k_binomialRF {binomialRF}

random forest feature selection based on binomial exact test

Description

k_binomialRF is the R implementation of the interaction feature selection algorithm by (Zaim 2019). k_binomialRF extends the binomialRF algorithm by searching for k-way interactions.

Usage

k_binomialRF(X, y, fdr.threshold = 0.05, fdr.method = "BY",
  ntrees = 2000, percent_features = 0.3, K = 2, cbinom_dist = NULL,
  sampsize = nrow(X) * 0.4)

Arguments

X

design matrix

y

class label

fdr.threshold

fdr.threshold for determining which set of features are significant

fdr.method

how should we adjust for multiple comparisons (i.e., p.adjust.methods =c("holm", "hochberg", "hommel", "bonferroni", "BH", "BY","fdr", "none"))

ntrees

how many trees should be used to grow the randomForest? (Defaults to 5000)

percent_features

what percentage of L do we subsample at each tree? Should be a proportion between (0,1)

K

for multi-way interactions, how deep should the interactions be?

cbinom_dist

user-supplied correlated binomial distribution

sampsize

user-supplied sample size for random forest

Value

a data.frame with 4 columns: Feature Name, Frequency Selected, Probability of Selecting it randomly, Adjusted P-value based on fdr.method

References

Zaim, SZ; Kenost, C.; Lussier, YA; Zhang, HH. binomialRF: Scalable Feature Selection and Screening for Random Forests to Identify Biomarkers and Their Interactions, bioRxiv, 2019.

Examples

set.seed(324)

###############################
### Generate simulation data
###############################

X = matrix(rnorm(1000), ncol=10)
trueBeta= c(rep(10,5), rep(0,5))
z = 1 + X %*% trueBeta
pr = 1/(1+exp(-z))
y = rbinom(100,1,pr)

###############################
### Run interaction model
###############################

require(correlbinom)

rho = 0.33
ntrees = 250
cbinom = correlbinom(rho, successprob =  calculateBinomialP_Interaction(10, .5,2), 
                               trials = ntrees, precision = 1024, model = 'kuk')

k.binom.rf <-k_binomialRF(X,y, fdr.threshold = .05,fdr.method = 'BY',
                      ntrees = ntrees,percent_features = .5,
                      cbinom_dist=cbinom,
                      sampsize=round(nrow(X)*rho))