predict_parts survshap runs out of memory when p >= 17

[hbaniecki]

For p=17, this line produces a square matrix of dim [131072, 131072]

survex/R/surv_shap.R

Line 106 in 64e96d6

W <- diag(shap_kernel_weights)

See rnorm(131072*131072).

Source:

library(survival)
library(survex)
library(ranger)

N <- 100
P <- 17

X <- matrix(rnorm(N*P), ncol=P, nrow=N)
time <- rnorm(N)
status <- rbinom(N, 1, 0.5)
df <- cbind(data.frame(time=time, status=status), data.frame(X))

dim(df)

model <- ranger(Surv(time, status) ~ ., data = df, num.trees = 50)

explainer <- explain(model, data = df[, -c(1, 2)], y = Surv(df$time, df$status))

model_performance(explainer)

predict_parts(explainer, df[1, -c(1, 2)])
# Error: cannot allocate vector of size 128.0 Gb

[krzyzinskim]

Fixed with #26