iterativehardthresholding's Introduction

IterativeHardThresholding

IterativeHardThresholding is part of the HADES.

Introduction

IterativeHardThresholding is an R package for performing L_0-based regressions using Cyclops

Features

Examples

Cox's Proportional Hazards Model

library(Cyclops)
library(IterativeHardThresholding)
library(survival)

## data dimension
p <- 20    # number of covariates
n <- 300   # sample size

## Cox model parameters
true.beta <- c(1, 0.1, 0, -1, 1, rep(0, p - 5))

## simulate data from an exponential model
x        <- matrix(rnorm(p * n, mean = 0, sd = 1), ncol = p)
ti       <- rweibull(n, shape = 1, scale = exp(-x%*%true.beta))
ui       <- runif(n, 0, 10) # Controls censoring
ci       <- rweibull(n, shape = 1, scale = ui * exp(-x%*%true.beta))
survtime <- pmin(ti, ci)
delta    <- ti == survtime; mean(delta)

cyclopsData <- createCyclopsData(Surv(survtime, delta) ~ x, modelType = "cox")
ihtPrior    <- createIhtPrior(K = 3, penalty = "bic")

cyclopsFit <- fitCyclopsModel(cyclopsData,
                             prior = ihtPrior)
coef(cyclopsFit)

Generalized Linear Model

library(Cyclops)
library(IterativeHardThresholding)

## data dimension
p <- 20    # number of covariates
n <- 300   # sample size

## logistic model parameters
itcpt     <- 0.2 # intercept
true.beta <- c(1, 0.3, 0, -1, 1, rep(0, p - 5))

## simulate data from logistic model
x <- matrix(rnorm(p * n, mean = 0, sd = 1), ncol = p)
y <- rbinom(n, 1, 1 / (1 + exp(-itcpt - x%*%true.beta)))


# fit BAR model
cyclopsData <- createCyclopsData(y ~ x, modelType = "lr")
ihtPrior    <- createIhtPrior(K  = 3, penalty = "bic", exclude = c("(Intercept)"))

cyclopsFit <- fitCyclopsModel(cyclopsData,
                             prior = ihtPrior)
coef(cyclopsFit)

Technology

System Requirements

Requires R (version 3.2.0 or higher).

Dependencies

Cyclops

Getting Started

On Windows, make sure RTools is installed.
In R, use the following commands to download and install IterativeHardThresholding:

install.packages("Cyclops")
install.packages("IterativeHardThresholding")

To perform a L_0-based Cyclops model fit with IHT, use the following commands in R:

library(IterativeHardThresholding)
cyclopsData <- createCyclopsData(formula, modelType = "modelType") ## TODO: Update
ihtPrior    <- createIhtPrior(K = 5, penalty = "bic")
cyclopsFit  <- fitCyclopsModel(cyclopsData, prior = ihtPrior)
coef(cyclopsFit) #Extract coefficients

Getting Involved

Package manual: IterativeHardThresholding manual
Developer questions/comments/feedback: OHDSI Forum
We use the GitHub issue tracker for all bugs/issues/enhancements

License

IterativeHardThresholding is licensed under Apache License 2.0.

Development

IterativeHardThresholding is being developed in R Studio.

Acknowledgments

This project is supported in part through the National Institutes of Health grant R01 HG006139.

iterativehardthresholding's People

Stargazers

Watchers

iterativehardthresholding's Issues

Recommended changes after thesis research

I would like to propose some changes after I have experimented with IHT for my master's thesis. I have tried to commit the changes to another branch, but this failed (I didn't seem to have access).

The most important change is the default value for the initialRidgeVariance (Prior.R, line 53). This is set to be 1E4. This caused an inability to converge for my PLP model. I recommend changing it to 1E-1.

Secondly, the todo's for step-halving can be removed. It did not produce any desirable effects.

Lastly, in my thesis, I changed the prior type for the IHT starting prior to "laplace" (Prior.R, line 209). The default value for the initialRidgeVariance was determined with this prior type (instead of "normal"). Naturally, this means that it is not the ridge variance (but that is just a name and doesn't matter as much), but it might also mean that another default value is better when the starting prior is "normal". Whether others want to continue using "laplace" or "normal" is up to them. I used "laplace" as I read in the literature that some IHT like algorithms were initiated with Lasso, but just "normal" could also be interesting.

ohdsi / iterativehardthresholding Goto Github PK