SLS: Fit phylogeny-regularized GLM with Laplacian smoothness...

In lichen-lab/SICS: Phylogeny-Constrained Regularization for Sparse Generalized Linear Models

Description

Fit a generalized linear model at grids of tuning parameter via penalized maximum likelihood. The regularization path is computed for a combination of sparse and smoothness penalty. The sparse penalty is MCP penalty and the smoothness penalty is Laplacian penalty derived from the phylogeny-induced correlation structure (C) based on patristic distance between OTU (D) in the form of C=exp(-2*pho*D). pho characterizes the evolutionary rate and performs phylogenetic grouping. Large pho groups OTUs into clusters at lower phylogenetic depth. By sparsifying C, we obtain Laplacian matrix (L) with different level of sparsity. The Laplacian penalty is in the form of beta^T*L*beta. Fits linear, logistic regression models. The estimator is called Sparse Laplacian Srinkage Estimator (SLS).

Usage

SLS(X,Y,D,family=c("gaussian","binomial"),phos=c(2^(c(seq(-5,5,length=5)))),lambda2=c(0,2^(seq(-5, 5, length=5))),sparsity=0.9, nfolds=5)

Arguments

X	Input matrix; each row is an observation vector.
Y	Response vector. Quantitative for family="gaussian" or binary(0/1) for family="binomial".
D	patristic distance between OTU.
family	Either "gaussian", "binomial", depending on the response.
pho	A user-specified sequence of pho values. pho characterizes the evolutionary rate and performs phylogenetic grouping. The default value is (0,2^(seq(-5, 5, length=5))).
lambda2	A user-specified sequence of lambda2 values.lambda2 is the tuning parameter for smoothness penalty. The default value is (0,2^(seq(-5, 5, length=5))).
sparsity	The sparsity level to convert C to L. It is between 0 and 1.
nfolds	folds to perform cross-validation. Default value is 5.

Value

beta	A vector of fitted coefficients not include intercept

Author(s)

Li Chen <[e-mail]>, Jun Chen <[e-mail]>

lichen-lab/SICS documentation built on May 6, 2019, 7:18 a.m.