sparseBC (version 1.2)

sparseBC.choosekr: Do tuning parameter K and R selection for sparse biclustering via cross-validation.

Description

Perform cross-validation to select K (number of row clusters) and R (number of column clusters) for sparse biclustering. We assume that lambda is known.

Usage

sparseBC.choosekr(x, k, r, lambda, percent = 0.1, trace=FALSE)

Arguments

x

Data matrix; samples are rows and columns are features. Cannot contain missing values.

k

A range of values of K to be considered. Values considered must be an increasing sequence.

r

A range of values of R to be considered. Values considered must be an increasing sequence.

lambda

Non-negative regularization parameter for lasso. lambda=0 means no regularization.

percent

Percentage of elements of x to be left out for cross-validation. 1 must be divisible by the specified percentage. The default value is 0.1.

trace

Print out progress as iterations are performed. Default is FALSE.

Value

estimated_kr

The chosen values of K and R based on cross-validation.

results.mean

Mean squared error for all values of K and R considered.

results.se

Standard error of the sum of squared error for all values of K and R considered.

Details

The function performs cross-validation as described in Algorithm (2) in Tan and Witten (2014) 'Sparse biclustering of transposable data'. Briefly, it works as follows: (1) some percent of the elements of x is removed at random from the data matrix - call those elements missing elements, (2) the missing elements are imputed using the mean of the other elements of the matrix, (3) sparse biclustering is performed with various values of K and R and the mean matrix is estimated, (4) calculate the sum of squared error of the missing values between x and the estimated mean matrix. This procedure is repeated 1/percent times. Finally, we select K and R based on the criterion described in Algorithm (2) in Tan and Witten (2014). A similar procedure is used in Witten et al (2009) 'A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis'.

Note that sparseBC is run with center=TRUE in this function.

References

KM Tan and D Witten (2014) Sparse biclustering of transposable data. Journal of Computational and Graphical Statistics 23(4):985-1008.

D Witten, R Tibshirani, and T Hastie (2009) A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis, Biostatistics 10(3), 515--534.

See Also

sparseBC sparseBC.BIC

Examples

# NOT RUN {
########### Create data matrix with K=2 R=4 row and column clusters 
#k <- 2
#r <- 4
#n <- 200
#p <- 200
#  mus<-runif(k*r,-3,3)
#  mus<-matrix(c(mus),nrow=k,ncol=r,byrow=FALSE)
#  truthCs<-sample(1:k,n,rep=TRUE)
#  truthDs<-sample(1:r,p,rep=TRUE)
#  x<-matrix(rnorm(n*p,mean=0,sd=2),nrow=n,ncol=p)
#  for(i in 1:max(truthCs)){
#     for(j in 1:max(truthDs)){ 
#         x[truthCs==i, truthDs==j] <- x[truthCs==i, truthDs==j] + mus[i,j]
#    }
#  }
#  x<-x-mean(x)

# Example is commented out for short run-time	
############ Perform sparseBC.choosekr to choose the number of row and column clusters
#sparseBC.choosekr(x,1:5,1:5,0,0.2)$estimated_kr

# }