It seems to be a well-established phenomenon that variable importance measures based on "increased node impurity" tend to be biased in favor of categorical variables with relatively many categories, and continuous variables in general. See for instance Strobl et al. 2007: https://doi.org/10.1186/1471-2105-8-25. This seems intuitive, as more categories and continuous variables allow for a more "fine-grained" partition of observations.

However, I am wondering whether this mechanism also would apply between continuous variables of different precision in their measurements? Or does the fact that all continuous variables may be made infinitely precise cause them all to compete on "fair grounds"?



Your Answer

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

Browse other questions tagged or ask your own question.