【挖掘模型】：R语言-决策树观察汽车信用贷款违约模式

背景：

目前有一批汽车信用贷款用户违约数据（客户属性 + 账号属性 + 消费行为 +还款行为），市场部门想根据这些数据建立决策树模型从而观察违约用户的违约模式，进而调整业务。

数据源：

data.csv（一份汽车贷款违约数据）
样本量：7193

24个观察指标

建模方法： 决策树-C5.0

建模结果：

违约模式

代码

> setwd("C:\\Users\\Administrator\\Desktop\\重新跑模型\\data\\")
> accepts<-read.csv("accepts.csv")
> accepts$bad_ind<-as.factor(accepts$bad_ind)
> names(accepts)
 [1] "application_id" "account_number" "bad_ind"        "vehicle_year"   "vehicle_make"   "bankruptcy_ind"
 [7] "tot_derog"      "tot_tr"         "age_oldest_tr"  "tot_open_tr"    "tot_rev_tr"     "tot_rev_debt"  
[13] "tot_rev_line"   "rev_util"       "fico_score"     "purch_price"    "msrp"           "down_pyt"      
[19] "loan_term"      "loan_amt"       "ltv"            "tot_income"     "veh_mileage"    "used_ind"      
> accepts=accepts[,c(3,7:24)]
> #根据业务理解生成更有意义的衍生变量，不过这些变量都是临时的，因为没有经过数据清洗，此处仅作一个示例
> #accepts$lti_temp=accepts$loan_amt/accepts$tot_income
> 
> set.seed(10)
> select<-sample(1:nrow(accepts),length(accepts$bad_ind)*0.7)
> train=accepts[select,]
> test=accepts[-select,]
> summary(train$bad_ind)
   0    1 
3233  858 
> ###################################
> ## Section 1: C50算法
> ###################################
> train<-na.omit(train)
> library(C50)
> #请注意，R中的C50包比较新，存在一些问题，比如遇到缺失值、字符类型变量会报错“c50 code called exit with value 1”
> ##建模
> tc<-C5.0Control(subset =F,CF=0.25,winnow=F,noGlobalPruning=F,minCases =20)
> model <- C5.0(bad_ind ~.,data=train,rules=F,control =tc)
> summary( model )

Call:
C5.0.formula(formula = bad_ind ~ ., data = train, rules = F, control = tc)


C5.0 [Release 2.07 GPL Edition]     Mon May 22 21:35:14 2017
-------------------------------

Class specified by attribute `outcome'

Read 3001 cases (19 attributes) from undefined.data

Decision tree:

fico_score > 661: 0 (2161/262)
fico_score <= 661:
:...tot_tr > 13:
    :...ltv <= 83: 0 (49/4)
    :   ltv > 83:
    :   :...fico_score <= 588: 1 (52/20)
    :       fico_score > 588: 0 (411/125)
    tot_tr <= 13:
    :...rev_util > 116: 1 (26/5)
        rev_util <= 116:
        :...used_ind > 0: 0 (181/78)
            used_ind <= 0:
            :...purch_price <= 25000: 1 (92/40)
                purch_price > 25000: 0 (29/5)


Evaluation on training data (3001 cases):

        Decision Tree   
      ----------------  
      Size      Errors  

         8  539(18.0%)   <<


       (a)   (b)    <-classified as
      ----  ----
      2357    65    (a): class 0
       474   105    (b): class 1


    Attribute usage:

    100.00% fico_score
     27.99% tot_tr
     17.06% ltv
     10.93% rev_util
     10.06% used_ind
      4.03% purch_price


Time: 0.0 secs

> #图形展示
> plot(model)
> C5imp(model)
              Overall
fico_score     100.00
tot_tr          27.99
ltv             17.06
rev_util        10.93
used_ind        10.06
purch_price      4.03
tot_derog        0.00
age_oldest_tr    0.00
tot_open_tr      0.00
tot_rev_tr       0.00
tot_rev_debt     0.00
tot_rev_line     0.00
msrp             0.00
down_pyt         0.00
loan_term        0.00
loan_amt         0.00
tot_income       0.00
veh_mileage      0.00
> #生成规则
> rule<- C5.0(bad_ind ~.,data=train,rules=T,control =tc)
> summary( rule )

Call:
C5.0.formula(formula = bad_ind ~ ., data = train, rules = T, control = tc)


C5.0 [Release 2.07 GPL Edition]     Mon May 22 21:35:15 2017
-------------------------------

Class specified by attribute `outcome'

Read 3001 cases (19 attributes) from undefined.data

Rules:

Rule 1: (2161/262, lift 1.1)
    fico_score > 661
    ->  class 0  [0.878]

Rule 2: (2015/301, lift 1.1)
    tot_tr > 13
    ->  class 0  [0.850]

Rule 3: (2879/531, lift 1.0)
    rev_util <= 116
    ->  class 0  [0.815]

Rule 4: (26/5, lift 4.1)
    tot_tr <= 13
    rev_util > 116
    fico_score <= 661
    ->  class 1  [0.786]

Default class: 0


Evaluation on training data (3001 cases):

            Rules     
      ----------------
        No      Errors

         4  563(18.8%)   <<


       (a)   (b)    <-classified as
      ----  ----
      2417     5    (a): class 0
       558    21    (b): class 1


    Attribute usage:

     96.80% rev_util
     72.88% fico_score
     68.01% tot_tr


Time: 0.1 secs

参考资料：CDA《信用风险建模》微专业