run_svd.R

#' Run an SVD for collaborative filtering and process the results to be more tidyverse-friendly
#' @param ratingsmatrix A item-user review matrix
#' @param ndimensions the number of dimensions to use, defaults to 10
#' @return a list with the original u, d, and v matrices from the svd function and
#'         item_values - a long-format tibble with the values per item per dimension
#'         user_values - a long-format tibble with the values per user per dimension
#'         predictions - a long-format tibble with the predictions per user per item
#' @note (c) 2022 Wouter van Atteveldt, license: CC-0
run_svd = function(ratingsmatrix, ndimensions=10) {

  # Run SVD and set rownames on the resulting matrices
  udv = svd(ratingsmatrix, ndimensions, ndimensions)
  rownames(udv$u) = rownames(ratingsmatrix)
  rownames(udv$v) = colnames(ratingsmatrix)
  
  # Convert the u and v matrices to long-format tibbles
  udv$item_values = as_tibble(udv$v, rownames='item') |> 
      pivot_longer(-item, names_to="dimension", values_to="item_value")
  udv$user_values = as_tibble(udv$u, rownames='user') |> 
    pivot_longer(-user, names_to="dimension", values_to="user_value")
    
  # Compute the predictions by re-multiplying the result of the decomposition 
  # and convert to long-format tibble
  udv$d = udv$d[1:ndimensions]
  p = (udv$u %*% diag(udv$d) %*% t(udv$v)) 
  udv$predictions = as_tibble(p, rownames='user') |>
    pivot_longer(-user, names_to="item", values_to="prediction")
    
  return(udv)
}