plot.sompdist: Plot distances between prototypes of a fitted Self-Organising...

In yasomi: Yet Another Self Organising Map Implementation

Description

Provide a u-matrix like visual representation of the distances between prototypes of neighbouring units of a fitted Self-Organising Map.

Usage

## S3 method for class 'sompdist'
plot(x, mode=c("mean","full"), ...)

Arguments

x	an object of class "sompdist" obtained via prototype.distances
mode	specifies how distances are aggregated for display (see below for details)
...	additional parameters transmitted to the low level plot function plot.somgrid

Details

This function provides a simple U-matrix like visualisation method for the distances between prototypes of direct neighbouring units of a fitted SOM. The main idea is to use colour coding of the cells of the prior structure to represent those distances. There are two modes for generating the picture:

"mean"

in this mode, the visualisation grid is identical to the grid used as the prior structure. Each cell (rectangular or hexagonal, depending on the grid type) is filled with a colour chosen to represent the average distance between the corresponding prototype and the prototypes of its neighbour units in the prior structure.

"full"

in this mode, the function uses a grid approximately four time as large as the original one. The visualisation is based on the insertion of additional fake units between each unit and its direct neighbours in the original grid. As in the "mean" mode, each original unit displays via its colour the average of the distances between its prototype and the neighbouring one. In addition, fake units display (again with a colour code) the actual distance between a prototype and its neighbour. In the case of an hexagonal grid, this can be done exactly. In the case of a rectangular grid, only horizontal and vertical neighbours can be represented exactly. Units added in diagonal are shared between two pair of prototypes and represent therefore the mean of the two corresponding distances (see also distance.grid for a similar solution to the same problem).

The additional parameters given to the function can be used to control the underlying plot.somgrid function, e.g. to change the default colour palette (heat.colors).

Author(s)

Fabrice Rossi

References

Ultsch, A. and Siemon, H. P. (1990) Kohonen's self organizing feature maps for exploratory data analysis, in: Proceedings of International Neural Network Conference (INNC'90).

See Also

See prototype.distances to get the distance structure, umatrix for direct access to this type of display and distance.grid for possibly smoother plots.

Examples

data(iris)
# scaling
data <- scale(iris[1:4])

# a medium hexagonal grid
sg <- somgrid(xdim=15,ydim=15,topo="hex")

# choose a good SOM via Kaski and Lagus' error measure
st <- som.tune(data,sg,som.tunecontrol(sg,criterion=error.kaskilagus))
som <- st$best.som

# compute the distance
pdist <- prototype.distances(som)

# simple mean based umatrix
plot(pdist)

# more complete display
plot(pdist,mode="full")

yasomi documentation built on May 2, 2019, 5:59 p.m.