birankr: Ranking Nodes in Bipartite and Weighted Networks

Documented in bipartite_rank

#' Bipartite Ranks
#' @description Estimate bipartite ranks (centrality scores) of nodes from an edge list or adjacency matrix. Functions as a wrapper for estimating rank based on a number of normalizers (algorithms) including HITS, CoHITS, BGRM, and BiRank. Returns a vector of ranks or (optionally) a list containing a vector for each mode. 
#'
#' @details If input data is an edge list, this function returns ranks ordered by the unique values in the supplied edge list. Data inputted as an edge list are always assumed to contain named vertex IDs rather than to reflect an index of vertex positions in a network matrix. Users who wish for their edge lists to reflect vertex indices are recommended to input their data as a matrix or as a sparse matrix. \cr \cr
#' Network isolates are assigned a value of \eqn{(1 - alpha) / (n\_columns)} or \eqn{(1 - beta) / (n\_rows)} depending on their mode in the network. These values will always be smaller than the minimum value assigned to non-isolated nodes in the given mode. However, estimates on network isolates are non-meaningful. Users are advised to treat isolates with caution. \cr \cr
#' For information about the different normalizers available in this function, see the descriptions for the HITS, CoHITS, BGRM, and BiRank functions. However, below outlines the key differences between the normalizers, with \eqn{K_d} and \eqn{K_p} representing diagonal matrices with generalized degrees (sum of the edge weights) on the diagonal (e.g. \eqn{(K_d)_{ii} = \sum_j w_{ij}} and \eqn{(K_p)_{jj} = \sum_i w_{ij}}).
#'\tabular{lll}{
#'   \strong{Transition matrix} \tab \strong{\eqn{S_p}} \tab \strong{\eqn{S_d}} \cr
#'           --------------------- \tab --------------------- \tab --------------------- \cr
#'   HITS \tab \eqn{W^T} \tab \eqn{W} \cr
#'   Co-HITS \tab \eqn{W^T K_d^{-1}} \tab \eqn{W K_p^{-1}} \cr
#'   BGRM \tab \eqn{K_p^{-1} W^T K_d^{-1}} \tab \eqn{K_d^{-1} W K_p^{-1}} \cr
#'   BiRank \tab \eqn{K_p^{-1/2} W^T K_d^{-1/2}} \tab \eqn{K_d^{-1/2} W K_p^{-1/2}}
#'}
#'
#' @md
#' @param data Data to use for estimating rank. Must contain bipartite graph data, either formatted as an edge list (class data.frame, data.table, or tibble (tbl_df)) or as an adjacency matrix (class matrix or dgCMatrix).
#' @param sender_name Name of sender column. Parameter ignored if data is an adjacency matrix. Defaults to first column of edge list.
#' @param receiver_name Name of sender column. Parameter ignored if data is an adjacency matrix. Defaults to the second column of edge list.
#' @param weight_name Name of edge weights. Parameter ignored if data is an adjacency matrix. Defaults to edge weights = 1.
#' @param rm_weights Removes edge weights from graph object before estimating rank. Parameter ignored if data is an edge list. Defaults to FALSE.
#' @param duplicates How to treat duplicate edges if any in data. Parameter ignored if data is an adjacency matrix. If option "add" is selected, duplicated edges and corresponding edge weights are collapsed via addition. Otherwise, duplicated edges are removed and only the first instance of a duplicated edge is used. Defaults to "add".
#' @param normalizer Normalizer (algorithm) used for estimating node ranks (centrality scores). Options include HITS, CoHITS, BGRM, and BiRank. Defaults to HITS.
#' @param return_mode Mode for which to return ranks. Defaults to "rows" (the first column of an edge list).
#' @param return_data_frame Return results as a data frame with node names in the first column and ranks in the second column. If set to FALSE, the function just returns a named vector of ranks. Defaults to TRUE.
#' @param alpha Dampening factor for first mode of data. Defaults to 0.85.
#' @param beta Dampening factor for second mode of data. Defaults to 0.85.
#' @param max_iter Maximum number of iterations to run before model fails to converge. Defaults to 200.
#' @param tol Maximum tolerance of model convergence. Defaults to 1.0e-4.
#' @param verbose Show the progress of this function. Defaults to FALSE.
#' @return A dataframe containing each node name and node rank. If return_data_frame changed to FALSE or input data is classed as an adjacency matrix, returns a vector of node ranks. Does not return node ranks for isolates.
#' @keywords Bipartite rank centrality HITS CoHITS BGRM BiRank
#' @export
#' @import Matrix data.table
#' @examples
#' #create edge list between patients and providers
#'     df <- data.table(
#'       patient_id = sample(x = 1:10000, size = 10000, replace = TRUE),
#'       provider_id = sample(x = 1:5000, size = 10000, replace = TRUE)
#'     )
#'
#' #estimate CoHITS ranks
#'     CoHITS <- bipartite_rank(data = df, normalizer = "CoHITS")

bipartite_rank <- function(
  data,
  sender_name = NULL,
  receiver_name = NULL,
  weight_name = NULL,
  rm_weights= FALSE,
  duplicates = c("add", "remove"),
  normalizer = c('HITS','CoHITS','BGRM','BiRank'),
  return_mode = c("rows", "columns", "both"),
  return_data_frame = TRUE,
  alpha = 0.85,
  beta = 0.85,
  max_iter = 200,
  tol = 1.0e-4,
  verbose = FALSE
){
  #a) if weight name = "unweighted", change to NULL
      if(!is.null(weight_name)){
          if(weight_name == "unweighted"){
            weight_name <- NULL
          }
      }

  #b) convert to sparse matrix if a dataframe or matrix
      if(any(class(data) == "data.frame")){
        data <- data.table(data)
        if(verbose) message("Converting to sparse matrix...")
        adj_mat <- sparsematrix_from_edgelist(
          data = data,
          sender_name = sender_name,
          receiver_name = receiver_name,
          weight_name = weight_name,
          duplicates = duplicates[1]
        )
      }else if(length(class(data)) == 1 & class(data) == "matrix"){
        adj_mat <- sparsematrix_from_matrix(data)
      }else if(class(data) == "dgCMatrix"){
        adj_mat <- data
      }
      else if(class(data) != "dgCMatrix"){
        stop('data is not a data.frame, tbl_df, data.table, matrix, or dgCMatrix')
      }

  #d) remove weights
      if(rm_weights){
          if(verbose) message("Removing edge weights...")
          adj_mat <- sparsematrix_rm_weights(adj_mat)
      }

  #e) estimate bipartite rank
      if(verbose) message("Estimating bipartite rank...")
      rank <- bipartite_pagerank_from_matrix(
        adj_mat = adj_mat,
        normalizer = normalizer[1],
        alpha = alpha,
        beta = beta,
        max_iter = max_iter,
        tol = tol,
        verbose = verbose,
        return_mode = return_mode[1]
      )

  #f) find rank labels
      #i) get labels if data is data frame
          if(any(class(data) == "data.frame")){
            #1) determine ID index
                if(is.null(sender_name) | is.null(receiver_name)){
                    id1 = 1
                    id2 = 2
                }else{
                    id1 = match(sender_name, names(data))
                    id2 = match(receiver_name, names(data))
                }
            #2) pull IDs for each mode in order of function
                edges <- data[, c(id1, id2), with = F]
                id_names1 <- as.character(unlist(unique(data[, id1, with = F])))
                id_names2 <- as.character(unlist(unique(data[, id2, with = F])))
          }
      #ii) get labels if data is matrix
          if(!any(class(data) == "data.frame")){
              return(rank)
          }

  #g) format results
      #i) get variable name id of a data.frame
          if(any(class(data) == "data.frame")){
            sender_name <- names(edges)[1]
            receiver_name <- names(edges)[2]
          } else{
            sender_name <- "ID"
            receiver_name <- "ID"
          }

      #ii) if return a data frame, format results as a dataframe
          if(return_data_frame){
            if(return_mode[1] == "rows"){
              rank <- data.frame(ID = id_names1, rank = rank)
              names(rank)[1] <- sender_name
            }
            if(return_mode[1] == "columns"){
              rank <- data.frame(ID = id_names2, rank = rank)
              names(rank)[1] <- receiver_name
            }
            if(return_mode[1] == "both"){
              rank <- list(
                rows = data.frame(ID = id_names1, rank = rank[[1]]),
                columns = data.frame(ID = id_names2, rank = rank[[2]])
              )
              names(rank$rows)[1] <- sender_name
              names(rank$columns)[1] <- receiver_name
            }
          }
      #iii) if not return data frame, format results as vector
          if(!return_data_frame){
            if(return_mode[1] == "rows"){
              names(rank) <- id_names1
            }
            if(return_mode[1] == "columns"){
              names(rank) <- id_names2
            }
            if(return_mode[1] == "both"){
              names(rank[[1]]) <- id_names1
              names(rank[[2]]) <- id_names2
            }
          }

  #h) return data
      return(rank)

}