reweight: Reweighting of Links from an Adjacency Weighted Matrix
In SyNet: Inference and Analysis of Sympatry Networks

Description Usage Arguments Details Value Author(s) References Examples

Transforms a weighted matrix into another one constrained into a common scale of relative strength. Values are reweighted by a combination of rewards and punishments. For a given link, its relative strength increases with the number of triads where it is the strongest and decreases with the number of triads where it is the weakest.

1 2	reweight(wm, similarity = TRUE, t1t2 = if(similarity) c(0, 1) else c(Inf, 0), normalized = TRUE)

`wm`	Square and symmetric matrix filled with valued score of spatial affinity.
`similarity`	Logical, depending on the existence of direct or inverse relationship between matrix scores and the spatial affinity between species (i.e. TRUE: high scores - high affinity; FALSE: low scores - high affinity).
`t1t2`	Non-negative numeric vector, conventionally of length 2.
`normalized`	Logical, if TRUE output values are linearly transformed into the scale [0, 1].

The vector t1t2 holds the thresholds to delimit intervals for meaningful association. If similarity is TRUE, any value of the original matrix lower than min(t1t2) is discarded for further consideration because it represents the lower bound for meaningful association. On the other hand, any value of the input matrix equal or higher than the upper bound (i.e. max(t1t2)) is deemed to be maximally similar and coded 1 in the output. If similarity is FALSE, any link lower than min(t1t2) represent close affinity between species, and conversely any link equal or higher than max(t1t2) denote no affinity at all between its endvertices.

The dichotomization is a common procedure to transform weighted matrices into binary ones. Values are set 1 or 0 depending on the location of some cutoff value. This function also enables us to perform a hard thresholding when the parameter t1t2 has been fed with a single value. Under this setting, the final reweighted matrix will be binary. If similarity is TRUE (FALSE), entries are 1 if original values are equal or higher (lower) than t1t2, otherwise entries are 0.

A similarity matrix composed of reweighted values. Values are constrained into the interval [0, 1] under the option of normalization. Here, null association is coded with a zero (0), whereas the unity (1) corresponds to the optimal meaningful association.

Daniel A. Dos Santos <dadossantos@csnat.unt.edu.ar>

Dos Santos D.A., Cuezzo M.G., Reynaga M.C., Dominguez E. 2011. Towards a Dynamic Analysis of Weighted Networks in Biogeography. Systematic Biology (in press).

  #Example to show an interesting numerical property of the reweighting formula, i.e.
  #its convergence to the same rescaled matrix after several self-iterations
  nitems <- 25
  ej <- matrix(0, nitems, nitems)
  ej[row(ej) > col(ej)] <- runif(nitems*(nitems - 1)/2)
  t(ej) + ej -> ej
  diag(ej) <- 1
  ### Display three graphics
  op <- par(mar = rep(3, 4), mfrow = c(3, 1))
  plot(unlist(ej),unlist(reweight(ej)), xlab = "Input matrix", ylab = "Reweighted matrix", main = "First reweighting")
  #Iterative reweighting
  histcor <- c()
  #Perform 100 iterations
  for(i in 1:100) {
    reweight(ej) -> rej
    b <- c()
    for(k in 1:(nrow(ej) - 1))
      for(j in (k + 1):nrow(ej)) {
        a <- (ej[k, j] >= ej[k, ]) + (rej[k, j] >= rej[k, ])
        b <- c(b, sum(a==2)/sum(a > 0))
    }
    histcor <- c(histcor, mean(b))
    ej <- rej
  }
  plot(unlist(ej),unlist(reweight(ej)), xlab = "Input matrix", ylab = "Reweighted matrix",
       main = "Reweighting after several iterations")
  plot(histcor, xlab = "Iteration", ylab = "Cross adjustment",
       main = "Evolution of the resemblance between input and reweigthed network")
  ## At end of plotting, reset to previous settings:
  par(op)