dataslingers / exclusivelasso Goto Github PK

It is interesting and exciting to see exclusivelasso package. It is helpful for the variable selection. I am a clinic researchers. Cox model is common and important to analyze the variable related to sample's survival status and time. So, I want to ask if the exclusivelasso package can apply to the cox model. I also scan the code of the function of the package. But it's difficult for me to understand because the programming ability.
It's very glad for your future reply.

It's not necessary to implement the full set of GLMs supported by glmnet, but we should have the "big three" of Gaussian, Logistic, and Poisson. No plans to implement Cox PH, multinomial and multi-response Gaussian for now.

• Coordinate Descent Algorithm (nested within IRLS like glmnet?)
• Prox Gradient Algorithm (will probably need back-tracking, so need to make sure prox is optimized)
• Methods for deviance function
• Other CV loss functions

By default, Rcpp transfers the RNG state to and from R when calling into C++. This is a little bit expensive and not necessary for us since we don't use RNGs in C++.

If we change our C++ attributes to // [[Rcpp::export(rng = false)]], things will be a smidge faster. (I'd imagine this is still dwarfed by compute time, but it can't hurt.)

See example at https://github.com/tidyverse/dplyr/blob/e08f0511986a10efda5b9fd991af2e049d801333/src/address.cpp#L18

The exclusive lasso proximal operator is relatively expensive to calculate, so an accelerated prox gradient scheme might be useful.

As discussed at [1], uint is not standard C++. We should use arma::uword consistently instead.

There's also a signed/unsigned comparison at [2] that should be fixed. I think switching to arma::uvec should suffice.

[1] https://stackoverflow.com/q/3552094
[2]

Multinomial and Multi-Response regression (with a group-lasso penalty)

Upper and lower bounds on each parameter. This should be a pretty minor algorithmic tweak: for CD, apply the constraint at each update; for PG, incorporate the constraint into the prox (itself a CD problem).

I am running into the error of

Error: Mat::init(): requested size is too large; suggest to enable ARMA_64BIT_WORD

I understand that this is due to large matrix error which should be the issue from the RcppArmadillo package. The sample code that run into issue as below.

library(glmnet)
library(ExclusiveLasso)

N = 80000 # number of observations
p = 35  # number of variables

# random generated X
X = matrix(rnorm(N*p), ncol=p)

# standardization : mean = 0, std=1
X = scale(X)

# artificial coefficients
beta = c(0.15,-0.33,0.25,-0.25,0.05,0,0,0,0.5,0.2,
         0.15,-0.33,0.25,-0.25,0.05,0,0,0,0.5,0.2,
         0.15,-0.33,0.25,-0.25,0.05,0,0,0,0.5,0.2,
         1, -0.2, 0.2, 0.1, 0.5)

# Y variable, standardized Y
y = X%*%beta + rnorm(N, sd=0.5)
#y = scale(y)

# group index for X variables
v.group <- rep(1:10, length.out = 35 )

#--------------------------------------------
# Model with a given lambda
#--------------------------------------------
# exclusive lasso
ex <- exclusive_lasso(X, y,lambda = 0.2, 
                      groups = v.group, family="gaussian", 
                      intercept = F)

Currently, if lambda is not supplied by the user, we use the same range as glmnet (more or less).

Does theory suggest a better default? Ideally, we want lambda_max to be the smallest value of lambda that gives exactly one non-zero element in each group, but there are problems for which no such lambda exists and even when it does, it's not clear how to calculate it quickly.

See DataSlingers/clustRviz#86 and DataSlingers/clustRviz#87

Recent versions of Armadillo (and hence RcppArmadillo) come with OpenMP support, used to speed
up certain expensive elementwise operators. Compiler support for OpenMP is inconsistent (particularly on MacOS) so we should disable this for now and re-enable it when the upstream fix is on CRAN.[1]

• Disable OpenMP for now
• Add dependency on recent RcppArmadillo when released to CRAN
• Re-enable OpenMP (conditionally)

[1] See RcppCore/RcppArmadillo#177 and RcppCore/RcppArmadillo#185, especially RcppCore/RcppArmadillo@dc294cb

Section 7 of Campbell and Allen (2017) contains a (simulated) NMR example data set.

This should be included and used in documentation / examples.

Add an option for partial L2 regularization (similar to glmnet's alpha argument) -- should only be a minor modification of the prox and CD steps.

using

library(devtools)
install_github("DataSlingers/ExclusiveLasso")

and it gives:

In file included from D:/R-4.2.2/library/Rcpp/include/RcppCommon.h:30,
                 from D:/R-4.2.2/library/RcppArmadillo/include/RcppArmadillo/interface/RcppArmadilloForward.h:25,
                 from D:/R-4.2.2/library/RcppArmadillo/include/RcppArmadillo.h:29,
                 from ExclusiveLasso.cpp:23:
D:/R-4.2.2/library/Rcpp/include/Rcpp/r/headers.h:66:10: fatal error: R.h: No such file or directory
   66 | #include <R.h>
      |          ^~~~~
compilation terminated.
make: *** [D:/R-4.2.2/etc/x64/Makeconf:260：ExclusiveLasso.o] 错误 1
ERROR: compilation failed for package 'ExclusiveLasso'
* removing 'D:/R-4.2.2/library/ExclusiveLasso'

i'm using r 4.2.2, win11, vscode

Is it possible to derive (safe or strong) screening rules for the Exclusive Lasso?