Different running configurations
In this section we will try different ways of calling the segclust() function. Minimal call only contains the data concerned (here the dataframe simulmode) as well as arguments lmin, ncluster and seg.var. The function will automatically complete the missing arguments with default values and communicate about it (not shown here).
mode_segclust <- segclust(simulmode,
Kmax = 20, lmin=10,
ncluster = c(2,3),
seg.var = c("speed","abs_spatial_angle"))
Additionally it is advised to scale the variables used to a mean of 0 and a variance of 1. This will automatically be done for segclust(), but can also be specified explicitly:
mode_segclust <- segclust(simulmode,
Kmax = 20, lmin=10,
ncluster = c(2,3),
seg.var = c("speed","abs_spatial_angle"),
scale.variable = TRUE)
Selecting the number of clusters and segments.
Once the segmentation/clustering have been successfully run. A summary like this will appear:
#> ✓ Best segmentation/clustering estimated with 3 clusters and 15 segments according to BIC
#> → Number of clusters should preferentially be selected according to biological
#> knowledge. Exploring the BIC plot with plot_BIC() can also provide advice to
#> select the number of clusters.
#> → Once number of clusters is selected, the number of segments can be selected
#> according to BIC.
#> → Results of the segmentation/clustering may further be explored with plot()
#> and segmap()
As stated, the number of clusters should preferentially be selected based on biological knowledge. Exploration of the BIC-based penalized log-likelihood can also help selecting an appropriate number of clusters, with the function plot_BIC(). Best-case scenario is as below, the BIC shows a steep increase up to a maximum and a slow decrease after the optimum and one number of clusters is clearly above the others.
With real data, a larger number of clusters almost always improves the penalized-likelihood so it will generally be a poor indication of the appropriate number of clusters. Such situation may be shown with our dataset simulmode indeed, if we allow the number of clusters to be tested between 2 and 5 instead of a maximum at three, we obtain the following BIC curve:
mode_segclust <- segclust(simulmode,
Kmax = 20, lmin=10,
ncluster = 2:5,
seg.var = c("speed","abs_spatial_angle"),
scale.variable = TRUE)
plot_BIC(mode_segclust)
In the curve above we see that the optimum number of clusters selected is 4, despite the difference being relatively small compared to ncluster = 3 and to the evolution with the number of segments. In addition, if we look at the results of the segmentation we can see identical segment distribution, the only difference is that state 2 in the segmentation with 3 clusters is divided into state 2 and 3 in the segmentation with 4 clusters:
plot(mode_segclust, ncluster = 3)
plot(mode_segclust, ncluster = 4)
More generally, as in this example, if the selected number of segments for a higher number of clusters is the same, then the lower number of clusters should be favored.
Once the number of clusters have been selected, the number of segments can be selected with the BIC criterium. When the segmentation-clustering is reliable, the selected optimum should be a maximum just before a linear drop of the penalized log-Likelihood as in the examples above. With real data the signal is seldom as clear and may require adjustment based on exploration of the realized segmentation/clustering.
