protoclass (version 1.0)

protoclass: Greedy algorithm for prototype selection

Description

Selects prototypes for each class in a greedy manner as described in 'Bien and Tibshirani (2011) Prototype Selection for Interpretable Classification. Annals of Applied Statistics. 5(4). 2403-2424.'

protoclass

Usage

protoclass(x, y, z, dxz, eps, lambda = 1/n)

Arguments

x

n by p matrix of training features (optional, see dxz).

y

n-vector of labels of the training data.

z

set of potential prototypes (optional, see dxz).

dxz

instead of x and z, you can give dxz, the matrix of pairwise dissimilarities between x and z, with ij-th element giving the dissimilarity between training point x_i and prototype-candidate z_j.

eps

size of covering balls.

lambda

cost of adding a prototype.

Value

An object of class "protoclass," which has the following elements:

• alpha: Matrix of dimensions nrow(z)-by-nclass. alpha[j,k] indicates whether jth potential prototype has been chosen as a prototype for class k.
• classes: Names of classes
• proto.order: The sequence of prototypes that were selected.
• ncovered: nproto-by-nclass matrix with ncovered[j,k] giving the number of class k training points covered by the jth prototype's ball.
• coverlist: n-by-nclass matrix with row i giving number of each type of prototype covering point i.
• uncovered: Indicates whether a training point is not covered by a prototype of its own class.
• wrongcover: Number of prototypes from other classes covering each training point.
• nproto: nclass-vector giving the number of prototypes in each class.

Details

It's more efficient to compute dxz just once on your own rather than have protoclass repeatedly compute the pairwise distances on each call.

See Also

predict.protoclass

Examples

# generate some data:
set.seed(1)
n <- 200
p <- 2
x <- matrix(rnorm(n * p), n, p)
y <- rep(c("A","B"), each=n/2)
x[y=="A", ] <- x[y=="A", ] + 3
itr <- sample(n, n/2)
xtr <- x[itr, ] # train
ytr <- y[itr]
xte <- x[-itr, ] # test
yte <- y[-itr]

# take prototype candidates identical to training points:
z <- xtr
dxz <- dist2(xtr, z)
# run protoclass:
prot <- protoclass(dxz=dxz, y=ytr, eps=2, lambda=1/n)
## Not run: 
# plot(prot,xtr,y=1+(ytr=="A"))
# ## End(Not run)
# get predictions on test data:
pred1 <- predict(prot, xte, z=xtr)
# get predictions on test data using pairwise distances:
pred2 <- predictwithd.protoclass(prot, dist2(xte, z))