Classification: MclustDA
library(condvis2)
#> Registered S3 method overwritten by 'seriation':
#> method from
#> reorder.hclust gclus
library(mclust)
#> Package 'mclust' version 5.4.10
#> Type 'citation("mclust")' for citing this R package in publications.data(banknote)
bankDA <- MclustDA(banknote[,-1], banknote[,1],verbose=F)This fits an XXX=ellipsoidal multivariate normal for the genuine notes and EVE2 = ellipsoidal, equal volume and equal shape with two components for counterfeits.
Condvis uses a generic CVpredict to provide a uniform
interface to predict methods. We will use it to find the
number of misclassifications:
table(banknote$Status, CVpredict(bankDA, banknote))
#>
#> counterfeit genuine
#> counterfeit 100 0
#> genuine 0 100svars <- c("Top", "Diagonal")
cvars <- setdiff(names(banknote)[-1], svars)
condvis(data = banknote, model = bankDA,
response="Status", sectionvars=svars,conditionvars=cvars,
pointColor="Status", showsim=TRUE
)The showsim=TRUE setting means that the condition plots
will mark points considered sufficiently near the section with a dark
outline. This is the default setting only when \(n <= 150\). In the condition plots, the
green points are the genuine ones, pink are counterfeit.
Here is a views of the result:
There is one point that appears to be mis-classified (though it is not).
However, as it’s size is small it is not very close to the selected
section.
Selecting Show probs: we see there is some uncertainty
about the classification on the classification boundaries.
A different, simpler fit is got using
bankDAe <- MclustDA(banknote[,-1], banknote[,1], modelType="EDDA",verbose=F)This uses an EVE1 for each class, and there is one mis-classification, at case 70.
We can use condvis to compare the two fits:
condvis(data = banknote, model = list(bankDA=bankDA, bankDAe=bankDAe),
response="Status", sectionvars=svars,conditionvars=cvars,
pointColor="Status", showsim=T
)If you go to Tour, Diff fits, you can move through views
where the conditioning points are those where the fits disagree. In this
case there is only one such point. Here Show probs is
selected so there is some uncertainty visible at the classification
boundaries.
The screenshot shows that in this view.
The mis-classified point is a genuine note (green) which is just inside the region classified as counterfeit by EDDA. Here also the Similarity threshold is made small so only the point on the section is visible in the section plot. From the condition plots it is evident this note has a low Diagonal size, like the counterfeits.
