getCover: Get Cover of partitioned SVM

In liquidSVM: A Fast and Versatile SVM Package

Description

If you use voronoi=3 or voronoi=4 this retrieves the voronoi centers that have been found.

Usage

getCover(model, task = 1)

Arguments

model	the model
task	the task between 1 and number of tasks

Value

the indices of the samples in the training data set that were used as Voronoi partition centers.

Note

This is not tested thoroughly so use in production is at your own risk.

Examples

## Not run: 
banana <- liquidData('banana-mc')
model <- mcSVM(Y~.,banana$train, voronoi=c(4,500),d=1)
# task 4 is predicting 2 vs 3
cover <- getCover(model,task=4)
centers <- cover$samples
# we are considering task 4 and hence only show labels 2 and 3:
bananaSub <- banana$train[banana$train$Y %in% c(2,3),]
plot(bananaSub[,-1],col=bananaSub$Y)
points(centers,pch='x',cex=2)

if(require(deldir)){
  voronoi <- deldir::deldir(centers$X1,centers$X2,rw=c(range(bananaSub$X1),range(bananaSub$X2)))
  plot(voronoi,wlines="tess",add=TRUE, lty=1)
  text(centers$X1,centers$X2,1:nrow(centers),pos=1)
}

# let us calculate for every sample in this task which cell it belongs to
distances <- as.matrix(dist(model$train_data))
cells <- apply(distances[model$train_labels %in% c(2,3),cover$indices],1,which.min)
# and you can check that the cell sizes are as reported in the training phase for task 4
table(cells)

## End(Not run)

liquidSVM documentation built on Sept. 15, 2019, 1:02 a.m.