pESCA: Penalized exponential family simultaneous component analysis...

In YipengUva/RpESCA: R implementaion of pESCA models with various concave penalties

Description

This is the main function for construncting a pESCA model on multiple data sets. The potential different data types in these data sets are tackled by the assumption of exponential family distribution. Gaussian for quantitative data, Bernoulli for binary data and Poisson for count data. Although the option for count data using Poisson distribution is included in the algorithm, we recommend to do variance stabilizing transformation on the count data, such as Next-Gen sequencing data, and then use the transformed data as quantitative data sets. The details of the developed algorithm can be found in https://arxiv.org/abs/1902.06241.

Usage

pESCA(dataSets, dataTypes, lambdas, penalty = "L2",
  fun_concave = "gdp", opts = list())

Arguments

dataSets	a list contains multiple matrices with same number of rows. Each matrix (samples * variables) indicates a data set.
dataTypes	a string indicates the data types of the multiple data sets.
lambdas	a numeric vector indicates the values of tuning parameters for each data set.
penalty	The name of the penalty you want to used. "L2": group-wise concave L2 norm penalty; "L1": group-wise concave L1 norm penalty; "element": element-wise concave penalty; "composite": the composition of group- and element-wise penalty.
fun_concave	a string indicates the used concave function. Three options are included in the algorithm. "gdp": GDP penalty; "lq": Lq penalty; "scad": SCAD penalty.
opts	a list contains the options of the algorithms. tol_obj: tolerance for relative change of joint loss function, default:1E-6; maxit: max number of iterations, default: 1000; gamma: hyper-parameter of the concave penalty, default: 1; R: the initial number of PCs, default: 0.5 0.5*min(I,J); rand_start: initilization method, random (1), SCA(0), default: 0; alphas: dispersion parameters of exponential dispersion families, default: 1. thr_path: the option to generate thresholding path, default: 0; quiet: quiet==1, not show the progress when running the algorithm, default: 0.

Value

This function returns a list contains the results of a pESCA mdoel.

• mu: offset term;

• A: score matrix;

• B: loading matrix;

• S: group sparse pattern of B;

• varExpTotals: total variation explained of each data set and the full data set;

• varExpPCs: variation explained of each data set and each component;

• Sigmas: the group length (the definition depends on the used input type). Only meaningful for group-wise sparisty;

• iter: number of iterations;

• diagnose$hist_objs: the value of loss function pluse penalty at each iteration;

• diagnose$f_objs: the value of loss function at each iteration;

• diagnose$g_objs: the value of penalty function at each iteration;

• diagnose$rel_objs: relative change of diagnose$hist_obj at each iteration;

• diagnose$rel_Thetas: relative change of Theta at each iteration.

Examples

## Not run: 
# Suppose we have three data sets X1, X2, X3
# They are quantitative, quantitative and binary matrices
pESCA(dataSets = list(X1, X2, X3),
              dataTypes = 'GGB',
              lambdas = c(20, 20, 10),
              penalty = 'L2',
              fun_concave = 'gdp',
              opts = list())

## End(Not run)

YipengUva/RpESCA documentation built on July 2, 2019, 6:41 p.m.