竞赛链接：https://www.kaggle.com/competitions/home-credit-default-risk/

认识数据集：application的两张表是申请人信息

通过id关联bureau：过去的借款、previous_application两张表

而bureau_balance则代表对应的还款信息

表之间的关系如下：

第一部分code我们做数据探索：看缺失值、异常值、相关性情况，并做填补及字段筛选，而后用逻辑回归和随机森林分别建立baseline，最终得分0.70，离0.85的第一名得分有较大差距，后几讲会再优化

具体代码如下：

coding: utf-8

In $1$ :

import numpy as np

In $2$ :

import pandas as pd

import time

In $3$ :

start_time=time.time()

In $4$ :

application_train=pd.read_csv('./application_train.csv',nrows=100000)

In $5$ :

application_test=pd.read_csv('./application_test.csv')

In $6$ :

previous_application=pd.read_csv('./previous_application.csv',nrows=100000)

In $7$ :

bureau_df=pd.read_csv('./bureau.csv',nrows=100000)

In $8$ :

bureau_balance=pd.read_csv('./bureau_balance.csv',nrows=100000)

In $9$ :

POS_CASH_balance=pd.read_csv('./POS_CASH_balance.csv',nrows=100000)

In $10$ :

credit_card_balance=pd.read_csv('./credit_card_balance.csv',nrows=100000)

In $11$ :

installments_payments=pd.read_csv('./installments_payments.csv',nrows=100000)

In $12$ :

application_train.memory_usage()

In $13$ :

print(f'application_train.shape:{application_train.shape}')

In $14$ :

class_counts=application_train $'TARGET'$ .value_counts()#

In $15$ :

import matplotlib.pyplot as plt

plt.pie(class_counts,labels=class_counts.index,autopct='%1.1f%%')#显示一位小数的百分比

In $16$ :

application_train.head()#发现训练集中缺失的id都出现在了测试集中

In $17$ :

application_train.select_dtypes('object')#看哪些是文本类型

In $18$ :

#看缺失情况

def missing(df):

missing_number=df.isnull().sum().sort_values(ascending=False)#sum看有几个缺失值，count看一共有几个值，如果直接count会踢掉缺失值再看有几个值

missing_percent=(df.isnull().sum()/df.isnull().count()).sort_values(ascending=False)

missing_values=pd.concat( $missing_number,missing_percent$ ,axis=1,keys= $'missing_number','missing_percent'$ )

return missing_values

In $19$ :

missing(application_train).sort_values(by='missing_percent',ascending=False)

In $20$ :

missing(application_train) $missing(application_train)\['missing_number'$ >0].index

In $21$ :

#想怎么填补

#类别型变量应该由众数填补

application_train $application_train\['NAME_TYPE_SUITE'$ =='Unaccompanied'] $'TARGET'$ .mean()

In $22$ :

application_train $application_train\['NAME_TYPE_SUITE'$ .isna()] $'TARGET'$ .mean()

In $23$ :

#发现二者违约率不一样，不适合这样填补，因此填成一个特殊的群体

application_train $'NAME_TYPE_SUITE'$ =application_train $'NAME_TYPE_SUITE'$ .fillna('Unknow')

In $24$ :

application_test $'NAME_TYPE_SUITE'$ =application_test $'NAME_TYPE_SUITE'$ .fillna('Unknow')

In $25$ :

application_train $'OWN_CAR_AGE'$ .isnull().sum()

In $26$ :

#想一个人为啥没车，可能FLAG_OWN_CAR也是N

application_train.loc $application_train\['OWN_CAR_AGE'$ .isnull()&(application_train $'FLAG_OWN_CAR'$ =='Y')] $\['OWN_CAR_AGE','FLAG_OWN_CAR'$ ]

In $27$ :

#填充没有车的人车龄为0

application_train.loc $application_train\['FLAG_OWN_CAR'$ =='N','OWN_CAR_AGE']=application_train.loc $application_train\['FLAG_OWN_CAR'$ =='N','OWN_CAR_AGE'].fillna(0)

application_test.loc $application_test\['FLAG_OWN_CAR'$ =='N','OWN_CAR_AGE']=application_test.loc $application_test\['FLAG_OWN_CAR'$ =='N','OWN_CAR_AGE'].fillna(0)

In $28$ :

#看填充结果

application_train $'OWN_CAR_AGE'$ .isna().sum()

In $29$ :

#再看上次换电话号码的时间,发现有大量是申请当天换的电话号码，这些是没有意义的

application_train $'DAYS_LAST_PHONE_CHANGE'$ .value_counts()

In $30$ :

#考虑把这些设为缺失值，后面再用均值或者中位数填补

application_train $'DAYS_LAST_PHONE_CHANGE'$ .replace(0,np.nan,inplace=True)

In $31$ :

#有时没有缺失值，但有XNA,测试集没有，因此可以把它删掉

application_train $'CODE_GENDER'$ .value_counts()

In $32$ :

application_train=application_train $application_train\['CODE_GENDER'$ !='XNA']

In $33$ :

#开始看异常值

#三类异常：看描述性统计，minmax是否远离均值/看箱线图，是否有离群点/3西格玛法则，看25%和75%分位数是否和minmax差别过大

application_train.describe()

In $34$ :

#观察发现DAYS_EMPLOYED最大值特别大

(application_train $'DAYS_EMPLOYED'$ /365).describe()

In $35$ :

application_train.loc $application_train\['TARGET'$ ==0,'DAYS_EMPLOYED'].hist()

In $36$ :

#直方图只适合离散值，连续值需要核密度估计图

import seaborn as sns

In $37$ :

sns.kdeplot(application_train.loc $application_train\['TARGET'$ ==0,'DAYS_EMPLOYED']/365,label='target'=='0')

In $38$ :

#写一个二分类的核密度直方图函数

def kde_plot(feature_name,df):

plt.figure(figsize=(8,6))

sns.kdeplot(df.loc $df\['TARGET'$ ==0,feature_name],label='target==0')

sns.kdeplot(df.loc $df\['TARGET'$ ==1,feature_name],label='target==1')

plt.legend()#显示曲线所代表的含义

plt.rcParams $'font.sans-serif'$ = $'SimHei'$

plt.rcParams $'axes.unicode_minus'$ =False

plt.show()

In $39$ :

kde_plot('DAYS_EMPLOYED',application_train)

#发现标签为0的异常值较多，因此

In $40$ :

#把异常值置空并留一列说明这些是异常值

application_train $'DAYS_EMPLOYED_ANOM'$ =application_train $"DAYS_EMPLOYED"$

application_train $'DAYS_EMPLOYED'$ .replace({365243:np.nan},inplace=True)

application_test $'DAYS_EMPLOYED_ANOM'$ =application_test $"DAYS_EMPLOYED"$

application_test $'DAYS_EMPLOYED'$ .replace({365243:np.nan},inplace=True)

In $41$ :

#看特征关联性,可视化;相关系数;特征重要性

In $42$ :

kde_plot('EXT_SOURCE_3',application_train)

#发现这个字段对标签影响比较大，特征工程时可以多考虑

In $43$ :

#再看看小提琴图，它既可以反应数据的分位数情况，也可以反应数据的密度情况

plt.figure(figsize=(10,8))

sns.violinplot(x='TARGET',y='EXT_SOURCE_3',data=application_train)

plt.show()

In $44$ :

#再看几个连续型变量

kde_plot('DAYS_BIRTH',application_train)

In $45$ :

#看相关性

correlations=application_train.corr() $'TARGET'$ .sort_values()

correlations

In $46$ :

correlations.tail(15)

#看正向最重要的15个特征

In $47$ :

#看绝对值

correlations_abs=abs(correlations).sort_values(ascending=False) $:11$

correlations_abs

In $48$ :

#特征间关系，热力图，选10个最强的特征来画

correlations=application_train.corr()

In $49$ :

plt.figure(figsize=(30,40))

sns.heatmap(correlations $correlations_abs.index.tolist()$ )

plt.show()

In $50$ :

#发现留个变量有较强相关性

ext_data=application_train $\['TARGET','DAYS_BIRTH','FLAG_EMP_PHONE','EXT_SOURCE_1','DAYS_EMPLOYED_ANOM'$ ]

In $51$ :

ext_data_corrs=ext_data.corr()

In $52$ :

plt.figure(figsize=(10,8))

sns.heatmap(ext_data_corrs,cmap='RdBu_r',annot=True,fmt=".2f")#颜色，把字写入

plt.show()

In $53$ :

application_train $application_train\['DAYS_EMPLOYED_ANOM'$ ==1] $'NAME_INCOME_TYPE'$ .value_counts()

In $54$ :

#发现新创建的这列DAYS_EMPLOYED_ANOM的信息可能已经被其它特征所反映，但如果能从业务角度挖掘出特别何原因，会对建模有很大帮助

In $55$ :

#验证EXT_SOURCE_1和DAYS_BIRTH有相关性，用六边形图

x=application_train $'EXT_SOURCE_1'$

y=application_train $'DAYS_BIRTH'$

plt.hexbin(x,y,gridsize=30)

plt.show()

In $56$ :

#海量数据处理的方法

import polars as pl

In $57$ :

df_pl=pl.read_csv('application_train.csv')

In $58$ :

df_pl.head()

In $59$ :

#建立baseline

bureau=pd.read_csv('./bureau.csv',nrows=100000)

In $60$ :

#先把类别型变量作数据编码。用label encoder，这对树模型不会有影响

#具体使用factorize，它对缺失值和异常值都会分配一个新值，防止自己先做填充出问题

#在合并时会遇到训练集和测试集对不齐（测试集多一列）的问题，解决方法是把训练集和测试集合起来再进行one-hot编码

#然后找到target是nan的

apply=application_train.append(application_test)

In $61$ :

object_col=apply.dtypes $apply.dtypes=='object'$ .index.to_list()

In $62$ :

for col in object_col:

if len(apply $col$ .unique())>2:

apply=pd.concat( $apply,pd.get_dummies(apply\[col$ ,prefix=col)],axis=1)#生成独热编码,prefix是前缀

apply.drop(columns= $col$ ,inplace=True)#inplace表示是否删副本

else:

apply $col$ =pd.factorize(apply $col$ ) $0$ #数值型编码

apply.head()

In $63$ :

#分割训练集和测试集，target为null的就是测试集

application_test=apply $apply\['TARGET'$ .isnull()]

application_test=application_test.drop('TARGET',axis=1)

application_train=apply $\~apply\['TARGET'$ .isnull()]

In $64$ :

#逻辑回归，需要填补缺失值，并进行缩放

from sklearn.preprocessing import MinMaxScaler

from sklearn.impute import SimpleImputer#用来算minmax

In $65$ :

train=application_train.drop(columns= $'TARGET','SK_ID_CURR'$ )#ID和TARGET作编号时无用

In $66$ :

features=list(train.columns)

In $67$ :

imputer=SimpleImputer(strategy='median')

In $68$ :

scaler=MinMaxScaler(feature_range=(0,1))

scaler

In $69$ :

#在训练集上进行拟合

imputer.fit(train.append(application_test $features$ ))

In $70$ :

train=imputer.transform(train)

test=imputer.transform(application_test $features$ )

train

In $71$ :

scaler.fit(train)

train=scaler.transform(train)

test=scaler.transform(test)

test

In $72$ :

#训练模型

from sklearn.linear_model import LogisticRegression

log_reg=LogisticRegression(C=0.0001)

log_reg.fit(train,application_train $'TARGET'$ )

In $73$ :

#进行预测，确保只获取第二列（为1的概率）

log_reg_pred=log_reg.predict_proba(test) $:,1$

test

In $74$ :

#获取特征的系数

coefficients=log_reg.coef_ $0$ #把数组转为整数

coefficients

In $75$ :

#看特征重要性

feature_importance=np.abs(coefficients)

In $76$ :

#给特征重要性排序,得出每个特征的重要性排名

sorted_indices=np.argsort(feature_importance) $::-1$

In $77$ :

for idx in sorted_indices:

print(f"{features $idx$ },IMPORTANCE：{feature_importance $idx$ }")

In $78$ :

np.argsort(feature_importance)

In $79$ :

coefficients $::-1$

In $80$ :

#保存结果

submit=application_test $\['SK_ID_CURR'$ ]

In $81$ :

submit $'TARGET'$ =log_reg_pred

In $82$ :

submit

In $83$ :

#保存结果

submit.to_csv('baseline_model_log_reg.csv',index=False)

In $84$ :

#再尝试下其它类型的模型，随机森林

#区别于逻辑回归，它不需要缩放

train=application_train.drop(columns= $'TARGET','SK_ID_CURR'$ )#ID和TARGET作编号时无用

features=list(train.columns)

imputer=SimpleImputer(strategy='median')

imputer.fit(train.append(application_test $features$ ))

train=imputer.transform(train)

test=imputer.transform(application_test $features$ )

In $85$ :

from sklearn.ensemble import RandomForestClassifier

random_forest=RandomForestClassifier(n_estimators=1000,random_state=2024,verbose=1,n_jobs=-1)

In $86$ :

random_forest.fit(train,application_train $'TARGET'$ )

#提取特征重要性

feature_importance_values=random_forest.feature_importances_

feature_importances=pd.DataFrame({'feature':features,'importance':feature_importance_values})

In $88$ :

#在测试数据上预测

predictions=random_forest.predict_proba(test) $:,1$

#并保存为提交文件

submit=application_test $\['SK_ID_CURR'$ ]

submit $'TARGET'$ =predictions

In $89$ :

#保存文件

submit.to_csv('baseline_model_random_forest.csv',index=False)

#0.703分，比逻辑回归稍好些

KAGGLE竞赛实战2-捷信金融违约预测竞赛-part1-数据探索及baseline建立

coding: utf-8

In1:

In2:

In3:

In4:

In5:

In6:

In7:

In8:

In9:

In10:

In11:

In12:

In13:

In14:

In15:

In16:

In17:

In18:

In19:

In20:

In21:

In22:

In23:

In24:

In25:

In26:

In27:

In28:

In29:

In30:

In31:

In32:

In33:

In34:

In35:

In36:

In37:

In38:

In39:

In40:

In41:

In42:

In43:

In44:

In45:

In46:

In47:

In48:

In49:

In50:

In51:

In52:

In53:

In54:

In55:

In56:

In57:

In58:

In59:

In60:

In61:

In62:

In63:

In64:

In65:

In66:

In67:

In68:

In69:

In70:

In71:

In72:

In73:

In74:

In75:

In76:

In77:

In78:

In $1$ :

In $2$ :

In $3$ :

In $4$ :

In $5$ :

In $6$ :

In $7$ :

In $8$ :

In $9$ :

In $10$ :

In $11$ :

In $12$ :

In $13$ :

In $14$ :

In $15$ :

In $16$ :

In $17$ :

In $18$ :

In $19$ :

In $20$ :

In $21$ :

In $22$ :

In $23$ :

In $24$ :

In $25$ :

In $26$ :

In $27$ :

In $28$ :

In $29$ :

In $30$ :

In $31$ :

In $32$ :

In $33$ :

In $34$ :

In $35$ :

In $36$ :

In $37$ :

In $38$ :

In $39$ :

In $40$ :

In $41$ :

In $42$ :

In $43$ :

In $44$ :

In $45$ :

In $46$ :

In $47$ :

In $48$ :

In $49$ :

In $50$ :

In $51$ :

In $52$ :

In $53$ :

In $54$ :

In $55$ :

In $56$ :

In $57$ :

In $58$ :

In $59$ :

In $60$ :

In $61$ :

In $62$ :

In $63$ :

In $64$ :

In $65$ :

In $66$ :

In $67$ :

In $68$ :

In $69$ :

In $70$ :

In $71$ :

In $72$ :

In $73$ :

In $74$ :

In $75$ :

In $76$ :

In $77$ :

In $78$ :

In $79$ :

In $80$ :