Negative impurity importance values with splitrule = "poisson"

[cmcrowley]

Variable importance values are maximum of 0 when using impurity measure of importance with Poisson split rule. Possibly related to #764

data(mtcars)
rf <- ranger::ranger(gear ~ cyl + mpg + disp + hp + drat + wt, 
splitrule = "poisson", poisson.tau = 1, 
importance = "impurity", 
data = mtcars) 

rf$variable.importance
       cyl        mpg       disp         hp       drat         wt 
 -44.54971 -127.57739 -151.34939 -143.82199 -243.64165 -162.35374 

Permutation importance measures are positive and appear to be on a different scale. Perhaps one computation is dealing with (minimization of a negative) log likelihood and one is doing something else?

> rf <- ranger::ranger(gear ~ cyl + mpg + disp + hp + drat + wt, 
+                      splitrule = "poisson", poisson.tau = 1, 
+                      importance = "permutation", 
+                      data = mtcars) 
> rf$variable.importance
       cyl        mpg       disp         hp       drat         wt 
0.02435822 0.03311226 0.12919229 0.05829103 0.29430501 0.15226425