mixOmics splsda - examine variable contributions across CV folds

CV_contrib.R

getBugsCV <- function(x, y, ncomp = 5, nfolds = 5, nvar = 5, direction = ">", seed = 42, nreps = 1){
  set.seed(seed)
  sigbugs <- list(nfolds*nreps)
  for(rep in 1:nreps){
    folds <- sample.int(nfolds, size = nrow(x), replace = TRUE)
    for(fold in 1:nfolds){
      train <- folds != fold
      test <- folds == fold
      fit <- splsda(x[train,], y[train], keepX = rep(nvar,ncomp), ncomp = ncomp)
      sigbugs[[(rep-1)*nfolds + fold]] <- names(fit$loadings$X[,1])[fit$loadings$X[,1] != 0]
    }
  }
  sigbugs
}

countlist <- function(list){
  tab <- list %>% unlist %>% table
  data.frame(counts = tab, stringsAsFactors = FALSE) %>% setNames(c("OTU", "freq")) %>% mutate(OTU = as.character(OTU), freq = freq/length(list)) %>% arrange(desc(freq))
}

buglist <- getBugsCV(x, y, ncomp = 1, nfolds = 5, nvar = 4, nreps = 20, seed = 100)
overview <- countlist(buglist) %>% mutate(type = as.vector(tax_table(xt1mlda)[OTU,"lowest"])) %>% mutate(label = paste(OTU, type) %>% gsub("OTU", "", .)) %>% mutate( label = factor(label, levels = rev(label)))
overview
ggplot(overview, aes(x = label, y = freq, group = 1)) + geom_point() + geom_line() + coord_flip() + scale_y_continuous(labels = scales::percent) + theme_bw() + xlab("Contributors") + ylab("Frequency present")