glmnet simulation -- compare using pca on predictors vs not

glmnet-sim

library(doMC)
library(glmnet)
library(tictoc)
library(tidyr)
library(ggplot2)
library(PCAtools)

y <- rgamma(10000, shape = 2, scale = 2000)

x <- y - matrix(rnorm(10000000, mean = 3000, sd = 100000), nrow = 10000)

options(future.globals.maxSize = 891289600 * 4)

registerDoMC(cores = 5)
tic()
mod <- cv.glmnet(x, y, parallel = T, nfolds = 5, trace.it = T)
toc()

preds <- predict(mod, newx = x)[,1]

x_pca <- pca(x)
mod_pcomp <- cv.glmnet(data.matrix(x_pca$loadings)[,1:30], y, parallel = T, nfolds = 5, trace.it = T)
preds_pcomp <- predict(mod_pcomp, newx = data.matrix(x_pca$loadings[,1:30]))[,1]

data.frame(preds, preds_pcomp, y) %>% 
  gather(key = metric, value = value) %>% 
  ggplot(aes(x = value, fill = metric)) + 
  geom_density(alpha = 0.3)