重要性（随机森林）和 RandomForest$importance 之间的区别

Question

我不明白随机森林模型的重要性函数（randomForest 包）和重要性值之间有什么区别：

我计算了一个简单的 RF 分类模型并尝试通过以下代码找到变量重要性：

 rf_model$importance
         0               1      MeanDecreaseAccuracy    MeanDecreaseGini
 X1  0.096886458    0.032546101    0.055488009             2472.172207
 X2  0.030985037    0.025615202    0.027530078             1338.378297
 X3  0.124302743    0.012551971    0.052402188             3091.891586

importance(rf_model)
            0            1      MeanDecreaseAccuracy    MeanDecreaseGini
 X1 159.9149603    175.6265625        242.424683          2472.172207
 X2 104.8273654    97.09338154        129.5084398         1338.378297
 X3 157.0207876    86.93847182        216.6374153         3091.891586

为什么 MeanDecreaseGini 相同时输出的前三列有差异？

Answer 1

默认情况下调用 importance(rf_model) 时，测量值除以它们的“标准误差”。考虑这个例子：

library(randomForest)
set.seed(4543)
data(mtcars)
mtcars.rf <- randomForest(mpg ~ ., data=mtcars, ntree=1000,
                          keep.forest=FALSE, importance=TRUE)

mtcars.rf$importance
#output
        %IncMSE IncNodePurity
cyl   7.3939431     162.38777
disp 10.0468306     257.46627
hp    7.6801388     200.22729
drat  1.0921653      65.96165
wt    9.7998328     250.94940
qsec  0.6066792      38.52055
vs    0.7048540      24.75183
am    0.6201962      17.27180
gear  0.4110634      16.33811
carb  1.0549523      27.47096

同上

importance(mtcars.rf, scale = FALSE)
        %IncMSE IncNodePurity
cyl   7.3939431     162.38777
disp 10.0468306     257.46627
hp    7.6801388     200.22729
drat  1.0921653      65.96165
wt    9.7998328     250.94940
qsec  0.6066792      38.52055
vs    0.7048540      24.75183
am    0.6201962      17.27180
gear  0.4110634      16.33811
carb  1.0549523      27.47096

default: 
importance(mtcars.rf)
       %IncMSE IncNodePurity
cyl  15.767986     162.38777
disp 19.885128     257.46627
hp   18.177916     200.22729
drat  7.002942      65.96165
wt   18.479239     250.94940
qsec  5.022593      38.52055
vs    4.427525      24.75183
am    6.435329      17.27180
gear  3.968845      16.33811
carb  8.207903      27.47096

最后：

importance(mtcars.rf, scale = FALSE)[,1]/mtcars.rf$importanceSD
      cyl      disp        hp      drat        wt      qsec        vs        am      gear      carb 
15.767986 19.885128 18.177916  7.002942 18.479239  5.022593  4.427525  6.435329  3.968845  8.207903

等同于importance(mtcars.rf)[,1]

all.equal(importance(mtcars.rf, scale = FALSE)[,1]/mtcars.rf$importanceSD,
          importance(mtcars.rf)[,1])
#output
TRUE

重要性（随机森林）和 RandomForest$importance 之间的区别

Difference between Importance(random forest) and RandomForest$importance

r

random-forest