It is sometimes stated that one does not care about the ''raw'' values of the variable importance outcome:
for each feature X_j using a random forest and rather should just compare PI(X_j) w.r.t their relative size or rank them.
Is there any intuitiv explanation for this? Or is this simply not correct?
I am conducting a ''variable importance study'' on some data set using the R party package by fitting a cforest and calculating the permutation importance (with missing data) using the function: varimp() as described by:
Alexander Hapfelmeier, Torsten Hothorn, Kurt Ulm, and Carolin Strobl. A new variable importance measure for random forests with missing data. Statistics and Computing, 24(1):21–34, 2014.
The obtained results are rather small i.e. in the range of $[0,0.0008]$, nevertheless comparing the relative sizes I can clearly distinguish between less and more important variables.
Thanks in advance