mayer79
commented
10 months ago When the number of trees plays a role, then maybe your system goes out of memory during the predict() call of ranger.
Random forest predictions are extremely expensive, which is a problem for expensive model-agnostic XAI methods. When accuracy matters, I usually prefer well-tuned boosted trees models over random forests.
When it is no option to increase RAM or use the smaller (still very large) random forest, you can test with {treeshap} instead of {kernelshap}. Would be interesting to see if the problem persists with TreeSHAP.
quantumlinguist
I'm trying to calculate SHAP-values of random forests from the ranger package. The random forests have between 1000 and 8000 trees and the datasets between 5000 and 37000 rows. The only model I managed to calculate SHAP-values for is a model that had 850 trees. The rest all crashed. I find it odd that none of the tutorials or vignettes says this could happen so I am wondering if there is something wrong with my models or whether this is expected. I cannot post the data for privacy issues (government data) but the variables are all numerical except one binary variable. One group of datasets have 8 variables (there are datasets for different years) and the other 32. The dependent variable is a 3-level factor. I'm using a 32GB Macbook pro and R 4.1.2.
This is the code I am using:
ntree=850
set.seed(199) model <- ranger(Satisfaction ~ ., data = df20, importance = 'permutation', num.trees = ntree20, mtry=3, probability = TRUE, case.weights = class_weights[df20$Satisfaction])
set.seed(199) xvars=colnames(df20[-9]) X= df20[sample(nrow(df20),500), xvars]
bg_X= df20[sample(nrow(df20),200),]
shap_rf= kernelshap(model, X, bg_X = bg_X)