1
$\begingroup$

I am running a model for which I am getting a very bad percentage detection of events in the confusion matrix (basically my true positives). Obviously that implies my false negatives are too hight. When I gave these dataset to a neural network node or a decision tree node, I see that the percentage detection by these two techniques is quite good (65% - 67% vis a vis mine 15%). I went to the decision tree diagram and saw the various cuts under which it divides the population. I obviously understand that the variable which falls on the root of the tree has highest importance and the leafs have lowest.

  • How can the decision tree "tree" help me create categorical variables or treat continuous variables so that the accuracy of my model improves?

To clarify, if a decision tree can generate a matrix with 65% detection, it would have some rule inside it to get such accuracy. These rules would display in the tree diagram we get as output.

  • Can we use this tree to create out variables in a different way and get close to the accuracy given by decision tree?
$\endgroup$
2
$\begingroup$

I see that you've accepted answers to just 3 of 9 questions...

Are you using the type of Decision Tree known as CHAID? If so, you will obtain an indication of one main effect and then any number of so-called interaction effects. You can try these effects in a regression, ANOVA, or general linear model. You build in the main effect. You build in the main effect for each variable involved in an interaction. And you build in the interactions. But before you do this, the interaction effects all need to be pre-tested, as I explained in a comment to this post.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ thanks..Yes,I did that..only that improved my accuracy from 12- 15%..But what I am unable to fathom is how is decision tree getting it to 65%..It seems like however good interaction I may try, I would not cross 17%-18%..Can we do only so much as learning the interactions from the Decision Tree ?? No other usage ? $\endgroup$ – ayush biyani May 12 '11 at 7:09
  • 1
    $\begingroup$ 65% vs. 17% is indeed a perplexing difference. Maybe one of us could figure out more if you posted that tree, or at least a list of the main and interaction effects involved. Also, is there an element of crossvalidation to your project? $\endgroup$ – rolando2 May 12 '11 at 15:42

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.