demo/korresp.R

In SOMbrero: SOM Bound to Realize Euclidean and Relational Outputs

library(ggplot2)
data(presidentielles2002)

# train the korresp algorithm
## details on the presidentielles2002 data set are provided with 'help(presidentielles2002)'
## default dimensions will be calculated by the algorithm (see 'help(trainSOM)')
presi.som <- trainSOM(x.data=presidentielles2002, type="korresp", scaling="chi2", dimension=c(8,8))

# prototypes overview can either be plotted for columns variables
plot(presi.som, what="prototypes", type="barplot", view="c")
# or for row variables
plot(presi.som, what="prototypes", type="barplot", view="r")
# Plots are ggplot2 so you can manipulate plot elements afterwards (if legend is too big for example)
plot(presi.som, what="prototypes", type="barplot", view="r") +    
  guides(fill=guide_legend(keyheight=0.6, ncol=2, label.theme=element_text(size=6))) + 
  theme(axis.text.x=element_blank(), axis.ticks.x=element_blank())

# the distances between prototypes can be displayed 
## either with smooth colors
plot(presi.som, what="prototypes", type="smooth.dist")
## or with a Multi Dimensional Scaling
plot(presi.som, what="prototypes", type="mds")

# observation distribution overviews are provided
## either with a plot
plot(presi.som, what="obs", type="hitmap")
## or with a table
table(presi.som$clustering)
## or more precisely by printing the variables names on the map
### NOTE: in the korresp SOM, both row and columns variables are considered
### WARNING: this graphic may produce some warnings when all row names can not fit on the plot
### this problem can be solved by enlarging your graphic device or by using the 'scale' argument (see 'help(wordcloud)')
plot(presi.som, what="obs", type="names")

# hierarchical clustering is also available
## compute 3 super clusters
presi.sc <- superClass(presi.som, k=3)
## plot its 3 dimensional dendrogram
plot(presi.sc, type="dendro3d")
## view the prototypes positions in the Multi Dimensional Scaling plot
plot(presi.sc, type="mds")

# two projection quality indicators are still available
quality(presi.som)