sda {sparseLDA}

Sparse discriminant analysis

Description

Performs sparse linear discriminant analysis. Using an alternating minimization algorithm to minimize the SDA criterion.

Usage

sda(x, ...)

## Default S3 method:
sda(x, y, lambda = 1e-6, stop = -p, maxIte = 100,
    Q = K-1, trace = FALSE, tol = 1e-6, ...)

Arguments

x

A matrix of the training data with observations down the rows and variables in the columns.

y

A matrix initializing the dummy variables representing the groups.

lambda

The weight on the L2-norm for elastic net regression. Default: 1e-6.

stop

If STOP is negative, its absolute value corresponds to the desired number of variables. If STOP is positive, it corresponds to an upper bound on the L1-norm of the b coefficients. There is a one to one correspondence between stop and t. The default is -p (-the number of variables).

maxIte

Maximum number of iterations. Default: 100.

Q

Number of components. Maximum and default is K-1 (the number of classes less one).

trace

If TRUE, prints out its progress. Default: FALSE.

tol

Tolerance for the stopping criterion (change in RSS). Default is 1e-6.

...

additional arguments

Details

The function finds sparse directions for linear classification.

Value

Returns a list with the following attributes:

beta

The loadings of the sparse discriminative directions.

theta

The optimal scores.

rss

A vector of the Residual Sum of Squares at each iteration.

varNames

Names on included variables

.

Author(s)

Line Clemmensen, modified by Trevor Hastie

References

Clemmensen, L., Hastie, T. Witten, D. and Ersboell, K. (2011) "Sparse discriminant analysis", Technometrics, To appear.

See Also

normalize, normalizetest, smda

Examples

## load data
data(penicilliumYES)

X <- penicilliumYES$X
Y <- penicilliumYES$Y
colnames(Y) <- c("P. Melanoconidium",
                 "P. Polonicum",
                 "P. Venetum")

## test samples
Iout<-c(3,6,9,12)
Iout<-c(Iout,Iout+12,Iout+24)

## training data
Xtr<-X[-Iout,]
k<-3
n<-dim(Xtr)[1]

## Normalize data
Xc<-normalize(Xtr)
Xn<-Xc$Xc
p<-dim(Xn)[2]

## Perform SDA with one non-zero loading for each discriminative
## direction with Y as matrix input
out <- sda(Xn, Y,
           lambda = 1e-6,
           stop = -1,
           maxIte = 25,
           trace = TRUE)

## predict training samples
train <- predict(out, Xn)

## testing
Xtst<-X[Iout,]
Xtst<-normalizetest(Xtst,Xc)

test <- predict(out, Xtst)
print(test$class)

## Factor Y as input
Yvec <- factor(rep(colnames(Y), each = 8))
out2 <- sda(Xn, Yvec,
            lambda = 1e-6,
            stop = -1,
            maxIte = 25,
            trace = TRUE)