2
$\begingroup$

I am using CART classification technique by dividing a dataset into train and test sets. I have been using Mis-classification error, KS by rank ordering, AUC and Gini as MPMs(model performance measures). The problem I am facing is that the MPM values are quite far apart.

I have tried with minsplit equal to anywhere from 20 to 1400 and minbucket from 5 to 100 but couldn't get expected results. I have also tried oversampling/undersampling through ROSE package but without any improvement. Moreover, the mis-classification error increased a lot. Following code is through which I could get the best values, but they were not enough.

#Reading Data
pdata = read.csv("PL_XSELL.csv", header = TRUE)

#Converting ACC_OP_DATE from type factor to date
pdata$ACC_OP_DATE<-as.Date(pdata$ACC_OP_DATE, format = "%d-%m-%Y")

#Paritioning the data into training and test dataset
set.seed(2000)
n=nrow(pdata)
split= sample(c(TRUE, FALSE), n, replace=TRUE, prob=c(0.70, 0.30))
ptrain = pdata[split, ]
ptest = pdata[!split,]

#CART Model
#Taking the minsplit, minbucket values as low as possible, so that pruning 
#can be done later. Higher values didn't allow any scope for pruning
r.ctrl = rpart.control(minsplit=20, minbucket = 5,  cp = 0, xval = 10)

#Calling the rpart function to build the tree
cartModel <- rpart(formula = TARGET ~ ., 
    data = ptrain[,-1], method = "class", 
    control = r.ctrl)

#Pruning Tree Code
cartModel<- prune(cartModel, cp= 0.00225  ,"CP")

#Predicting class and scores
ptrain$predict.class <- predict(cartModel, ptrain, type="class")
ptrain$predict.score <- predict(cartModel, ptrain, type="prob")

Results that I got-: Train data Mis-classification error-.103 AUC - 0.679 KS - 0.259 Gini - 0.313

Test data Mis-classification error-.113 AUC - 0.664 KS - 0.226 Gini - 0.307

Is it due to the dataset or am I doing something wrong. I am new to Data Analytics. It is a part of my academic project, so I need to use CART technique only. I will put separate questions for Random Forest and Neural Networks. Kindly help.

$\endgroup$
0
$\begingroup$

Your model suffers of a slight overfitting, however it doesn't seem too dramatic.

Performance on train set are always better than test set if random sampled (when you have statistically significant volumes)

Maybe you can reduce the gap of performance by controlling the CP parameter, try setting a higher cp when you prune the tree (like 0.01) or by using the parameter maxdepth that prune according to the length of the tree.

$\endgroup$
  • $\begingroup$ I will try that but increasing cp to prune the tree reduces KS to lower than 0.20, which is not desirable. Also I've heard that the MPMs should be within 10% of difference between the train and test sets. What is the acceptable difference to you? $\endgroup$ – Manu Vats Mar 25 at 12:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.