iRF: iteratively grows weighted random forests, finds stable...

In Jromero1208/PBDRiRF: iterative Random Forests

Description

Using repeated calls to iRF::randomForest, this function iteratively grows weighted ensembles of decision trees. Optionally, for every iteration, returns stable feature interactions by analyzing feature usage on decision paths of large leaf nodes. Implemented only for binary classification with numeric predictors and response taking values in 0,1.

Usage

  iRF(x, y, xtest=NULL, ytest=NULL, 
      n.iter=5, 
      ntree=500, 
      n.core=1,
      mtry.select.prob = rep(1/ncol(x), ncol(x)),
      keep.impvar.quantile=NULL, 
      interactions.return=NULL,
      wt.pred.accuracy=FALSE, 
      cutoff.unimp.feature = 0,
      rit.param=list(depth=5, ntree=100, nchild=2, 
                     class.cut=NULL, class.id=1), 
      varnames.grp=NULL,
      n.bootstrap=30,
      bootstrap.forest=TRUE, 
      verbose=TRUE, ...
     )

Arguments

x, xtest	numeric matrices of predictors
y, ytest	factor with two levels: 0, 1
n.iter	number of weighted random forest fits
ntree	number of trees to grow in each iteration
n.core	number of cores across which tree growing should be distributed
mtry.select.prob	initial weights specified for first random forest fit, defaults to equal weights
keep.impvar.quantile	a nonnegative fraction q. If provided, all the variables with Gini importance in the top 100*q percentile are retained during random splitting variable selection in the next iteration
interactions.return	a numeric vector specifying which iterations to calculate interactions for. Note: interaction computation is computationally intensive particularly when n.bootstrap is large.
wt.pred.accuracy	Should leaf nodes be sampled proportional to both size and decrease in variabiliy of responses?
cutoff.unimp.feature	a non-negative fraction r. If provided, only features with Gini importance score in the top 100*(1-r) percentile are used to find feature interactions
class.id	which class of observations will be used to find class-specific interaction? Choose between 0 or 1. Default is set to 1.
rit.param	a named list, containing entries to specify depth: depth of random intersection trees ntree number of random intersection trees nchild: number of children in each split of a random intersection tree class.cut: threshold to determine leaf nodes that are used to find interactions. Any leaf nodes with prediction greater than specified threshold will be used as input to RIT. If NULL, all leaf nodes from regression iRF will be used as input to RIT. Ignored if classification forest

class.idwhich class of observations will be used to find class-specific interaction? Choose between 0 or 1. Default is set to 1. Ignored if regression forest.

varnames.grp	If features can be grouped based on their demographics or correlation patterns, use the group of features or “hyper-feature”s to conduct random intersection trees
n.bootstrap	Number of bootstraps replicates used to calculate stability scores of interactiosn obtained by RIT
bootstrap.forest	Should a new Random Forest be constructed for each bootstrap sample to evaluate stability? Setting to FALSE results in faster runtime.
verbose	Display progress messages and intermediate outputs on screen?
...	additional arguments passed to iRF::randomForest

Value

A list containing the following items:

rf.list	A list of n.iter objects of the class randomForest
interaction	A list of length n.iter. Each element of the list contains a named numeric vector of stability scores, where the names are candidate interactions (feature names separated by "_"), defined as frequently appearing features and feature combinations on the decision paths of large leaf nodes

See Also

randomForest, readForest

Jromero1208/PBDRiRF documentation built on May 4, 2019, 10:56 a.m.