2
$\begingroup$

I'm using the good old decision tree (CRT or CHAID algorithm, depending on the situation) in order to predict voting behavior and extract some profile (e.g: Women who live in the suburbs, who are not married and have a high income are a target group for Democrats).

I'd like to do the same exercise with the results of a random forest model, but I have no idea how I can extract such multivariate profiles (IF female, AND IF suburbs, AND IF not married, AND IF high income, odds are high that one votes Democratic). I do know that you can list the general importance of each variabele, but that's about it. This variable importance is not informative on how a combination of certain values results in a certain outcome. The whole reason I'm using decision trees here and not logistic regression is precisely because I'm interested in how the combination of certain values leads to a certain outcome. And since there are - let's say - 500 trees, with in each of them a subset of variables and a subset of cases, it's useless to look at these trees separately in order to formulate such profiles.

So, here's my question: is there any way to summarize the results of a random forest in order to extract profiles based on multiple variables?

Thanks!

PS: It's my first question on StackExchange, I hope I'm doing this right!

$\endgroup$
3
  • $\begingroup$ Try this: stats.stackexchange.com/questions/172163/… $\endgroup$ Nov 9, 2017 at 8:14
  • $\begingroup$ @KBoghe, it is a common question, and the common answer is that RF is rather a black box than interpretable model. I however worked on extracting knowledge from it analyzing input variables one-by-one. It was implemented by assigning mean values to all input terms except one and then building a scatter plot of VS. RF output. You can also try to turn inputs into categorical variables and analyze how RF output differs across the combination of the most important categorical inputs. $\endgroup$ Nov 9, 2017 at 9:38
  • $\begingroup$ @KarelMacek Thanks for the answers, guys. I'll be able to work something out! $\endgroup$
    – KBoghe
    Nov 10, 2017 at 10:59

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.