pairs2svmData: pairs2svmData

In tdhock/rankSVMcompare: Support Vector Machines for Ranking and Comparing

Description

We do several preprocessing steps on the input data:

1. centering and scaling.

2. for all pairs i such that yi=-1, we flip Xi and Xip and set yi=1.

3. for all pairs i such that yi=0, we generate new features Xi <- Xip and Xip <- Xi with corresponding new yi=0 labels.

4. We take the difference Xip-Xi of the resulting scaled, flipped, augmented feature matrices.

5. We map 0 -> -1 in the resulting label vector, creating an integer vector with elements in c(-1,1).

Usage

pairs2svmData(Pairs)

Arguments

Pairs

Value

data suitable for plugging into an SVM solver:

center	center of the input features.
scale	scale of input features.
Xi	inputs feature matrix.
Xip	inputs feature matrix.
yi	outputs -1 -> 1, 0 -> -1, 1 -> 1.

Author(s)

Toby Dylan Hocking

Examples

p <- list(Xi=cbind(var=c(3,0,1)),
          Xip=cbind(var=c(0,-2,0)),
          yi=as.integer(c(-1,1,0)))
(result <- pairs2svmData(p))
## Inequality pairs such that yi=1 or -1 are mapped to 1, and
## equality pairs such that yi=0 are duplicated and mapped to -1.
stopifnot(result$labels == c(1,1,-1,-1))
## The duplicate equality features are flipped.
stopifnot(result$Xi[3]==result$Xip[4])
stopifnot(result$Xi[4]==result$Xip[3])

tdhock/rankSVMcompare documentation built on May 31, 2019, 7:38 a.m.