2
$\begingroup$

My question is an extension to the question asked here. How does one identify the parity of predictor/feature/variable impact on response/outcome in a data mining model. Is there a standard procedure to find the 'direction' of impact after one does feature selection and derives variable importance using methods such as regularized random forest or lasso/elasticnet?

I know this question may sound quite naive, but I really wanted to know and I have searched SO and other materials but couldn't find a convincing answer.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$
  • I don't thing there is a standard approach.
  • "Boosted Regression Trees for ecological modeling" is a commonly cited reference that briefly discusses some of these issues in the context of boosting. Partial dependence plots are available in many packages.
  • Rminer is a package the uses sensitivity analysis to extract information from models. Underused and has the benefit that you can use almost any model with this methodology.
  • Soren Welling is an active member on this site - has authored the [forestFloor][2] package and goes into some depth into getting information from blackbox models in the following stackexchange answer: Getting Information out of Blackbox Models - RandomForest / XGBoost
  • I haven't looked at the ggforest package, but this also offers to open the black box.
Improve this answer
$\endgroup$
2
  • $\begingroup$ Thx for the referral :) - pkg is called forestFloor $\endgroup$
    – Soren Havelund Welling
    Commented Jan 3, 2016 at 15:06
  • $\begingroup$ aaaah. Sorry about that. $\endgroup$
    – charles
    Commented Jan 3, 2016 at 23:45

Your Answer

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

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