CODEWITHSUNDEEP
rrandom-forest
pkumar0
pkumar0

Reputation: 2239

R randomForest importance

It seems that in R, if we fit a randomForest and try to figure out importance, we get different results depending on if we use $importance or the function importance for %IncMSE. Not sure why?

> rf.model = randomForest(Weight ~ ., data = XYWeightTrain, importance = TRUE, ntree = 500, xtest = XYWeightTest[, -1], ytest = XYWeightTest[, 1])
> rf.model$importance
                       %IncMSE IncNodePurity
Wrist.Diam           0.8212594      305.3484
Wrist.Girth          2.8674595     1244.2349
Forearm.Girth       14.7491374     6681.7611
Elbow.Diam           1.0207908      427.8362
Bicep.Girth          7.9351242     4636.7848
Shoulder.Girth       9.5574023     5108.2292
Biacromial.Diam      0.9785278      347.9064
Chest.Depth          2.0081676      873.7349
Chest.Diam           1.9936859     1330.1593
Chest.Girth         24.2663570     9815.0322
Navel.Girth          2.1440752      648.8285
Waist.Girth         31.6001879    12512.4992
Pelvic.Breadth       0.5893632      227.1361
Bitrochanteric.Diam  1.1661954      346.4844
Hip.Girth            8.0548212     2178.3831
Thigh.Girth          2.8990134      726.3200
Knee.Diam            3.6684350     1207.6730
Knee.Girth           6.3831885     2258.2849
Calf.Girth           5.6392469     1972.3754
Ankle.Diam           0.7002560      199.7919
Ankle.Girth          2.0712253      684.4244
> importance(rf.model)
                      %IncMSE IncNodePurity
Wrist.Diam           7.541535      305.3484
Wrist.Girth          9.240727     1244.2349
Forearm.Girth       12.534953     6681.7611
Elbow.Diam           8.742194      427.8362
Bicep.Girth          9.966211     4636.7848
Shoulder.Girth      11.263877     5108.2292
Biacromial.Diam      6.680291      347.9064
Chest.Depth         10.196696      873.7349
Chest.Diam           6.846195     1330.1593
Chest.Girth         15.979216     9815.0322
Navel.Girth         12.194066      648.8285
Waist.Girth         20.320096    12512.4992
Pelvic.Breadth       6.575887      227.1361
Bitrochanteric.Diam  9.568542      346.4844
Hip.Girth           20.481270     2178.3831
Thigh.Girth         15.100160      726.3200
Knee.Diam           16.784600     1207.6730
Knee.Girth          19.353398     2258.2849
Calf.Girth          18.927534     1972.3754
Ankle.Diam           6.360296      199.7919
Ankle.Girth          9.868660      684.4244

Upvotes: 2

Views: 1906

Answers (1)

Jeffrey Evans
Jeffrey Evans

Reputation: 2397

If you just print the importance object from the model they are the raw importance values. However, when you use the importance function, the default for the scale argument is TRUE which returns the importance values divided by the standard error. If you use importance(rf.model scale=FALSE) the values should be the same.

I would highly recommend using the %IncMSE and not the GINI (IncNodePurity). The %IncMSE is permuted, at the nodes, and is a more stable representation of variable importance.

Upvotes: 4

Related Questions