PO.EN: A robust presence-only model with Elastic Net penalty

In PO.EN: An Elastic-Net Regularized Presence-Only Model

Description

Fit a logistic regression with presence-only response via penalized maximum likelihood. The regularization path is computed for the elastic-net penalty at a pair values of lambda and the prevalence parameter.

Usage

PO.EN(x,y,o.iter=5, i.iter=5, lambda=.01,alpha=.5,
true.prob=0.5,beta_start,epsilon=1e-4, gram.input=FALSE,XtX.input=0,
ytx.input=0,XtX_reduce.input)

Arguments

x	Input design matrix. Should not include the intercept vector.
y	Response variable. Should be a binary vector, such that 0 represents background observations and 1 represents presence observations.
o.iter	Number of outer loop iteration.
i.iter	Number of inner loop iteration.
lambda	A user supplied Elastic Net penalty parameter.
alpha	The elastic net mixing parameter, where 0≤alpha≤ 1.
true.prob	The prevalence parameter, should be provided by users. Can be tuned in the cross-validation function.
beta_start	A user supplied starting coefficients vector.
epsilon	The threshold for stopping the coordinate descent algorithm.
gram.input	The function allows users to feed the gram matrix for fasting computation. The default setting is False, and the function compute the gram matrix for computation.
XtX.input	If gram.input is TRUE, users should supply the corresponding gram matrix X'X.
ytx.input	If gram.input is TRUE, users should supply the product of y'X.
XtX_reduce.input	If gram.input is TRUE, users should supply a matrix of X'X without the diagnol entries.

Details

The function fits a presence-only model with an elastic net penalty.

Value

beta The fitting vector of the coefficients, the intercept is included.

Examples

data(example.data) # example datasets, including training dataset and testing dataset
train_data<-example.data$train.data
y_train=train_data$response;x_train=train_data[,-1]  # response and design matrix of training data
test_data<-example.data$test.data
y_test=test_data$response;x_test=test_data[,-1]  # response and design matrix of testing data
PO.EN.beta<-PO.EN(x_train,y_train,lambda=0.1,
           true.prob=sum(y_train)/length(y_train),beta_start=rep(0,ncol(x_train)+1))
predictions<-PO.EN.predict(x_test,PO.EN.beta)
pROC::roc(y_test~predictions)

PO.EN documentation built on Aug. 19, 2020, 9:06 a.m.