R/ranM2.R

In shibiaowan/SHARP: hyperfast and accurate single-cell RNA-Seq data clustering

#' Prepare a random matrix
#'
#' This function is to prepare a random matrix to be used in parallel computation, so that the same random matrix will be applied to different partitions/groups of datasets in the same application of random projection. The difference between ranM and ranM2 is that the former requires the input expression matrix, whereas the latter only requires the number of genes.
#'
#' @param m number of genes for the input scRNA-seq matrix
#' @param p the dimension to be reduced to
#'
#' @import Matrix
#'
#' @export
ranM2 <- function(m, p, seedn) {
    # for random projection; note scdata: m*n, m is the feature dimensions, n is the
    # sample number; p is the reduced dimension
    
#     m = nrow(scdata)  #the number of features
#     n = ncol(scdata)  #number of samples/cells

    # set.seed(123)#a flag to fix the random number
    s = sqrt(m)  #according to the paper 'Very Sparse Random Projection'
    # x0 = sample(c(sqrt(s),0, -sqrt(s)), size= m*p, replace=TRUE, prob=c(1/(2*s), 1
    # - 1/s, 1/(2*s)))
    if (seedn%%1 == 0) {
        # integer
        set.seed(seedn)
        x0 = sample(c(sqrt(s), 0, -sqrt(s)), size = m * p, replace = TRUE, prob = c(1/(2 * 
            s), 1 - 1/s, 1/(2 * s)))
    } else {
        x0 = sample(c(sqrt(s), 0, -sqrt(s)), size = m * p, replace = TRUE, prob = c(1/(2 * 
            s), 1 - 1/s, 1/(2 * s)))
    }
    # return(x)
    x <- Matrix(x0, nrow = m, byrow = TRUE, sparse = TRUE)  #reshape a vector to a sparse matrix
    
    return(x)  #the same format, feature*sample
}