MSGLasso {MSGLasso}R Documentation

A function to fit the Multivariate Sparse Group Lasso with an arbitrary group structure (MSGLasso)

Description

A function to fit the Multivariate Sparse Group Lasso with an arbitrary group structure using the mixed coordinate descent algorithm.

Usage

MSGLasso(X.m, Y.m, grp.WTs, Pen.L, Pen.G, PQ.grps, GR.grps, grp_Norm0,
 lam1, lam.G, Beta0 = NULL)

Arguments

X.m

numeric predictor matrix (n by p): columns correspond to predictor variables and rows correspond to samples. Missing values are not allowed.

Y.m

numeric predictor matrix (n by q): columns correspond to response variables and rows correspond to samples. Missing values are not allowed.

grp.WTs

user specified adaptive group weighting matrix of g by r, for putting different penalization levels on different groups. Missing values are not allowed.

Pen.L

user specified single-entry level penalization indictor mateix of p by q. 1 for being penalized and 0 for not. Missing values are not allowed.

Pen.G

user specified group level penalization indictor mateix of g by r. 1 for being penalized and 0 for not. Missing values are not allowed.

PQ.grps

the group attributing matrix of (p+q) by (gmax+1), where gmax is max number of different groups a single variable belongs to. Each row corresponds to a (predictor or response) varaible, and starts with group indexes the variable belongs to and followed by 999.

GR.grps

the variable attributing matrix of (g+r)*(cmax+1), where cmax is max number of variables a single group contains. Each row corresponds to a (predictor or response) group, and starts with variable indexes the group contains to and followed by 999.

grp_Norm0

a numeric matrix (g by r) containing starting L2 group norm values. Should be calculated from the Beta starting value matrix Beta0.

lam1

lasso panelty parameter scaler.

lam.G

group penalty parameter matrix (g by r).

Beta0

numeric matrix (p by q) containing starting beta values. By defalt use a zero matrix.

Details

Uses the mixed coordinate descent algorithm for fitting the multivariate sparse group lasso in a multivariate-response-multiple-predictor linear regression setting, with an arbitrary group structure on the regression coefficient matrix (Li, Nan and Zhu 2014).

Value

A list with five components:

Beta

the estimated regression coefficient matrix (p by q).

grpNorm

the L2 group norm matrix (g by r) of the estimated regression coefficient matrix.

E

residual matrix (n by q).

rss.v

a vector of length q recording the resisual sum square for each of the q responses.

rss

a scaler of overall residual sum of square.

iter

a positive interger recording the number of iterations till convergence.

Author(s)

Yanming Li, Bin Nan, Ji Zhu

References

Y. Li, B. Nan and J. Zhu (2015) Multivariate sparse group lasso for the multivariate multiple linear regression with an arbitrary group structure. Biometrics. DOI: 10.1111/biom.12292

Examples


#####################################################
# Simulate data
#####################################################

set.seed(sample(1:100,1))
G.arr <- c(0,20,20,20,20,20,20,20,20,20,20)

data("Beta.m")

######## generate data set for model fitting

simDataGen<-function(N, Beta, rho, s, G.arr, seed=1){

P<-nrow(Beta)
Q<-ncol(Beta)
gsum<-0
X.m<-NULL

set.seed(seed)

Sig<-matrix(0,P,P)
jstart <-1

for(g in 1:length(G.arr)-1){
X.m<-cbind(X.m, matrix(rnorm(N*G.arr[g+1]),N,G.arr[g+1], byrow=TRUE))

for(i in 2:P){ for(j in jstart: (i-1)){

    Sig[i,j]<-rho^(abs(i-j))

    Sig[j,i]<-Sig[i,j]

}}
jstart <- jstart + G.arr[g+1]
}


diag(Sig)<-1
R<-chol(Sig)

X.m<-X.m%*%R

SVsum <-0

for (q in 1:Q){SVsum <-SVsum+var(X.m %*% Beta[,q])}
sdr =sqrt(s*SVsum/Q)

E.m <- matrix(rnorm(N*Q,0,sdr),N, Q, byrow=TRUE)

Y.m<-X.m%*%Beta+E.m

return(list(X=X.m, Y=Y.m, E=E.m))
}

N <-150

rho=0.5; 
s=4;

Data <- simDataGen(N, Beta.m,rho, s, G.arr, seed=sample(1:100,1))
X.m<-Data$X
Y.m<-Data$Y


############################################################
## fit model for one set of (lam1, lam.G) using example data
############################################################

P <- dim(Beta.m)[1]
Q <- dim(Beta.m)[2]
G <- 10
R <- 10

gmax <- 1
cmax <- 20
GarrStarts <- c(0,20,40,60,80,100,120,140,160,180)
GarrEnds <- c(19,39,59,79,99,119,139,159,179,199)
RarrStarts <- c(0,20,40,60,80,100,120,140,160,180)
RarrEnds <- c(19,39,59,79,99,119,139,159,179,199)

tmp <- FindingPQGrps(P, Q, G, R, gmax, GarrStarts, GarrEnds, RarrStarts, RarrEnds)
PQgrps <- tmp$PQgrps

tmp1 <- Cal_grpWTs(P, Q, G, R, gmax, PQgrps)
grpWTs <- tmp1$grpWTs

tmp2 <- FindingGRGrps(P, Q, G, R, cmax, GarrStarts, GarrEnds, RarrStarts, RarrEnds)
GRgrps <- tmp2$GRgrps

Pen_L <- matrix(rep(1,P*Q),P,Q, byrow=TRUE)
Pen_G <- matrix(rep(1,G*R),G,R, byrow=TRUE)
grp_Norm0 <- matrix(rep(0, G*R), nrow=G, byrow=TRUE)

MSGLassolam1 <- 1.6
MSGLassolamG <- 0.26
MSGLassolamG.m <- matrix(rep(MSGLassolamG, G*R),G,R,byrow=TRUE)
 
system.time(try <-MSGLasso(X.m, Y.m, grpWTs, Pen_L, Pen_G, PQgrps, GRgrps, grp_Norm0, 
   MSGLassolam1, MSGLassolamG.m, Beta0=NULL))

## Not run: 
################################################
## visulizing model fitting results
################################################

########visulizing selection effect using heatmaps

MYplotBW <- function(Beta){
colorNP <- ceiling(abs(max(Beta)))+2
ColorValueP <- colorRampPalette(c("gray50", "black"))(colorNP)
colorNN <- ceiling(abs(min(Beta)))+2
ColorValueN <- colorRampPalette(c("gray50", "white"))(colorNN)
P <-nrow(Beta)
Q <-ncol(Beta)
Xlim <- c(0,2*(P+1))
Ylim <- c(0,2*(Q+1))
plot(0, type="n", xlab="", ylab="", xlim=Xlim, ylim=Ylim, cex.lab=1.0, 
  bty="n", axes=FALSE)
for (p in 1:P){
for (q in 1:Q){
        k0 <- Beta[p,q]
if(k0==0){
      rect(2*(P-p+1)-1,2*(Q-q+1)-1, 2*(P-p+1)+1, 2*(Q-q+1)+1, col="white", border=NA)
}
if(k0>0){
      k <- ceiling(k0)+1
      if(k>2) {k <- k+1}
          rect(2*(P-p+1)-1,2*(Q-q+1)-1, 2*(P-p+1)+1, 2*(Q-q+1)+1, 
               col="black", border=NA)
}
if(k0<0){
      k <- ceiling(abs(k0))+1
      if(k>2) {k <- k+1}
          rect(2*(P-p+1)-1,2*(Q-q+1)-1, 2*(P-p+1)+1, 2*(Q-q+1)+1, 
               col="black", border=NA)
}
}
}

rect(1,1,2*P, 2*Q, lty=2)
}

MYplotBW(try$Beta)

rect(1,1,40,40, lty=2)
rect(41,41,80,80, lty=2)
rect(81,81,120,120, lty=2)
rect(121,121,160,160, lty=2)
rect(161,161,200,200, lty=2)
rect(201,201,240,240, lty=2)
rect(241,241,280,280, lty=2)
rect(281,281,320,320, lty=2)
rect(361,1,400,400, lty=2)

######## visulizing the true Beta matrix

#X11()

MYplotBW(Beta.m)

rect(1,1,40,40, lty=2)
rect(41,41,80,80, lty=2)
rect(81,81,120,120, lty=2)
rect(121,121,160,160, lty=2)
rect(161,161,200,200, lty=2)
rect(201,201,240,240, lty=2)
rect(241,241,280,280, lty=2)
rect(281,281,320,320, lty=2)
rect(361,1,400,400, lty=2)


## End(Not run)

[Package MSGLasso version 2.1 Index]