BestOverlap {shipunov}

Calculates the best overlap

Description

Uses multiple datasets, measures overlaps between class-related convex hulls and reports the best dataset, the best overlap table and summary with confidence intervals. Can be used to assess bootstrap or jackknife results, to compare different dimension reduction and/or clustering methods, and to average results of stochastic methods.

Usage

BestOverlap(xylabels, ci="95%", round=4)

Arguments

xylabels

List of data frames, each with at least 3 columns named exactly as: "x" for x coordinates, "y" for y coordinates and "labels" for class labels

ci

Confidence interval (character string with percent sign)

round

How to round numbers in summary table

Details

'BestOverlap()' requires object, typically created after bootstrapping or similar procedure (see below for examples). This 'xylabels' object must contain at least three columns named exactly as c("x", "y", "labels"), in any order.

Please note that label types must be the same between data frames inside 'xylabels' list. For consistency, first data frame is used as a label standard. If any next data frame contain label types different from standard, it will be ignored.

Value

List with three components: 'best' data frame, 'best.overlap' table and 'summary' data frame.

Author(s)

Alexey Shipunov

Examples

 
## Bootstrap PCA
B <- 100
xylabels <- vector("list", length=0)
for (n in 1:B) {
ROWS <- sample(nrow(iris), replace=TRUE)
tmp <- prcomp(iris[ROWS, -5])$x[, 1:2]
xylabels[[n]] <- data.frame(x=tmp[, 1], y=tmp[, 2], labels=iris[ROWS, 5])
}
BestOverlap(xylabels)

## Jacknife PCA
B <- nrow(iris)
xylabels <- vector("list", length=0)
for (n in 1:B) {
ROWS <- (1:B)[-n]
tmp <- prcomp(iris[ROWS, -5])$x[, 1:2]
xylabels[[n]] <- data.frame(x=tmp[, 1], y=tmp[, 2], labels=iris[ROWS, 5])
}
BestOverlap(xylabels)

## Stochastic method: Stochastic Proximity Embedding
library(tapkee)
B <- 100
xylabels <- vector("list", length=0)
for (n in 1:B) {
tmp <- Tapkee(iris[, -5], method="spe")
xylabels[[n]] <- data.frame(x=tmp[, 1], y=tmp[, 2], labels=iris[, 5])
}
BestOverlap(xylabels)

## Diverse dimension reduction methods
library(tapkee)
B <- c("lle", "npe", "ltsa", "lltsa", "hlle", "la", "lpp", "dm", "isomap", "l-isomap")
xylabels <- vector("list", length=0)
for (n in B) {
tmp <- Tapkee(iris[, -5], method=n, add="-k 50")
xylabels[[n]] <- data.frame(x=tmp[, 1], y=tmp[, 2], labels=iris[, 5])
}
BestOverlap(xylabels)

## One dimension reduction but many clusterings
B <- 100
xylabels <- vector("list", length=0)
tmp1 <- prcomp(iris[, -5])$x[, 1:2]
for (n in 1:B) {
tmp2 <- kmeans(iris[, -5], centers=3)$cluster
xylabels[[n]] <- data.frame(x=tmp1[, 1], y=tmp1[, 2], labels=letters[tmp2])
}
BestOverlap(xylabels)