shipunov (version 1.13)

Misclass: Misclassification (confusion) table

Description

Misclassification (confusion) table

Usage

Misclass(pred, obs, best=FALSE, ignore=NULL, quiet=FALSE, force=FALSE, ...)

Arguments

pred

Predicted class labels

obs

Observed class labels

best

Perform a search for the classification table with minimal misclassification error?

ignore

Vector of class labels to ignore (convert into NAs)

quiet

Output summary?

force

Override the restriction of class number in 'best=TRUE' and speed up code?

...

Arguments to 'table'

Value

Invisibly returns the table of class comparison

Details

'Misclass()' produces misclassification (confusion) 2D table based on two classifications.

The simple variant ('best=FALSE') assumes that class labels are concerted (same number of corresponding classes).

Advanced variant ('best=TRUE') can search for the best classification table (with minimal misclassification rate), this is especially useful in case of unsupervised classifications which typically return numeric labels. It therefore assumes that the table is a result of some non-random process. However, internally it generates all permutations of factor levels and could be very slow if there are 8 and more class labels. Therefore, more than 8 classes are not allowed. It is possible nevertheless to override this restriction with 'force=TRUE'; this option also uses the experimental code which replaces internal table() with tabulate() and is much faster with many labels.

Variant with 'best=TRUE' might also add empty rows (filled with zeros) to the table in case if numbers of classes are not equal.

Additional arguments could be passed for table(), for example, 'useNA="ifany"'. If supplied data contains NAs, there will be also note in the end. Note that tabulate()-based code (activated with force="TRUE") does not take table()-specific arguments, so if this is a case, warning will be issued.

It is possible to ignore (convert into NAs) some class labels with 'ignore=...', this is useful for methods like DBSCAN which output special label for outliers. In that case, note about missing data is also issued.

Alternatives: confusion matrix from caret::confusionMatrix() which is more feature rich but much less flexible. See in examples how to implement some statistics used there.

Note that partial "Misclassification errors" are reverse sensitivities, and "Mean misclassification error" is a reverse accuracy.

If you want to plot misclassification table, Cohen-Friendly association plot, assocplot() is probably the best. On this plot, note rectangles which are big, tall and black (check help(assocplot) to know more). Diagonal which is black and other cells red indicate low misclassification rates.

See Also

Adj.Rand, link{assocplot}

Examples

# NOT RUN {
iris.dist <- dist(iris[, -5], method="manhattan")
iris.hclust <- hclust(iris.dist)
iris.3 <- cutree(iris.hclust, 3)
Misclass(iris.3, iris[, 5])

set.seed(1)
iris.k <- kmeans(iris[, -5], centers=3)
Misclass(iris.k$cluster, iris[, 5])
Misclass(iris.k$cluster, iris[, 5], best=TRUE)

res <- Misclass(iris.k$cluster, iris[, 5], best=TRUE, quiet=TRUE)
## how to calculate statistics from caret::confusionMatrix()
binom.test(sum(diag(res)), sum(res))$conf.int
mcnemar.test(res) # to avoid NA's, add small number to 'res'
## how to plot misclassification table
assocplot(res)
## how to use Misclass() for Recode()
nn <- Recode(iris.k$cluster, from=dimnames(res)$pred, to=dimnames(res)$obs)
head(nn)

library(dbscan)
iris.db <- dbscan(iris[, -5], eps=0.3)
Misclass(iris.db$cluster, iris$Species, ignore=0, best=TRUE)

set.seed(NULL)
# }