Open see24 opened 6 years ago
Has variable importance for multiclass classification officially been implemented yet? If not, @see24 did you find any issues with the code which you posted on SO?
I did continue using this code and I published an article that used the by class variable importance. In this article the importance of different variables for different classes mostly aligned with our ecological expectations so I felt pretty good about it. However, nobody with in depth knowledge of the gbm package or the statistics behind it has reviewed the code as far as I know.
This is not possible in the old gbm but I recently wrote a function to calculate it. It can yield very different results for the most important variables for different classes. See this question and my answer for the code Hopefully you would just need to translate this from my dplyr/purrr version to base to implement it. If it doesn't make sense to do this this way please let me know!