Fix ranger.unify for probability forests

[ck37]

This is a minor fix to allow ranger.unify() to work correctly when probability = TRUE in a ranger::ranger() model, i.e. a probability forest.

This example code will now work correctly:

data = fifa20$data[, 1:8]
data$target_high = as.integer(fifa20$target > median(fifa20$target))
rf_model = ranger::ranger(target_high ~ ., data = data, num.trees = 10L, probability = TRUE)
unified = unify(rf_model, data)
treeshap1 = treeshap(unified, data[1:100, ], verbose = 0)
plot_feature_importance(treeshap1, max_vars = 6)

[mayer79]

This fix will work only if the response is 0-1 coded.

library(ranger)

fit <- ranger(Species ~ ., data = iris, probability = TRUE)
colnames(treeInfo(fit, 1))


# "nodeID"          "leftChild"       "rightChild"      "splitvarID"      "splitvarName"    
#"splitval"        "terminal"        "pred.setosa"     "pred.versicolor" "pred.virginica"
`` 

[ck37]

Agreed, but I don't believe this package handles categorical outcomes. Looking at the different unify functions it's expecting there to be a single prediction column.

[mayer79]

Good point...

[ck37]

@krzyzinskim Are you able to merge or propose any revisions?

[ck37]

@krzyzinskim Checking back on this PR - are you able to review?