SemiSupervised (version 1.0)

predict.s4pm: Out-of-Sample Predict Procedure for s4pm

Description

This implements the out-of-sample prediction for an ‘s4pm’ object.

Usage

# S4 method for s4pm
predict(object,xnew,gnew,type=c("vector","response","prob","terms"),pow=1,…)

Arguments

object

an existing ‘s4pm’ object.

xnew

an object of class ‘data.frame’, ‘vector’, or ‘matrix’. This is not always necessary and depends on call context (refer to details below).

gnew

the ‘matrix’ of new graph links between the data to predict and the data used for training. This is not always necessary and depends on call context (refer to details below).

type

the type of prediction to return. In the case of terms the fits are returned and the individual functions with f1 ss the graph and f2 as the linear term.

pow

tuning parameter for the weighted power in the interpolation predictions.

…

mop up additional arguments.

Value

If type(object) is ‘r’, a vector of predicted values is returned. If type(object) is ‘c’, the object returned depends on the type argument.

Details

The prediction inputs are dependent upon how one calls the original s4pm generic function. The cases are discussed next:

1) y~.: This is the default and most common case. Set xnew to your new hold-out data set and do not initialize gnew.

2) y~dG(G): The gnew argument will [most likely] be a non-symmetric ‘matrix’ of adjacencies between some new set of observations and the original x data.

3) y~sG(G): The gnew argument will [most likely] be a non-symmetric ‘matrix’ of similarity adjacencies [most likely] observed directly.

4) y~.+dG(G) or y~.+sG(G): The xnew data set must be given, but it need not correspond to the data used to generate the graph.

5) Non-formula call: xnew will either be provided or NULL depending on context, and gnew will have to be provided in this case.

Examples

# NOT RUN {
## Prediction depends on the nature of the call. Consider some examples.
library(mlbench)
data(Sonar)

n=dim(Sonar)[1]
p=dim(Sonar)[2]

nu=0.2
set.seed(100)
L=sort(sample(1:n,ceiling(nu*n)))
U=setdiff(1:n,L)
U1=sample(U,ceiling(0.5*n))

y.true<-Sonar$Class
Sonar$Class[U]=NA

## Typical, call to s4pm and predict

g.s4pm<-s4pm(Class~.,data=Sonar[c(L,U1),])
p.s4pm<-predict(g.s4pm,xnew=Sonar[U,-p])
tab=table(y.true[U],p.s4pm)
1-sum(diag(tab))/sum(tab)
# }
# NOT RUN {
## Predict the graph only case
Dij<-x.scaleL(Sonar[,-p],L)
Dij<-as.matrix(cosineDist(Dij))
Dij1<-Dij[c(L,U1),c(L,U1)]
attr(Dij1,"metric")=attr(Dij,"metric")
attr(Dij1,"distance.graph")=attr(Dij,"distance.graph")
g.s4pm<-s4pm(Class~dG(Dij1),data=Sonar[c(L,U1),])
p.s4pm<-predict(g.s4pm,gnew=Dij[U,c(L,U1)])
tab=table(y.true[U],p.s4pm)
1-sum(diag(tab))/sum(tab)
# }