Classproj {shipunov}

Class projection

Description

Class projection which preserves distances between class centers

Usage

Classproj(data, classes, method="DMS")

Arguments

data

Data: must be numeric and convertible into matrix

classes

Class labels (correspond to data rows), NAs are allowed (sic!)

method

Either "DMS" for Dhillon et al., 2002 or "QJ" for Qiu and Joe, 2006

Details

'Classproj' is the leveraged (supervised) or educated (semi-supervised) manifold learning (dimension reduction). See examples for the variety of its uses.

It uses classes to determine centers and then tries to preserve distances between centers; two methods are possible: "DMS" which is slightly faster, and "QJ" which frequently finds a better projection.

The code is based on the functions from 'clusterGeneration' package from Weiliang Qiu.

Value

Returns list with 'proj' coordinates of projected data points and 'centers' coordinates of class centers.

Author(s)

Alexey Shipunov

References

Dhillon I.S., Modha D.S., Spangler W.S. 2002. Class visualization of high-dimensional data with applications. Computational Statistics and Data Analysis. 41: 59–90.

Qiu W.-L., Joe H. 2006. Separation index and partial membership for clustering. Computational Statistics and Data Analysis. 50: 585–603.

Examples

## Leveraged approach (all classes are known)
iris.dms <- Classproj(iris[, -5], iris$Species, method="DMS")
plot(iris.dms$proj, col=iris$Species)
text(iris.dms$centers, levels(iris$Species), col=1:3)

iris.qj <- Classproj(iris[, -5], iris$Species, method="QJ")
plot(iris.qj$proj, col=iris$Species)
text(iris.qj$centers, levels(iris$Species), col=1:3)

## Educated approach (classes are known only for 10 data points per class)
sam <- Class.sample(iris$Species, 10)
newclasses <- iris$Species
newclasses[!sam] <- NA

iris.dms <- Classproj(iris[, -5], newclasses)
plot(iris.dms$proj, col=iris$Species, pch=ifelse(sam, 19, 1))
text(iris.dms$centers, levels(iris$Species), col=1:3)

iris.qj <- Classproj(iris[, -5], newclasses, method="QJ")
plot(iris.qj$proj, col=iris$Species, pch=ifelse(sam, 19, 1))
text(iris.qj$centers, levels(iris$Species), col=1:3)

## Automated approach (classes calculated automatically)
## Good to visualize _any_ clustering or learning
iris.km <- kmeans(iris[, -5], 3)

iris.dms <- Classproj(iris[, -5], iris.km$cluster)
plot(iris.dms$proj, col=iris.km$cluster)
text(iris.dms$centers, labels=1:3, col=1:3, cex=2)

iris.qj <- Classproj(iris[, -5], iris.km$cluster, method="QJ")
plot(iris.qj$proj, col=iris.km$cluster)
text(iris.qj$centers, labels=1:3, col=1:3, cex=2)