summary(spe.KM.cascade)
spe.KM.cascade$results
The minimum of SSE is the criterion used by the algorithm to find the optimal
grouping of the objects for a given value of k, while calinski and ssiare good
criteria to find the optimal value of k.
Remember that the different partitions in k = {2, 3, …, 10} groups are computed
independently of one another. Examining the plot from bottom to top is NOT
equivalent to examining a dendrogram because groups of successive partitions
are not necessarily nested .
After defining site clusters, it is time to examine their contents. The simplest way
is to define subgroups of sites on the basis of the typology retained and compute
basic statistics. You can run the following example based upon the k-means 4-group
partition.
# Reorder the sites according to the k-means result
spe.kmeans.g <- spe.kmeans$cluster
spe[order(spe.kmeans.g), ]
# Reorder sites and species using function vegemite()
ord.KM <- vegemite(spe, spe.kmeans.g)
spe[ord.KM$sites, ord.KM$species]
4.8.1.2 Use of k-means Partitioning to Optimize an Independently
Obtained Classification
From another point of view, k-means partitioning can be used to optimise the result
of a hierarchical clustering. Indeed, the nature of an agglomeration algorithm such as
those explored in the previous sections prevents an object that has been included in a
group to be translocated to another that appeared later in the agglomeration process,
even if the latter would now be more appropriate. To overcome this potential
problem, one can provide the k-means algorithm with prior information derived
from the clustering to be optimized, either in the form of a k  p matrix of mean
values of the p variables in the k groups obtained with the method, or as a list of
k objects, one per group, considered to be “typical” for each group. In the latter case,
these “typical” objects, called medoids, are used as seeds for the construction of
groups in the technique presented in the next Section. Let us apply this idea to verify
if the four groups obtained by a Ward clustering of the fish data are modified by a kmeans partitioning.
100
4 Cluster Analysis
spe.KM.cascade$results
The minimum of SSE is the criterion used by the algorithm to find the optimal
grouping of the objects for a given value of k, while calinski and ssiare good
criteria to find the optimal value of k.
Remember that the different partitions in k = {2, 3, …, 10} groups are computed
independently of one another. Examining the plot from bottom to top is NOT
equivalent to examining a dendrogram because groups of successive partitions
are not necessarily nested .
After defining site clusters, it is time to examine their contents. The simplest way
is to define subgroups of sites on the basis of the typology retained and compute
basic statistics. You can run the following example based upon the k-means 4-group
partition.
# Reorder the sites according to the k-means result
spe.kmeans.g <- spe.kmeans$cluster
spe[order(spe.kmeans.g), ]
# Reorder sites and species using function vegemite()
ord.KM <- vegemite(spe, spe.kmeans.g)
spe[ord.KM$sites, ord.KM$species]
4.8.1.2 Use of k-means Partitioning to Optimize an Independently
Obtained Classification
From another point of view, k-means partitioning can be used to optimise the result
of a hierarchical clustering. Indeed, the nature of an agglomeration algorithm such as
those explored in the previous sections prevents an object that has been included in a
group to be translocated to another that appeared later in the agglomeration process,
even if the latter would now be more appropriate. To overcome this potential
problem, one can provide the k-means algorithm with prior information derived
from the clustering to be optimized, either in the form of a k  p matrix of mean
values of the p variables in the k groups obtained with the method, or as a list of
k objects, one per group, considered to be “typical” for each group. In the latter case,
these “typical” objects, called medoids, are used as seeds for the construction of
groups in the technique presented in the next Section. Let us apply this idea to verify
if the four groups obtained by a Ward clustering of the fish data are modified by a kmeans partitioning.
100
4 Cluster Analysis