2

Data is a subset of my original DF. Snip of it is attached

I use R, Party package in order to fit prediction model ("classifier") for

"Converted.clicks" as response variable.

The rest of vars are used as explaining variables in the model.

Here is the relevant part of my code:

table(DF$Converted.clicks)

"0" = 31456              
"1" =  39  
"2" =  6


Formula<-Converted.clicks ~ Day.of.week 
                          + Device
                          + Keyword 
                          + Quality.score
                          + Network..with.search.partners. 
                          + Ad.group
                          + Match.type

ct<-ctree(Formula,data=DF) 

#######################################

Issue:

The Converted.clicks variable is highly imbalanced.The majority of the observations has

class "zero". So after ctree function is applied,all the predictions are "zero",there are

no classes "1" and "2" predicted.

My questions are:

  1. Is the classifier Decision Tree model is appropriate model to predict

    as.factor(DF$Converted.clicks)?

  2. If so, how can I balance the response var (i.e.to give the chance the two rest classes

    "1" and "2" to be predicted?) - if I need to use weights, I need an

    example,please.

  3. Is there any other appropriate model to predict # of Converted.clicks? I understand

    that Regression Decision Tree is only for continuous response variable, but in my case

    I have an integer response var, please advise.

Alex
  • 21
  • 3
  • 1- There is no `Ctree` package. You are using the `party` package. 2- Clicks is a integer so I'm not sure why are you converting it to a categorical variable. 3- I don't think that `Cost...converted.click` is independant variable. It is actually a dependant variable of clicks. – David Arenburg Feb 08 '15 at 11:27
  • I haven't used the `party` package. `rpart` has a simple cost function...as does `C5.0`. Would be surprised if there isn't a cost function hidden there somewhere, but if you have got lots of data could always downsample. Some might advocate using probabilities from `ctree` -not sure if that makes sense (ESL 9.2.5) – charles Feb 09 '15 at 01:46
  • I have posted a related answer here: http://stats.stackexchange.com/questions/28029/training-a-decision-tree-against-unbalanced-data – rf7 Apr 07 '17 at 06:41

0 Answers0