Table of contents
Edit Proportion of usage of coupons
Write-off status and average collection distance of different types of coupons
Distribution of merchants with different write-off rates (proportion)
Distribution of users with different write-off rates (proportion)
Average coupon write-off cycle and write-off rate on different working days
1. Import and understand data characteristics
2. Data preprocessing: analysis of Titanic shipwreck data
2.1 Read the data and randomly obtain 5 lines of observed data characteristics
2.2 Check the missing values of each field
2.3 Missing value analysis and missing value processing
3. Analysis of data characteristics
3.1.1 Summary of survival and death of all members
3.1.3 Analysis of other characteristic data
3.1.4 Effects of continuous-valued features (age, ticket fare) on survival outcomes
3.1.7 Feature Association Analysis
4.1 Label encoding preprocessing
4.2 Remove meaningless features
4.3 Remove similar duplicate features
4.4 Visually explore the distribution of each feature
5.2. Establish model training and evaluation functions
6.5 Accuracy comparison between different models
SQL
Statistical data overview: calculate the total number of samples, the total number of merchants, the total number of users, the total number of consumption, the total number of coupons, etc.
select
count(User_id) as '样本总数',
count(distinct Merchant_id) as '商家总数',
count(distinct User_id) as '用户总数',
count(Date) as '消费总数',
count(Date_received) as '领券总数',
(select count(*) from ddm.offline_train as a where a.Date_received is not null and a.Date is not null) as '领券消费总数',
(select count(*) from ddm.offline_train as a where a.Date_received is null and a.Date is not null) as '无券消费总数',
(select count(*) from ddm.offline_train as a where a.Date_received is not null and a.Date is not null)/count(Date_received) as '核销率'
from ddm.offline_train
Statistics at different distances: the number of coupon recipients, the number of coupon consumers, and the write-off rate
# 查找各距离的领券人数/用券消费人数/核销率
select
Distance,
count(Coupon_id) as get_coupons_num,
sum(if(Date_received is not null and Date is not null,1,0)) as user_coupons_num,
sum(if(Date_received is not null and Date is not null,1,0)) /count(Coupon_id) as use_coupons_rate
from ddm.offline_train
where Distance is not null
group by Distance
order by distance
Proportion of coupon usage
# 消费券使用情况占比
with temp as (
select
case
when Date_received is not null and Date is not null then '有券消费'
when Date_received is not null and Date is null then '有券未消费'
when Date_received is null and Date is not null then '无券消费'
end as flag
from ddm.offline_train
)
select
flag as '优惠券使用情况',
concat(round(count(flag)/(select count(*) from temp)*100,2),'%') as '百分比'
from temp
group by flag
order by count(flag)/(select count(*) from temp) 
with as is also called the subquery part, which is similar to a view or a temporary table. It can be used to store a part of the SQL statement query results. It must be used together with other query statements, and there should be no semicolon in the middle. Currently, in oracle, sql server, Both hive and others support the usage of with as, but mysql does not support it!
Write-off status and average collection distance of different types of coupons
# 不同优惠券类型的核销情况和平均领取距离
select
Discount_rate as '折扣',
avg(Distance) as '平均距离',
count(Date_received) as '领券人数',
sum(if(Date_received is not null and Date is not null,1,0)) as '有券消费人数',
sum(if(Date_received is not null and Date is not null,1,0))/count(Date_received) as '核销率'
from ddm.offline_train
where Date_received is not null
group by Discount_rate
order by '有券消费人数' desc
Write-offs with different full reduction thresholds
# 不同满减门槛的核销情况
select
mk as '门槛',
count(*) as '领券数量',
sum(if(Date is not null,1,0)) as '用券消费数量',
concat(round(sum(if(Date is not null,1,0))/count(*)*100,2),'%') as '核销率'
from(select
DATE,
convert(if(Discount_rate like '%.%',0,Discount_rate),signed) as mk
from ddm.offline_train) as aa
where mk is not null
group by mk
order by mk
Distribution of merchants with different write-off rates (proportion)
# 不同核销率用户分布
with temp as (
select
Merchant_id,
count(Date_received) as get_num,
sum(if(Date is not null and Date_received is not null,1,0)) as use_num,
sum(if(Date is not null and Date_received is not null,1,0))/count(Date_received) as Merchant_rate
from ddm.offline_train
where Date_received is not null
group by Merchant_id
)
select
tag,
concat(round(count(*)/(select count(*) from temp)*100,2),'%') as Merchant_percent
from(
select
Merchant_id,
case
when Merchant_rate = 0 then '核销率：0'
when Merchant_rate > 0 and Merchant_rate < 0.2 then '核销率：0-20%'
when Merchant_rate >= 0.2 and Merchant_rate< 0.3 then '核销率：20%-30%'
when Merchant_rate >= 0.3 and Merchant_rate< 0.5 then '核销率：30%-50%'
when Merchant_rate >= 0.5 then '核销率：50%以上'
end as tag
from temp
)aa
group by tag
order by Merchant_percent desc
The distribution of merchants with different coupon collection times (average write-off rate/proportion)
# 不同领券次数用户分布-平均核销率/占比
with temp as (
select
Merchant_id,
count(Date_received) as get_num,
sum(if(Date is not null and Date_received is not null,1,0))/count(Date_received) as user_rate,
sum(if(Date is not null and Date_received is not null,1,0)) as use_num,
case
when count(Date_received)>100 then '100次以上'
when count(Date_received)=0 then '0次'
when count(Date_received) between 1 and 10 then '1-10次'
when count(Date_received) between 11 and 50 then '11-50次'
when count(Date_received) between 51 and 100 then '51-100次'
else '其他次'
end as flag
from ddm.offline_train
group by Merchant_id
)
select
flag as '被领券次数',
concat(round(avg(user_rate)*100,2),'%') as Merchant_avg_use_rate,
concat(round(count(*)/(select count(*) from temp)*100,2),'%') as Merchant_percent
from temp
group by flag
order by (count(*)/(select count(*) from temp)) desc
Distribution of users with different write-off rates (proportion)
# 不同核销率用户分布
with temp as (
select
User_id,
count(Date_received) as get_num,
sum(if(Date is not null and Date_received is not null,1,0)) as use_num,
sum(if(Date is not null and Date_received is not null,1,0))/count(Date_received) as user_rate
from ddm.offline_train
where Date_received is not null
group by User_id
)
select
tag,
concat(round(count(*)/(select count(*) from temp)*100,2),'%') as user_percent
from(
select
User_id,
case
when user_rate = 0 then '核销率：0'
when user_rate > 0 and user_rate < 0.3 then '核销率：0-30%'
when user_rate >= 0.3 and user_rate< 0.5 then '核销率：30%-50%'
when user_rate >= 0.5 then '核销率：50%以上'
end as tag
from temp
)aa
group by tag
order by user_percent desc
Number of Coupons Received / Number of Verifications / Verification Rate of Coupons in Different Months
# 不同月份领券次数/核销次数/核销率
select
`month`,
coupons_get_num,
coupons_use_num,
concat(round(coupons_use_num/coupons_get_num*100,2),'%') as coupons_use_rate
from(select
month(Date_received) as `month`,
count(*) as coupons_get_num
from ddm.offline_train
where Date_received is not null
group by month(Date_received)) as a
inner join(
select
month(Date) as `month`,
count(*) as coupons_use_num
from ddm.offline_train
where Date_received is not null and Date is not null
group by month(Date)
)as b using(`month`)
order by `month`
Average coupon write-off cycle and write-off rate on different working days
# 工作日平均核销间隔、核销率
with get_coupons as(
select
weekday(Date_received)+1 as coupons_day,
count(*) as coupons_get_num
from ddm.offline_train
where Date_received is not null
group by weekday(Date_received)+1
),
use_coupons as(
select
weekday(Date)+1 as coupons_day,
count(*) as coupons_use_num,
round(avg(datediff(Date,Date_received)),2) as use_interval
from ddm.offline_train
where Date is not null and Date_received is not null
group by weekday(Date)+1
)
select
coupons_day,
use_interval,
concat(round(coupons_use_num/coupons_get_num*100,2),'%') as coupons_use_rate
from get_coupons
inner join use_coupons using(coupons_day)
order by coupons_day
Python
library package
- pandas : supports loading data from CSV, JSON, SQL, and Microsoft Excel, and can merge, reshape, and select data of various dimensions, as well as data cleaning and data processing features
- numpy : Packages for storing and manipulating matrices and array calculations: array calculations, logical operations, Fourier transforms and graphics operations, operations related to linear algebra
- Matplotlib : Drawing tools can draw line graphs, scatter plots, contour plots, bar charts, histograms, 3D graphics, and even graphic animations.
- sklearn : It covers six modules of classification, regression, clustering, dimensionality reduction, model selection, and data preprocessing, lowers the threshold for machine learning practice, integrates complex mathematical calculations into simple functions, and provides many public data sets and Study case.
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from pylab import *
import seaborn as sns
from sklearn import model_selection, preprocessing, naive_bayes, metrics, svm
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn import ensemble, tree
# 忽略警告提示
import warnings
warnings.filterwarnings('ignore')1. Import and understand data characteristics
data = pd.read_csv(r"/Users/xll/Documents/服务数据质量/测试数据1.csv")
#设置python中dataframe的print结果不带省略号
pd.set_option('display.max_columns',5, 'display.max_rows', 100)
#返回数据集行和列的元组，其中data.shape[0]代表返回行数，data.shape[1] 代表返回列数
print(data.shape)
#返回数据集的所有列名
data.columns
#随机返回样本5行
data.sample(5)
#返回前5行
print(data.head(5))
#返回浮点型和正行字段的均值、最大值、等统计数据
print(data.describe())
#numpy包的这个方法也是可以得到同样的结果
import numpy as np
print(data.describe(include=[np.number]))
2. Data preprocessing: analysis of Titanic shipwreck data
2.1 Read the data and randomly obtain 5 lines of observed data characteristics
data = pd.read_csv('titanic.csv')
print(data.shape) #输出：(891, 15)
data.sample(5)
2.2 Check the missing values of each field
data.isnull().sum()
2.3 Missing value analysis and missing value processing
There are missing values in age, deck, embarked, and embark_town, which need to be processed.
(1) age has an impact on the survival rate, which cannot be ignored, and is filled with the average value;
(2) There are a total of 891 pieces of information, and deck has 688 missing values, so the classification label of deck is removed;
(3) Embarked and embark_town have fewer missing values, both of which are 2, and one of them is randomly selected to fill.
//用平均年龄补充年龄缺失的字段
data['age']=data['age'].fillna(data['age'].median())
//因该特征缺失占比是99%，所以删除deck特征
del data['deck']
//只有两条记录缺失，随机补充
data['embarked']=data['embarked'].fillna('S')
data['embark_town']=data['embark_town'].fillna('Southampton')
//再次查询下空值的记录总和，返回都是0
data.isnull().sum()3. Analysis of data characteristics
3.1.1 Summary of survival and death of all members
//定义一个survive的矩阵：死亡为0，生存为1；统计死亡人数、生存人数和死亡及生存率
survived = data['survived'].value_counts().to_frame().reset_index().rename(columns={'index': 'label', 'survived': 'counts'})
#计算存活率
survived_rate = round(342/891, 2)
survived['rate'] = [1-survived_rate, survived_rate]
print(survived)
3.1.2 Draw a pie chart
mpl.rcParams['axes.unicode_minus'] = False #处理无法显示中文的问题
mpl.rcParams['font.sans-serif'] = ['SimHei']
fig=plt.figure(1,figsize=(6,6))
ax1=fig.add_subplot(1,1,1)
label=['遇难','存活']
color=['#C23531','#F5DEB3']
explode=0.05,0.05 #扇区间隔
patches,l_text,p_text = ax1.pie(survived.rate,labels=label,colors=color,startangle=90,autopct='%1.0f%%',explode=explode,shadow=True)
for t in l_text:
t.set_size(20)
for t in p_text:
t.set_size(20)
ax1.set_title('全体成员的生存情况', fontsize=20) 
3.1.3 Analysis of other characteristic data
fig = plt.figure(figsize=(15,10))
fig.set(alpha=0.3) # 设定图表颜色alpha参数(透明度)
plt.subplot2grid((2,3),(0,0))
data.survived.value_counts().plot(kind='bar')
plt.title("获救情况 (1为获救)")
plt.ylabel("人数")
plt.subplot2grid((2,3),(0,1))
data.pclass.value_counts().plot(kind="bar")
plt.ylabel("人数")
plt.title("乘客等级分布")
plt.subplot2grid((2,3),(0,2))
plt.scatter(data.survived, data.age)
plt.ylabel("年龄")
plt.grid(b=True, which='major', axis='y')
plt.title("按年龄看获救分布 (1为获救)")
plt.subplot2grid((2,3),(1,0), colspan=2)
data.age[data.pclass == 1].plot(kind='kde')
data.age[data.pclass == 2].plot(kind='kde')
data.age[data.pclass == 3].plot(kind='kde')
plt.xlabel("年龄")
plt.ylabel("密度")
plt.title("各等级的乘客年龄分布")
plt.legend(('头等舱', '2等舱','3等舱'),loc='best')
plt.subplot2grid((2,3),(1,2))
data.embarked.value_counts().plot(kind='bar')
plt.title("各登船口岸上船人数")
plt.ylabel("人数")
plt.show()
3.1.4 Effects of continuous-valued features (age, ticket fare) on survival outcomes
fig = plt.figure(figsize=(15,4))
plt.subplot2grid((2,2),(0,0))
data.age[data.survived == 0].plot(kind='box', vert=False, patch_artist=True, notch = True, color='#C23531', fontsize=15)
plt.grid(linestyle="--", alpha=0.8)
plt.title("遇难", fontsize=15)
plt.subplot2grid((2,2),(0,1))
data.fare[data.survived == 0].plot(kind='box', vert=False, patch_artist=True, notch = True, color='#C23531', fontsize=15)
plt.grid(linestyle="--", alpha=0.8)
plt.title("遇难", fontsize=15)
plt.subplot2grid((2,2),(1,0))
data.age[data.survived == 1].plot(kind='box', vert=False, patch_artist=True, notch = True, color='#F5DEB3', fontsize=15)
plt.grid(linestyle="--", alpha=0.8)
plt.xlabel("存活", fontsize=15)
plt.subplot2grid((2,2),(1,1))
data.fare[data.survived == 1].plot(kind='box', vert=False, patch_artist=True, notch = True, color='#F5DEB3', fontsize=15)
plt.grid(linestyle="--", alpha=0.8)
plt.xlabel("存活", fontsize=15)
3.1.5 The impact of passenger class and gender on survival outcomes (from the perspective of age distribution)
mpl.rcParams.update({'font.size': 14})
fig,axes=plt.subplots(2,2,figsize=(18, 12))
sns.violinplot("pclass","age", hue="survived", data=data, palette='autumn',ax=axes[0][0]).set_title('Pclass and Age vs Survived')
sns.swarmplot(x="pclass", y="age",hue="survived", data=data,palette='autumn',ax=axes[1][0]).legend(loc='upper right').set_title('survived')
sns.violinplot("sex","age", hue="survived", data=data, palette='winter', ax=axes[0][1]).set_title('Sex and Age vs Survived')
sns.swarmplot(x="sex", y="age",hue="survived", data=data,palette='winter',ax=axes[1][1]).legend(loc='upper right').set_title('survived') 
3.1.6 The impact of passenger class and gender on survival outcomes (from the distribution of ticket fees)
fig,axes=plt.subplots(2,2,figsize=(18, 12))
sns.violinplot("pclass","fare", hue="survived", data=data, palette='autumn',ax=axes[0][0]).set_title('Pclass and Age vs Survived')
sns.stripplot("pclass", "fare",hue="survived", data=data,palette='autumn',ax=axes[1][0]).legend(loc='upper right').set_title('survived')
sns.violinplot("sex","fare", hue="survived", data=data, palette='winter', ax=axes[0][1]).set_title('Sex and Age vs Survived')
sns.stripplot("sex", "fare",hue="survived", data=data,palette='winter',ax=axes[1][1]).legend(loc='upper right').set_title('survived')
3.1.7 Feature Association Analysis
sns.heatmap(data.corr(),annot=True,cmap='RdYlGn',linewidths=0.2)
fig=plt.gcf()
fig.set_size_inches(10,8)
plt.show()
4. Feature engineering
4.1 Label encoding preprocessing
Among all labels, survived is a categorical label and the remaining 14 variables are categorical features. Since the values of features and labels have unstructured types, feature engineering processing, that is, string encoding processing, is required. Personal understanding is actually to convert non-integer or floating-point types into integers
print(data.info())
Before processing: the data of the first 5 rows is like this
print(data.head())
Initialize encoder processing: the first 5 lines after processing, all become numeric
le = preprocessing.LabelEncoder()
for col in data.columns:
data[col] = le.fit_transform(data[col])
data.head()
data.to_csv('Preprocessing_Titanic.csv')
4.2 Remove meaningless features
Names have little effect on survival, so remove the who tag
del data['who']4.3 Remove similar duplicate features
eg: Gender in column 1: Difficult, female and 0, 1 in column 2, this is based on the encoding in the previous step and processed into a numeric type; after removal, it is found that there are three repeated features
data_ = data.T.drop_duplicates().T
print('去重前：', len(data.columns))
print('去重后：', len(data_.columns))
for a in data.columns:
if a not in data_.columns:
for b in data_.columns:
if list(data[b].values) == list(data[a].values):
print(f'重复标签: {a} 和 {b}')
data = data_
#drop_duplicates方法，它用于返回一个移除了重复行的DataFrame
#所以这里先做了转置的操作，剔除后，再转置回来的
#保留第一个重复行 data.drop_duplicates()
#去除所有重复行 data.drop_duplicates(keep=False)
#合并起来再去重，只剩下真的重复行 pd.concat([data0_1,data0_2]).drop_duplicates(keep=False)
4.4 Visually explore the distribution of each feature
result_plot = data.hist(bins=50, figsize=(14, 12))
Judging from the above visualization, there is no need to standardize the features.
# 对数据进行标准化
# X = StandardScaler().fit_transform(X)5. Model training
5.1 Split the data set with 10-fold cross-validation, 9 for training and 1 for testing, to ensure that there is no intersection between training and test data
# data第一列是survived分类标签，作为y输出结果，其他类都是分类的特征，用于输入X
X = data.iloc[:, 1:]
y = data.iloc[:, 0]
x_train, x_test, y_train, y_test = train_test_split(X, y,test_size=0.1,shuffle=True,random_state=20)
print(x_train.shape)
print(x_test.shape)
print(y)
print(X)
5.2. Establish model training and evaluation functions
5.2.1 Modeling: training the model
model, train_score, test_score, roc_auc = [], [], [], [] # 存储相关模型信息，以便后续分析
def train_model(classifier, x_train, y_train, x_test):
lr = classifier # 初始化
lr.fit(x_train, y_train) # 训练
y_pred_lr = lr.predict(x_test) # 预测
if '.' in str(classifier):
model_name = str(classifier).split('(')[0].split('Classifier')[0].split('.')[1]
print('\n{:=^60}'.format(model_name))
else:
model_name = str(classifier).split('(')[0].split('Classifier')[0]
print('\n{:=^60}'.format(model_name))
model.append(model_name)
# 性能评估
print('\n>>>在训练集上的表现：', lr.score(x_train, y_train))
print('\n>>>在测试集上的表现：', metrics.accuracy_score(y_test, y_pred_lr))
print('\n>>>预测的 Roc_auc：%.4f' % metrics.roc_auc_score(y_test, y_pred_lr))
print('\n>>>混淆矩阵'),show_confusion_matrix(metrics.confusion_matrix(y_test,y_pred_lr))
train_score.append(lr.score(x_train, y_train))
test_score.append(metrics.accuracy_score(y_test, y_pred_lr))
roc_auc.append(metrics.roc_auc_score(y_test, y_pred_lr))5.2.2 Plotting Categorical Matrix Functions
def show_confusion_matrix(cnf_matrix):
plt.matshow(cnf_matrix,cmap=plt.cm.YlGn,alpha=0.7)
ax=plt.gca()
ax.set_xlabel('Predicted Label',fontsize=16)
ax.set_xticks(range(0,len(survived.label)))
ax.set_xticklabels(survived.label,rotation=45)
ax.set_ylabel('Actual Label',fontsize=16,rotation=90)
ax.set_yticks(range(0,len(survived.label)))
ax.set_yticklabels(survived.label)
ax.xaxis.set_label_position('top')
ax.xaxis.tick_top()
for row in range(len(cnf_matrix)):
for col in range(len(cnf_matrix[row])):
ax.text(col,row,cnf_matrix[row][col],va='center',ha='center',fontsize=16)
plt.show()5.2.3 Draw the ROC curve function
def show_roc_line(classifier, x_train, y_train):
y_train_prob=classifier.predict_proba(x_train)
y_pred_prob=y_train_prob[:,1] #正例率
fpr,tpr,thresholds=metrics.roc_curve(y_train,y_pred_prob) #计算ROC曲线
auc=metrics.auc(fpr,tpr) #计算AUC
plt.plot(fpr,tpr,lw=2,label='ROC curve (area={:.2f})'.format(auc))
plt.plot([0,1],[0,1],'r--')
plt.xlim([-0.01, 1.02])
plt.ylim([-0.01, 1.02])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic')
plt.legend(loc='lower right')
plt.show() 6. Training predictions
6.1 Decision tree model
classifier = tree.DecisionTreeClassifier()
train_model(classifier, x_train, y_train, x_test) #建模
Plotting important features of a random forest
labels = X.columns
importances = classifier.feature_importances_ # 获取特征权重值
indices = np.argsort(importances)[::-1] # 打印特征等级
features = [labels[i] for i in indices]
weights = [importances[i] for i in indices]
print("Feature ranking:")
for f in range(len(features)):
print("%d. %s (%f)" % (f + 1, features[f], weights[f])) # 绘制随机森林的特征重要性
plt.figure()
plt.title("Feature importances")
plt.bar(features, np.array(weights), color='r')
plt.xticks(rotation=90)
plt.title('Feature Weights')
plt.show()
Data analysis: As can be seen from the visualization above, there are only four features that have a great impact on the survival rate: fare (ticket fee), adult_male (adult male), age (age), and pclass (passenger class).
6.2 KNN model
classifier = KNeighborsClassifier(n_neighbors=20)
train_model(classifier, x_train, y_train, x_test)
show_roc_line(classifier, x_train, y_train)
6.3 SVC model
classifier = svm.SVC()
train_model(classifier, x_train, y_train, x_test)
6.4 Naive Bayes
classifier = naive_bayes.GaussianNB()
train_model(classifier, x_train, y_train, x_test) #建模
show_roc_line(classifier, x_train, y_train) #绘制ROC曲线
6.5 Accuracy comparison between different models
df = pd.DataFrame()
df['model'] = model
df['Roc_auc'] = roc_auc
df['train_score'] = train_score
df['test_score'] = test_score
df
6.5.1 Conclusion:
Judging criteria: the area under the AUC (Area under Curve) ROC curve, between 0.1 and 1, when its value is used as a criterion for judging the quality of the classifier, the larger the value, the better;
So it can be seen that the naive Bayesian classification prediction effect is better, and the decision tree performs better on the training set, but its performance on the test machine is average, which means overfitting
6.5.2 Overfitting parameter adjustment
param = [{'criterion':['gini'],'max_depth': np.arange(20,50,10),'min_samples_leaf':np.arange(2,8,2),
'min_impurity_decrease':np.linspace(0.1,0.9,10)},
{'criterion':['gini','entropy']},
{'min_impurity_decrease':np.linspace(0.1,0.9,10)}]
clf = GridSearchCV(tree.DecisionTreeClassifier(),param_grid=param,cv=10)
clf.fit(x_train,y_train)
print('最优参数:', clf.best_params_)
print('最好成绩:', clf.best_score_)
6.5.3 Generate a decision tree with the above optimal parameters
model = tree.DecisionTreeClassifier(criterion= 'gini', max_depth=20, min_impurity_decrease=0.1, min_samples_leaf= 2)
model.fit(x_train, y_train)
y_pred = model.predict(x_test)
print('train score:', clf.score(x_train, y_train))
print('test score:', clf.score(x_test, y_test))
print("查准率：", metrics.precision_score(y_test,y_pred))
print('召回率:',metrics.recall_score(y_test,y_pred))
print('f1分数:', metrics.f1_score(y_test,y_pred)) #二分类评价标准
Judging from the results, it has improved, but it is still worse than the prediction score of Naive Bayes
reference:
https://blog.csdn.net/twlve/article/details/128609147?spm=1001.2014.3001.5502
https://blog.csdn.net/twlve/article/details/128626526?spm=1001.2014.3001.5502
O2O Coupon Verification-Data Analysis
https://blog.csdn.net/weixin_47068543/article/details/126151816