Evaluation of sparse version of SuSiE

In susieR: Sum of Single Effects Linear Regression

knitr::opts_chunk$set(collapse = TRUE,comment = "#",fig.width = 4.5,
                      fig.height = 3,fig.align = "center",
                      fig.cap = " ",dpi = 120)

Set up environment

library(Matrix)
library(susieR)
set.seed(1)

Overview

In this vignette, we provide line profiles for revised version SuSiE, which allows for a sparse matrix structure. We compare speed performance when the form of the matrix X is dense and sparse.

In this minimal example, we observe that given a large sparse matrix, if it is in the dense form, the speed is around 40% slower than that in a sparse form.

Simulate data

We randomly simulate a n=1000 by p=1000 dense matrix and a sparse matrix at sparsity $99\%$, i.e. $99\%$ entries are zeros.

create_sparsity_mat = function(sparsity, n, p) {
  nonzero          <- round(n*p*(1-sparsity))
  nonzero.idx      <- sample(n*p, nonzero)
  mat              <- numeric(n*p)
  mat[nonzero.idx] <- 1
  mat              <- matrix(mat, nrow=n, ncol=p)
  return(mat)
}

n <- 1000
p <- 1000
beta <- rep(0,p)
beta[c(1,300,400,1000)] <- 10 
X.dense  <- create_sparsity_mat(0.99,n,p)
X.sparse <- as(X.dense,"CsparseMatrix")
y <- c(X.dense %*% beta + rnorm(n))

X in a dense form

susie.dense <- susie(X.dense,y)

X in a sparse form

susie.sparse <- susie(X.sparse,y)

Further step

We encourage people who are insterested in improving SuSiE can get insights from those line profiles provided.

susieR documentation built on March 7, 2023, 6:11 p.m.