protection stack overflow about ranger

[mnwright]

This is probably due to the formula interface. Try to use the alternative interface with dependent.variable.name.

[shihuang047]

I tried dependent.variable.name="y", it is still not working for me...

[mnwright]

With the same error message? That should only occur with the formula interface. Please give a reproducible example or at least your ranger call and some information to the data.

[shihuang047]

I just solved the problem by converting the matrix into sparse matrix. But I also encountered a similar problem when using importance_pvalue(). It only allows data.frame/formula as inputs. Is it possible to make it also compatible with the sparse matrix?

The codes I used:
x <- data.frame(rbind(t(rmultinom(7000, 75000, c(.201,.5,.02,.18,.099))),
t(rmultinom(8000, 75000, c(.201,.4,.12,.18,.099))),
t(rmultinom(15000, 75000, c(.011,.3,.22,.18,.289))),
t(rmultinom(15000, 75000, c(.091,.2,.32,.18,.209))),
t(rmultinom(15000, 75000, c(.001,.1,.42,.18,.299)))))
y <- factor(c(rep("A", 15000), rep("B", 15000), rep("C", 15000), rep("D", 15000)))
data<-data.frame(y, x)
sparse_data <- Matrix(data.matrix(data), sparse = TRUE)
rf.model <- ranger::ranger(dependent.variable.name="y", data=sparse_data, keep.inbag=TRUE, importance='permutation')

[mnwright]

The importance_pvalues() function only needs the formula/data when using the permutation approach ("altmann"). The idea was that this method is so slow that no one wants to run it on data that's to large for the formula interface. I will check whether this is possible with sparse data.

In the meantime, the permutation p-values are so simple, you can just do it yourself:

library(ranger)

num.permutations <- 100

# Run RF
rf <- ranger(dependent.variable.name = "Species", data = iris, importance = "permutation")

# Permute and compute importance again (be sure to use same parameters as above)
vimp <- replicate(num.permutations, {
  dat <- iris
  dat[, "Species"] <- dat[sample(nrow(dat)), "Species"]
  ranger(dependent.variable.name = "Species", data = dat, importance = "permutation")$variable.importance
})

# Compute p-values
pval <- sapply(1:nrow(vimp), function(i) {
  (sum(vimp[i, ] >= rf$variable.importance[i]) + 1)/(ncol(vimp) + 1)
})

res <- cbind(rf$variable.importance, pval)
colnames(res) <- c("importance", "pvalue")

res

(this is just copy&paste from the importance_pvalues() function)

[shihuang047]

Thanks so much!