Retrieving individual predictions for each tree about ranger

[mnwright]

Added for classification and regression. Is this also needed for probability prediction and survival?

[mllg]

Thanks! I personally need it only for regression at the moment, but I guess it would be consistent to have this for all forests.

[krz]

this may be related, so I'm posting it here. The out of bag predictions in the ranger class object $predictions, are these also majority votes (for classification) or averages (for regression) ?
I would be interested in all oob predictions from all trees.

[mnwright]

Sorry for the delay. Yes these are also majority votes or averages. I will try to set the predict.all option for growing, too.

[ajai-sharma-backup]

Is it possible for the predict method to predict with a fewer number of trees than the original model? This may be useful for model selection.

[mnwright]

You could use predict.all = TRUE and compute manual majority votes or averages on a subset of the trees.

[crarlus]

Hi, I was wondering if "predict.all" could work for probability trees as wel, i.e. a probability result per tree. This would be useful for statistical stability analysis via sub-sampling and aggregating results as well as for easy estimation of the optimal tree number (similar to the initial post). Thanks.

[mnwright]

Yes that's possible and the reason why this issue is still open. I'll definitely add it for probability prediction and survival.

[crarlus]

Ok, great.
I devised an interim solution by slicing the forest into smaller sub-forest and making individual predictions on these sub-forests. One can then aggregate a number of sub-predictions for robustness studies etc.

A quick implementation reads as follows

# the original forest is called "rf"

slice.size <- 10 # size of sub-forests
slice.number <- ceiling(rf$num.trees/slice.size) # number of sub-forests

rf_slice <- list()
for(slice.index in 1:slice.number) {

  print(paste("Predicting slice",slice.index,"of",length(slice.number)))

  From <- (slice.index - 1 )* slice.size + 1
  To <- min(slice.index * slice.size,rf$num.trees)

  rf_temp <- rf # copy forest object

# extract "slice.size" trees from original forest
  rf_temp$num.trees <- slice.size 
  rf_temp$forest$child.nodeIDs <- rf$forest$child.nodeIDs[From:To]
  rf_temp$forest$split.values <- rf$forest$split.values[From:To]
  rf_temp$forest$split.varIDs <- rf$forest$split.varIDs[From:To]
  rf_temp$forest$terminal.class.counts <- rf$forest$terminal.class.counts[From:To]
  rf_temp$forest$num.trees <- slice.size 

  rf_slice[[slice.index]] <- rf_temp

}

# Prediction

  Prediction_slice <- list()
  for(slice in 1:length(rf_slice)) {

    print(paste("Predicting slice",slice,"of",length(rf_slice)))
    Prediction_temp <- predict(object = rf_slice[[slice]], data = testFeatures, num.threads = 1, predict.all = F)$prediction[,2] # extracts only probability of second class
    Prediction_slice[[slice]] <- Prediction_temp 
  }
  Prediction_slice <- do.call(cbind,Prediction_slice) # results in a matrix of predictions

I opted for a slice.size of 10, i.e. each of the sub-forests contains 10 trees. Note that this piece of code only extracts the probabilities of the second class - however extension to more classes is straightforward.

Marvin, do you think this is a possible (quick) solution?

[mnwright]

Done.