dissimilarity

0th

Percentile

Dissimilarities and Correlations Between Seriation Orders

Calculates dissimilarities/correlations between seriation orders in a list.

Keywords
cluster
Usage
ser_cor(x, y = NULL, method = "spearman", reverse = TRUE, test = FALSE)
ser_dist(x, y = NULL, method = "spearman", reverse = TRUE, ...)
ser_align(x, method = "spearman")
Arguments
x
set of seriation orders as a list with elements which can be coerced into ser_permutation_vector objects.
y
if not NULL then a single seriation order can be specified. In this case x has to be a single seriation order and not a list.
method
a character string with the name of the used measure. Available measures are: "kendall", "spearman", "manhattan", "euclidean", "hamming", "ppc" (positional proximity coefficient), and "aprd" (absolute pairwise rank differences).
reverse
a logical indicating if the orders should also be checked in reverse order and the best value (highest correlation, lowest distance) is reported. This only affect ranking-based measures and not precedence invariant measures (e.g., ppc, aprd).
test
a logical indicating if a correlation test should be performed.
...
Further arguments passed on to the method.
Details

ser_cor calculates the correlation between two sequences (orders). Note that a seriation order and its reverse are identical and purely an artifact due to the method that creates the order. This is a major difference to rankings. For ranking-based correlation measures (Spearman and Kendall) the absolute value of the correlation is returned for reverse = TRUE (in effect returning the correltation for the reversed order). If test = TRUE then the appropriate test for association is performed and a matrix with p-values is returned as the attribute "p-value". Note that no correction for multiple testing is performed.

For ser_dist, the correlation coefficients (Kendall's tau and Spearman's rho) are converted into a dissimilarity by taking one minus the correlation value. Note that Manhattan distance between the ranks in a linear order is equivalent to Spearman's footrule metric (Diaconis 1988). reverse = TRUE returns the pairwise minima using also reversed orders.

The positional proximity coefficient (ppc) is a precedence invariant measure based on product of the squared positional distances in two permutations defined as (see Goulermas et al 2016):

$$d_{ppc}(R, S) = 1/h \sum_{j=2}^n \sum_{i=1}^{j-1} (\pi_R(i)-\pi_R(j))^2 * (\pi_S(i)-\pi_S(j))^2,$$

where \(R\) and \(S\) are two seriation orders, \(pi_R\) and \(pi_S\) are the associated permutation vectors and \(h\) is a normalization factor. The associatied generalized correlation coefficient is defined as \(1-d_{ppc}\). For this precedence invariant measure reverse is ignored.

The absolute pairwise rank difference (aprd) is also precedence invariant and defined as a distance measure:

$$d_{aprd}(R, S) = \sum_{j=2}^n \sum_{i=1}^{j-1} | |\pi_R(i)-\pi_R(j)| - |\pi_S(i)-\pi_S(j)| |^p,$$

where \(p\) is the power which can be passed on as parameter p and is by default set to 2. For this precedence invariant measure reverse is ignored.

ser_align tries to normalize the direction in a list of seriations such that ranking-based methods can be used. We add for each permutation also the reversed order to the set and then use a modified version of Prim's algorithm for finding a minimum spanning tree (MST) to choose if the original seriation order or its reverse should be used. We use the orders first added to the MST. Every time an order is added, its reverse is removed from the possible remaining orders.

Value

ser_dist returns an object of class dist. ser_align returns a new list with elements of class ser_permutation.

References

P. Diaconis (1988): Group Representations in Probability and Statistics. Institute of Mathematical Statistics, Hayward, CA.

J.Y. Goulermas, A. Kostopoulos, and T. Mu (2016): A New Measure for Analyzing and Fusing Sequences of Objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(5):833-48.

See Also

ser_permutation_vector

Aliases
  • dist
  • ser_dist
  • ser_cor
  • ser_align
Examples
set.seed(1234)
## seriate dist of 50 flowers from the iris data set
data("iris")
x <- as.matrix(iris[-5])
x <- x[sample(1:nrow(x), 50),]
rownames(x) <- 1:50
d <- dist(x)

## Create a list of different seriations
methods <- c("HC_single", "HC_complete", "OLO", "GW", "R2E", "VAT",
  "TSP", "Spectral", "SPIN", "MDS", "Identity", "Random")

os <- sapply(methods, function(m) {
  cat("Doing", m, "... ")
  tm <- system.time(o <- seriate(d, method = m))
  cat("took", tm[3],"s.\n")
  o
})

## Compare the methods using distances. Default is based on
## Spearman's rank correlation coefficient. Reverse orders are considered
## equivalent.
ds <- ser_dist(os)
hmap(ds, margin=c(7,7))

## Compare using actual correlation between orders. Reversed orders have
## negative correlation!
cs <- ser_cor(os, reverse = FALSE)
hmap(cs, margin=c(7,7))

## Also check reversed seriation orders.
## Now all but random and identity are highly positive correlated
cs2 <- ser_cor(os, reverse = TRUE)
hmap(cs2, margin=c(7,7))

## Use Manhattan distance of the ranks (i.e., Spearman's foot rule)
ds <- ser_dist(os, method="manhattan")
plot(hclust(ds))
Documentation reproduced from package seriation, version 1.2-2, License: GPL-3

Community examples

Looks like there are no examples yet.