ddegras/sparsegfl

Hybrid Approach to Sparse Group Fused Lasso

sparseGFL

Hybrid Approach to Sparse Group Fused Lasso

Fast hybrid algorithm for joint model segmentation and sparse estimation in high-dimensional piecewise linear regression models.
The algorithm finds global solutions to the sparse group fused lasso (SGFL) problem whose objective function is the sum of a squared loss function (to control fit to the data), an elastic net penalty (to promote sparsity in regression coefficients), and a total variation penalty (to promote parsimony in detected change points or segments).

Package installation

install.packages(c("foreach", "glmnet", "Matrix", "Rcpp", "RcppArmadillo")) # package dependencies and imports
install.packages("devtools")
library(devtools)  
install_github("ddegras/sparseGFL") 

Main functions

• SGFL: solve SGFL for piecewise regression model y(t) = X(t) b(t) + e(t) (matrix predictors X, vector regression coefficients b)
• SGFL.AXY: solve SGFL for piecewise regression model y(t) = A(t) x(t) + e(t) (vector predictors x, matrix regression coefficients A)
• SGFL.VAR: solve SGFL for piecewise vector autoregressive (VAR) model y(t) = A1(t) y(t-1) + ... + Ap(t) y(t-p) + e(t)

Reference

Degras, D. "Sparse group fused lasso for model segmentation: a hybrid approach." Advances in Data Analysis and Classification 15, 625–671 (2021).

sparsegfl's Issues

Documentation

I've tested out your spatiotemporal lasso, and it is fantastically fast. Two questions:

1. Using the below code, how does one actually plot out the results? I see the beta matrix there, in the results, after this code, but it's not conformable with the X matrix, so unlike the genlasso package... I can't exactly plot the results.
   `
   m <- 5 # number of (time-varying) observation variables
   n <- 100 # time series length
   p <- 10 # number of (time-varying) predictor variables

x <- array(rnorm(mnp),c(m,p,n)) # random coefficients
beta <- matrix(0,p,n) # random coefficients with 50% sparsity
beta[1:5,1:50] <- runif(5) # and one change point at time 51
beta[6:p,51:n] <- runif(5)
y <- matrix(,m,n)
for (i in 1:n){ y[,i] <- x[,,i] %*% beta[,i] + rnorm(m,sd=.01)}

lambda1 <- .05
lambda2 <- 1.5
alpha <- .9

result <- SGFL(x, y, lambda1, lambda2, alpha,
intercept=FALSE, verbose=TRUE)

`
2. Is there an example that uses parallelization? Does one do this with SNOW?

type definitions not known

What are the definitions of the types
mat (a 2 dimensional array?)
cube (a 3 dimensional array?)
uword (an integer?)
in the allmethods.cpp?