1
$\begingroup$

For example, I got an address feature like '32, 5th Avenue', and I could map it to a new feature as '5th Avenue' by only considering about the street part. The latter one will have far less possible values than the origin one. My question is will it perform better in a classification model, especially the tree models [including id3,rf,gdbt,etc.], and why this works/does not work.

Thanks for helping.

$\endgroup$
2
  • 3
    $\begingroup$ This question has been downvoted, but I see it as a great opportunity to answer about splitting factors without losing information and how classification models deal with nested features (e.g. about splitting street address in street and number). $\endgroup$
    – Pere
    Mar 25 '17 at 8:48
  • $\begingroup$ It is not possible to answer this because it is unclear what you are asking that is related to this site's topic. $\endgroup$ Mar 25 '17 at 13:08
2
$\begingroup$

Letting a feature to have more possible values results in more flexible models. In other words, it can find more subtle regularities in the domain. But, you should be cautious of the overfitting problem. If you have not enough training data, the more flexible model can overfit your training data rather than finding the true regularities in the domain. But if you have enough training data, having features with more possible values lets to find more interesting subtle regularities in the domain.

$\endgroup$

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.