0

I am trying to create partial dependence plots for my multinomial gbm predictions but I haven't been able to figure out how to produce the correct plots, the ones that I am getting have a single line instead of a line for every level of my response variable (in my case are 3 different species names). I have seen several examples but they require objects created with other packages (not gbm objects) and most of the examples don't include multinomial variables.

gbm fit

gbm.fit.final<-readRDS(file = "gbm_fit_final1_organism.rds")

getting table with variable importance

summary.gbm<-summary(
  gbm.fit.final, 
  cBars = 10,
  method = relative.influence, 
  las = 2)

The table looks like this:

          var   rel.inf
     MA0356.1 22.641689
     MA1071.1 21.707397
     MA0311.1 16.010605
     MA0210.1  7.249431
     MA0271.1  4.958186

I used the following code to generate the partial dependence plot for the most important predictor variable:

gbm.fit.final %>%
  partial(pred.var = "MA0356.1", n.trees = gbm.fit.final$n.trees, grid.resolution = 100, prob=T) %>%
  autoplot(rug = TRUE, train = motifs_train.100) +
  scale_y_continuous()

motifs_train.100 is the training data that I used to create the gbm fit (gbm.fit.final), I am not sure if it is necessary to add the training data.

I got the following plot:

plot with single line

I would like to get a plot like this one (I think I need to get marginal probabilities):

plot with a line for each level of response variable

I am very new to gbm package. I don't know if there is an argument of the function partial that I am omitting, or if there is a better function to do this. Any help would be appreciated. Thanks!

Carina Paola Cornejo Paramo
  • 1
  • 1
  • what stops you from getting the predictions for all your variables, and *then* plotting everything ? – RoB Mar 27 '20 at 14:13
  • Nothing, I want the plots for the most important variables, I could make plots for all of them (they are 100) but still I don't know how to plot the partial dependence including the classes of the response variables. This is just an example of what I want to to. – Carina Paola Cornejo Paramo Mar 29 '20 at 04:07
  • I mean, the predictor variables are not the issue, I want to include the response variable (species) – Carina Paola Cornejo Paramo Mar 29 '20 at 04:08

0 Answers0