R/elbow.R

In celda: CEllular Latent Dirichlet Allocation

# https://stackoverflow.com/questions/35194048/using-r-how-to-calculate
#-the-distance-from-one-point-to-a-line
# http://mathworld.wolfram.com/Point-LineDistance2-Dimensional.html
# Kimberling, C. "Triangle Centers and Central Triangles." Congr.
# Numer. 129, 1-295, 1998.
.dist2d <- function(a, b, c) {
  v1 <- b - c
  v2 <- a - b
  m <- cbind(v1, v2)
  d <- abs(det(m)) / sqrt(sum(v1 * v1))
  return(d)
}


.secondDerivativeEstimate <- function(v) {
  nv <- length(v)
  res <- rep(NA, nv)
  for (i in seq(2, nv - 1)) {
    res[i] <- v[i + 1] + v[i - 1] - (2 * v[i])
  }
  return(res)
}


.curveElbow <- function(var, perplexity, pvalCutoff = 0.05) {
  len <- length(perplexity)
  a <- c(var[1], perplexity[1])
  b <- c(var[len], perplexity[len])
  res <- rep(NA, len)
  for (i in seq_along(var)) {
    res[i] <- .dist2d(c(var[i], perplexity[i]), a, b)
  }
  elbow <- which.max(res)
  ix <- var > var[elbow]
  perplexitySde <- .secondDerivativeEstimate(perplexity)
  perplexitySdeSd <- stats::sd(perplexitySde[ix], na.rm = TRUE)
  perplexitySdeMean <- mean(perplexitySde[ix], na.rm = TRUE)
  perplexitySdePval <-
    stats::pnorm(perplexitySde,
      mean = perplexitySdeMean,
      sd = perplexitySdeSd,
      lower.tail = FALSE
    )
  # other <- which(ix & perplexitySdePval < pvalCutoff)
  return(list(elbow = var[elbow]))
}