/somKernels

Extra distance calculation Rcpp pointers for the kohonen package

Primary LanguageC++GNU General Public License v3.0GPL-3.0

somKernels

Extra distance Rcpp pointers for the kohonen package.

Enables generation of self-organizing maps (SOM) with additional distance measures provided as the dist.fcts argument to som() and supersom() functions from kohonen package. Run get_kernel_info() to get available distance measures. The distance pointers are called by their name (e.g. dist.fcts = 'dice').

Installation

You may easily fetch the package with devtools:

devtools::install_github('PiotrTymoszuk/somKernels')

Terms of use

The package is available under a GPL-3 license.

Contact

The package maintainer is Piotr Tymoszuk.

Acknowledgements

soucer uses tools provided by Rcpp. Many thanks to the developers, maintainers and contributors.

Usage

We'll start with loading the packages needed to test functionality of the distance pointers:

  library(kohonen)
  library(somKernels)

The available distances include most of distances covered by the R's philentropy package. The distance list is returned by get_kernel_info():

> get_kernel_info()

 [1] "BrayCurtis"        "cosine"            "dice"              "jaccard"           "smc"               "sumofsquares"     
 [7] "euclidean"         "manhattan"         "tanimoto"          "canberra"          "chebyshev"         "soergel"          
[13] "kulczynski_d"      "lorentzian"        "intersection"      "wavehedges"        "czekanowski"       "motyka"           
[19] "ruzicka"           "inner_product"     "harmonic_mean"     "hassebrook"        "fidelity"          "bhattacharyya"    
[25] "hellinger"         "matusita"          "squared_chord"     "squared_euclidean" "pearson"           "neyman"           
[31] "squared_chi"       "prob_symm"         "divergence"        "clark"             "additive_symm"     "jeffreys"         
[37] "k_divergence"      "topsoe"            "jensen_difference" "taneja"

Usage of the distance pointers is quite simple: you just pass the pointer name to the dist.fcts argument of the kohonen's som() or supesom() function. We'll check them out with the popular iris data set. The data set will be clustered with $5 \times 4$ hexagonal SOM with Canberra and Ruzicka distance. Of importance, simirarity coefficients such as cosine or Ruzicka coefficient are donverted to distances with a simple $dist = 1 - simil$ formula.

  test_data <-
    iris[c('Sepal.Length', 'Sepal.Width', 'Petal.Length', 'Petal.Width')]

  ## generating a SOM grid object required by `som()` and `supersom()`

  som_grid <- somgrid(xdim = 5,
                      ydim = 4,
                      topo = 'hexagonal',
                      neighbourhood.fct = 'gaussian',
                      toroidal = TRUE)

  ## fitting SOMs with Canberra and Ruzicka distances

  canberra_som <- som(as.matrix(test_data),
                      grid = som_grid,
                      rlen = 1000,
                      dist.fcts = 'canberra')

  ruzicka_som <- som(as.matrix(test_data),
                     grid = som_grid,
                     rlen = 1000,
                     dist.fcts = 'ruzicka')

Let's check if the distances worked by examining plots of their training process - mean distances to the SOM nodes at subsequent iterations. A decrease in mean distance with progressing iterations followed by a plateau indicates algorithm convergence:

plot(canberra_som, type = 'changes')

image

plot(ruzicka_som, type = 'changes')

image