e+02 1.21032418e+02 8.30848288e+016.89599635e+01 3.62791414e+01 1.44027941e+01 1.24044853e+017.79687146e+00 6.80796010e+00 5.35458829e+00 3.44523057e+00]

recu = redio.cumsum()
print(recu)
x = np.arange(len(recu))
plt.plot(recu, color='r')

[0.9978438  0.99976405 0.99988417 0.99993987 0.99996666 0.999982120.99999143 0.99999582 0.99999884 0.99999968 0.99999981 0.999999910.99999994 0.99999997 0.99999999 1.        ][<matplotlib.lines.Line2D at 0x1c21dadada0>]

png

得到降维后的数据

pca.set_params(n_components=10)
pac3 = pd.DataFrame(pca.fit_transform(pac2))
pacsse = pac3.copy()
pac3.head()

	0	1	2	3	4	5	6	7	8	9
0	2706.266005	-100.824346	-1.874787	-1.577536	12.481591	-2.394320	9.770878	7.807535	0.021273	-2.169596
1	2373.811140	147.314930	-16.386795	-8.428655	10.019577	-3.004725	6.009771	0.961469	-1.598531	2.144615
2	-1171.733361	-5.493081	0.744995	0.542033	-0.785251	-5.756412	-1.012336	-1.778067	7.256884	0.343277
3	-2738.903900	-50.468487	2.328491	2.965415	-5.794347	11.891289	2.965366	-1.182413	0.065619	1.245358
4	-1493.642618	58.686385	-10.807612	11.777973	7.664692	9.312968	4.376429	1.994214	-1.568050	0.426246

6.KMeans 进行聚类

from sklearn.cluster import KMeanskm = KMeans(n_clusters=5)
km.fit(pac3)

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)

clu = km.labels_
pac3['clu'] = clu
pac3.head()

	0	1	2	3	4	5	6	7	8	9	clu
0	2706.266005	-100.824346	-1.874787	-1.577536	12.481591	-2.394320	9.770878	7.807535	0.021273	-2.169596	0
1	2373.811140	147.314930	-16.386795	-8.428655	10.019577	-3.004725	6.009771	0.961469	-1.598531	2.144615	0
2	-1171.733361	-5.493081	0.744995	0.542033	-0.785251	-5.756412	-1.012336	-1.778067	7.256884	0.343277	1
3	-2738.903900	-50.468487	2.328491	2.965415	-5.794347	11.891289	2.965366	-1.182413	0.065619	1.245358	4
4	-1493.642618	58.686385	-10.807612	11.777973	7.664692	9.312968	4.376429	1.994214	-1.568050	0.426246	1

pac3.groupby('clu')[2].count()

clu
0     90
1    113
2    122
3    109
4    123
Name: 2, dtype: int64

#### palette 的颜色风格：
Accent, Accent_r, Blues, Blues_r, BrBG, BrBG_r, BuGn, BuGn_r, BuPu, BuPu_r, CMRmap, CMRmap_r, Dark2, Dark2_r, GnBu, GnBu_r, Greens, Greens_r, Greys, Greys_r, OrRd, OrRd_r, Oranges, Oranges_r, PRGn, PRGn_r, Paired, Paired_r, Pastel1, Pastel1_r, Pastel2, Pastel2_r, PiYG, PiYG_r, PuBu, PuBuGn, PuBuGn_r, PuBu_r, PuOr, PuOr_r, PuRd, PuRd_r, Purples, Purples_r, RdBu, RdBu_r, RdGy, RdGy_r, RdPu, RdPu_r, RdYlBu, RdYlBu_r, RdYlGn, RdYlGn_r, Reds, Reds_r, Set1, Set1_r, Set2, Set2_r, Set3, Set3_r, Spectral, Spectral_r, Vega10, Vega10_r, Vega20, Vega20_r, Vega20b, Vega20b_r, Vega20c, Vega20c_r, Wistia, Wistia_r, YlGn, YlGnBu, YlGnBu_r, YlGn_r, YlOrBr, YlOrBr_r, YlOrRd, YlOrRd_r, afmhot, afmhot_r, autumn, autumn_r, binary, binary_r, bone, bone_r, brg, brg_r, bwr, bwr_r, cool, cool_r, coolwarm, coolwarm_r, copper, copper_r, cubehelix, cubehelix_r, flag, flag_r, gist_earth, gist_earth_r, gist_gray, gist_gray_r, gist_heat, gist_heat_r, gist_ncar, gist_ncar_r, gist_rainbow, gist_rainbow_r, gist_stern, gist_stern_r, gist_yarg, gist_yarg_r, gnuplot, gnuplot2, gnuplot2_r, gnuplot_r, gray, gray_r, hot, hot_r, hsv, hsv_r, icefire, icefire_r, inferno, inferno_r, jet, jet_r, magma, magma_r, mako, mako_r, nipy_spectral, nipy_spectral_r, ocean, ocean_r, pink, pink_r, plasma, plasma_r, prism, prism_r, rainbow, rainbow_r, rocket, rocket_r, seismic, seismic_r, spectral, spectral_r, spring, spring_r, summer, summer_r, tab10, tab10_r, tab20, tab20_r, tab20b, tab20b_r, tab20c, tab20c_r, terrain, terrain_r, viridis, viridis_r, vlag, vlag_r, winter, winter_r

plt.figure(figsize=(13,7))
sns.scatterplot(x=0, y=1, data=pac3,style='clu',hue='clu', palette='autumn')

<matplotlib.axes._subplots.AxesSubplot at 0x1c21db35438>

png

将分类后的类别添加至原数据中

pac4 = pac2.copy()
pac4['cluster'] = list(pac3.clu)
pac4.head()

	A2	A11	A12	A13	A20	A23	A24	A43	A44	A46	A48	A49	A50	A51	A53	A54	cluster
user_id
425530	0	19	2	19	0	0	0.5	23	0.92174	20	347	36	22	4	0.0	71	0
776120	0	38	8	19	0	0	0.0	20	0.90256	38	684	37	15	0	0.0	312	0
276197	0	15	3	11	0	0	0.0	10	0.92000	15	4226	15	0	8	0.0	95	1
221572	0	0	0	3	0	0	0.0	2	0.85714	24	5792	4	0	0	0.0	21	4
692433	0	0	0	2	0	0	0.0	11	0.73684	28	4549	16	8	0	0.0	154	1

# 计算每个类的平均值
clu5 = pac4.groupby('cluster').mean()

# 删除相关性较高的列
clu5.drop(columns='A53',inplace=True)
c5cor = clu5.corr()
plt.figure(figsize=(15,8))
sns.heatmap(c5cor,annot=True)

<matplotlib.axes._subplots.AxesSubplot at 0x1c21d92a780>

png

ccrp = pd.DataFrame(np.tril(c5cor,-1))
ccrp.columns = clu5.columns
cccc = clu5.loc[:,(ccrp.abs()<.95).all()]
cccc

	A2	A20	A23	A24	A44	A46	A50	A51	A54
cluster
0	0.022222	0.322222	0.655556	0.167691	0.858193	27.600000	10.666667	2.011111	166.711111
1	0.079646	0.274336	0.362832	0.095231	0.844027	20.159292	3.008850	1.469027	102.106195
2	0.073770	0.377049	0.336066	0.070628	0.849343	24.737705	4.286885	1.844262	121.909836
3	0.018349	0.229358	0.284404	0.098252	0.845981	24.119266	5.266055	1.733945	146.871560
4	0.203252	0.292683	0.243902	0.063686	0.775076	18.983740	2.130081	0.975610	84.032520

from sklearn.preprocessing import scaleccccc = pd.DataFrame(scale(cccc))

ccccc.columns = cccc.columns
ccccc

	A2	A20	A23	A24	A44	A46	A50	A51	A54
0	-0.855590	0.468859	1.918400	1.862020	0.785882	1.422970	1.867773	1.118457	1.424282
1	0.002962	-0.503392	-0.094337	-0.104961	0.315530	-0.940402	-0.688647	-0.381093	-0.746672
2	-0.084884	1.582038	-0.278379	-0.772826	0.492038	0.513827	-0.261998	0.656909	-0.081200
3	-0.913505	-1.416613	-0.633601	-0.022944	0.380387	0.317394	0.064879	0.351742	0.757602
4	1.851016	-0.130892	-0.912083	-0.961289	-1.973837	-1.313789	-0.982007	-1.746015	-1.354012

plt.figure(figsize=(8,8))
# 极坐标的分割分数
N = ccccc.shape[1]
# 设置雷达图的角度，用于平分切开一个圆面
angles = np.linspace(0, 2*np.pi, N, endpoint=False)
# 使雷达图一圈封闭起来
angles = np.concatenate((angles,[angles[0]]))
for i in range(len(ccccc)):# 构造数据values = ccccc.loc[i,:]# 为了使雷达图一圈封闭起来values = np.concatenate((values,[values[0]]))# 绘制plt.polar(angles, values, 'o-', linewidth=2)
plt.legend(ccccc.index, loc='lower right')
# 添加极坐标的标签
plt.thetagrids(angles * 180/np.pi, labels=list(ccccc.columns))
plt.title('重要指标雷达图呈现')

Text(0.5,1.05,'重要指标雷达图呈现')

png

不进行预处理的降维

dfp = acc.iloc[:,3:(len(acc.columns)-1)]
dfp.index=acc.user_id
dfp.head()

	A1	A2	A3	A4	A5	A6	A7	A8	A9	A10	...	A45	A46	A47	A48	A49	A50	A51	A52	A53	A54
user_id
425530	0	0	0	0	10	1	233	58.25	288	230	...	19	20	180543	347	36	22	4	0	0.0	71
776120	0	0	0	0	9	0	0	0.00	325	195	...	19	38	142214	684	37	15	0	0	0.0	312
276197	0	0	0	0	7	0	58	7.25	150	100	...	15	15	54602	4226	15	0	8	0	0.0	95
221572	0	0	0	0	1	0	0	0.00	40	14	...	24	24	39891	5792	4	0	0	0	0.0	21
692433	0	0	0	0	6	0	0	0.00	102	95	...	15	28	50706	4549	16	8	0	0	0.0	154

5 rows × 54 columns

from sklearn.decomposition import PCApca = PCA(whiten=False)
pca.fit(dfp)

PCA(copy=True, iterated_power='auto', n_components=None, random_state=None,svd_solver='auto', tol=0.0, whiten=False)

retio = pca.explained_variance_ratio_
# print(retio) 
# print(pca.singular_values_)  rec = retio.cumsum()
print(rec)
x = np.arange(len(rec))
plt.plot(rec, color='r')

[0.9996008  0.99995245 0.99997489 0.99999016 0.9999933  0.999995640.99999759 0.99999838 0.99999897 0.9999995  0.99999962 0.999999720.99999979 0.99999986 0.9999999  0.99999993 0.99999996 0.999999970.99999997 0.99999998 0.99999998 0.99999999 0.99999999 0.999999990.99999999 1.         1.         1.         1.         1.1.         1.         1.         1.         1.         1.1.         1.         1.         1.         1.         1.1.         1.         1.         1.         1.         1.1.         1.         1.         1.         1.         1.        ][<matplotlib.lines.Line2D at 0x1c21f406780>]

png

pca.set_params(n_components=10)
pacsse = pd.DataFrame(pca.fit_transform(dfp))
pacsse.head()

	0	1	2	3	4	5	6	7	8	9
0	94938.293061	-342.891655	-161.442878	-199.616210	1.830692	73.107938	153.124982	124.440657	-34.371612	46.548951
1	56613.313155	-960.580156	-38.560364	-45.836571	13.670166	90.767620	-145.846645	-40.255134	10.508203	16.287863
2	-31060.195159	388.005529	-6.932692	-0.948812	-5.332728	18.237293	11.393467	14.689011	-7.994909	32.398532
3	-45806.252443	1579.357883	-81.812845	-96.488345	-18.477649	-90.059217	31.377291	-22.865193	-19.724837	16.293640
4	-34963.135693	611.858506	-18.187490	-16.454233	-5.597209	-9.722257	-63.112236	-3.943266	7.222725	-10.889839

手肘法获取最优 K 值

from sklearn.cluster import KMeansdf_features = pacsse # 读入数据
# '利用SSE选择k'
SSE = []  # 存放每次结果的误差平方和
for k in range(1,9):estimator = KMeans(n_clusters=k)  # 构造聚类器estimator.fit(df_features)SSE.append(estimator.inertia_)
X = range(1,9)
plt.xlabel('k')
plt.ylabel('SSE')
plt.plot(X,SSE,'o-')

[<matplotlib.lines.Line2D at 0x1c2211cac50>]

png

显然，先标准化数据是不合适的

# 显然，先标准化数据是不合适的df_features = pd.DataFrame(scale(pacsse)) SSE = []  
for k in range(1,9):estimator = KMeans(n_clusters=k) estimator.fit(df_features)SSE.append(estimator.inertia_)
X = range(1,9)
plt.xlabel('k')
plt.ylabel('SSE')
plt.plot(X,SSE,'o-')

[<matplotlib.lines.Line2D at 0x1c2213bc438>]

png

km = KMeans(n_clusters=4)
km.fit(pacsse)
clu = km.labels_
pacsse['clu'] = clu
pacsse.head()

	0	1	2	3	4	5	6	7	8	9	clu
0	94938.293061	-342.891655	-161.442878	-199.616210	1.830692	73.107938	153.124982	124.440657	-34.371612	46.548951	2
1	56613.313155	-960.580156	-38.560364	-45.836571	13.670166	90.767620	-145.846645	-40.255134	10.508203	16.287863	0
2	-31060.195159	388.005529	-6.932692	-0.948812	-5.332728	18.237293	11.393467	14.689011	-7.994909	32.398532	1
3	-45806.252443	1579.357883	-81.812845	-96.488345	-18.477649	-90.059217	31.377291	-22.865193	-19.724837	16.293640	1
4	-34963.135693	611.858506	-18.187490	-16.454233	-5.597209	-9.722257	-63.112236	-3.943266	7.222725	-10.889839	1

pacsse.groupby('clu')[2].count()

clu
0    153
1    344
2     54
3      6
Name: 2, dtype: int64

plt.figure(figsize=(13,7))
sns.scatterplot(x=0, y=1, data=pacsse,style='clu',hue='clu', palette='autumn')

<matplotlib.axes._subplots.AxesSubplot at 0x1c22118b668>

png

显然，不进行预处理的数据聚类是有问题的，第一主成分和第二主成分显然是相关的

pac4 = pac2.copy()
pac4['cluster'] = list(pacsse.clu)
pac4.head()clu5 = pac4.groupby('cluster').mean()
clu5.drop(columns='A53',inplace=True)
c5cor = clu5.corr()
plt.figure(figsize=(15,8))
sns.heatmap(c5cor,annot=True)

<matplotlib.axes._subplots.AxesSubplot at 0x1c22145a4e0>

png

ccrp = pd.DataFrame(np.tril(c5cor,-1))
ccrp.columns = clu5.columns
cccc = clu5.loc[:,(ccrp.abs()<.95).all()]
cccc

	A12	A20	A51	A54
cluster
0	3.398693	0.228758	1.810458	146.287582
1	1.938953	0.316860	1.433140	101.531977
2	4.592593	0.407407	1.870370	169.777778
3	2.166667	0.166667	1.666667	213.833333

from sklearn.preprocessing import scaleccccc = pd.DataFrame(scale(cccc))ccccc.columns = cccc.columns
ccccc

	A12	A20	A51	A54
0	0.352533	-0.562784	0.684599	-0.285229
1	-1.021705	0.406288	-1.555764	-1.388557
2	1.476502	1.402249	1.040338	0.293858
3	-0.807330	-1.245753	-0.169173	1.379928

plt.figure(figsize=(8,8))
# 极坐标的分割分数
N = ccccc.shape[1]
# 设置雷达图的角度，用于平分切开一个圆面
angles = np.linspace(0, 2*np.pi, N, endpoint=False)
# 使雷达图一圈封闭起来
angles = np.concatenate((angles,[angles[0]]))
for i in range(len(ccccc)):# 构造数据values = ccccc.loc[i,:]# 为了使雷达图一圈封闭起来values = np.concatenate((values,[values[0]]))# 绘制plt.polar(angles, values, 'o-', linewidth=2)
plt.legend(ccccc.index, loc='lower right')
# 添加极坐标的标签
plt.thetagrids(angles * 180/np.pi, labels=list(ccccc.columns))
plt.title('重要指标雷达图呈现')

Text(0.5,1.05,'重要指标雷达图呈现')

png

转载于:https://www.cnblogs.com/cvlas/p/9537532.html

本章节的目的是 【明确目标用户群】 ，以更好的服务现有用户。

【知识点】

1.作图

2.数据库操作

3.批量读取文件

4.groupby()以及agg() 的联合使用，应对不同列使用不同的函数

5.KMeans 聚类的使用

6.主成分分析

一、库导入以及matplotlib显示中文

二、批量读取文件

三、数据预处理

1.合并 DAU DPU

2.按月统计

4.使用 Kmeans 进行分类， 得到排名靠前的用户，即 重度用户/中度用户/轻度用户

对 A47 列进行聚类，分为3类

限定排名靠前的用户，即得分较高的重度和中度用户，以便接下来进行分析

5.主成分分析

1.删掉 0 值比较多的列

2.删掉相关性较强的列

进行主成分分析

得到降维后的数据

6.KMeans 进行聚类

将分类后的类别添加至原数据中

不进行预处理的降维

手肘法获取最优 K 值

显然，先标准化数据是不合适的

显然，不进行预处理的数据聚类是有问题的， 第一主成分和第二主成分 显然是相关的

相关文章

本章节的目的是【明确目标用户群】，以更好的服务现有用户。

4.使用 Kmeans 进行分类，得到排名靠前的用户，即重度用户/中度用户/轻度用户

显然，不进行预处理的数据聚类是有问题的，第一主成分和第二主成分显然是相关的