gRIT: Generalized random intersection trees

In karlkumbier/iRF2.0: Iterative Random Forests

Generalized random intersection trees

Description

Run RIT across decision paths of a fitted random forest.

Usage

gRIT(
  x,
  y,
  rand.forest = NULL,
  read.forest = NULL,
  rit.param = list(depth = 5, ntree = 500, nchild = 2, class.id = 1, min.nd = 1,
    class.cut = NULL),
  varnames.grp = colnames(x),
  weights = rep(1, nrow(x)),
  signed = TRUE,
  oob.importance = TRUE,
  ints.idx.eval = NULL,
  ints.eval = NULL,
  n.core = 1
)

Arguments

x	numeric feature matrix
y	response vector. If factor, classification is assumed.
rand.forest	an object of class randomForest. Required if read.forest is NULL.
read.forest	output of readForest. Required if rand.forest is NULL.
rit.param	named list specifying RIT parameters. Entries include depth: depths of RITs, ntree: number of RITs, nchild: number of child nodes for each RIT, class.id: 0-1 indicating which leaf nodes RIT should be run over, min.nd: minimum node size to run RIT over, class.cut: threshold for converting leaf nodes in regression to binary classes.
varnames.grp	grouping "hyper-features" for RIT search. Features with the same name will be treated as identical for interaction search.
weights	numeric weight for each observation. Leaf nodes will be sampled for RIT with probability proprtional to the total weight of observations they contain.
signed	if TRUE, signed interactions will be returned
oob.importance	if TRUE, importance measures are evaluated on OOB samples.
ints.idx.eval	like ints.eval, but specifies the indice of the interactions instead of their names. Intended for internal use only.
ints.eval	interactions to evaluate. If specified, importance metrics will be evaluated for these interactions instead of those recovered by RIT.
n.core	number of cores to use. If -1, all available cores are used.

Value

a data table containing the recovered interactions and importance scores.

karlkumbier/iRF2.0 documentation built on April 12, 2025, 1:44 p.m.