flsa {flsa}

Fused Lasso Signal Approximator

Description

These functions are the main interface functions for calculating FLSA solutions

Usage

flsa(y, lambda1=0, lambda2=NULL, connListObj = NULL, splitCheckSize=1e+09, 
    verbose=FALSE, thr = 1e-09, maxGrpNum=4*length(y))
flsaTopDown(y, lambda1=0, groups=1:length(y), lambda2=NULL)
flsaGetSolution(solObj, lambda1=0, lambda2=NULL, dim=NULL)

Arguments

y

response variable; numeric

lambda1

penalty parameter vector (non-negative) for the absolute values uf the coefficients; numeric

lambda2

penalty parameter vector (non-negative) for the difference of certain coefficients; numeric

groups

Return solutions for which the given number of groups is present - solutions found exactly at the breakpoint

connListObj

an object specifying which differences are to be penalized by lambda2. If NULL, then the dimensionalty of y is being used. If y is a vector, the differences of neighbouring coefficients are penalized. If y is a matrix, differences of neighbouring coefficients in 2 dimensions are being penalized. For more information see connListObj

splitCheckSize

a parameter specifying from which size on, groups of variables are not being checked for breaking up; can be used to reduce computation time; may lead to inaccurate results

solObj

Solution object as returned by FLSA if lambda2=NULL

dim

dimensions how the result should be formatted for a specific lambda. Used to format the 2-dimensional FLSA as a matrix in the response. For this, just include the dimensions of y as dim

verbose

print status messages during fitting

thr

the error threshold used in the algorithm

maxGrpNum

if every step of the algorithm, a group with a higher number is generated; this limits the number of steps the algorithm can take

Details

flsa is the main function for calculate a flsa fit. If lambda2=NULL, then it returns an object that encodes the whole solution path. Solutions for specific values of lambda1 and lambda2 can then be obtained by using flsaGetSolution.

flsaTopDown calculates the solution of the 1-dimensional FLSA, starting at large values of lambda2. If only solutions for large values of lambda2 are needed, this is more efficient.

Author(s)

Holger Hoefling

See Also

connListObj

Examples

library(flsa)
# generate some artificial data, 1 and 2 dimensional
y <- rnorm(100)
y2Dim = matrix(rnorm(100), ncol=10)

### apply function flsa and get solution directly
lambda2= 0:10/10
res <- flsa(y, lambda2=lambda2)
res2Dim <- flsa(y2Dim, lambda2=lambda2)

### apply the function and get the solution later
resSolObj <- flsa(y, lambda2=NULL)
resSolObjTopDown <- flsaTopDown(y)
resSolObj2Dim <- flsa(y2Dim, lambda2=NULL)

res2 <- flsaGetSolution(resSolObj, lambda2=lambda2)
### here note that the solution object does not store that the input was 2 dimensional
### therefore, res2Dim does not give out the solution as a 2 
### dimensional matrix (unlike the direct version above)
res2Dim2 <- flsaGetSolution(resSolObj2Dim, lambda2=lambda2)