iRF: iteratively grows weighted random forests, finds stable...
In iRF: iterative Random Forests

Description Usage Arguments Value Author(s) See Also

Using repeated calls to iRF::randomForest, this function iteratively grows weighted ensembles of decision trees. Optionally, return stable feature interactions for select iterations by analyzing feature usage on decision paths of large leaf nodes. For details on the iRF algorithm, see https://arxiv.org/abs/1706.08457.

  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.id=1, class.cut=NULL), 
      varnames.grp=NULL,
      n.bootstrap=30,
      bootstrap.forest=TRUE, 
      verbose=TRUE, 
      ...
     )

`x, xtest`	numeric matrices of predictors
`y, ytest`	response vectors
`n.iter`	number of weighted random forest iterations
`ntree`	number of trees to grow in each iteration
`n.core`	number of cores across which tree growing and reading 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 evaluate 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 accuracy (`TRUE`) or just size (`FALSE`)?
`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
`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.id`: which 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. `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
`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

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

Sumanta Basu sumbose@berkeley.edu, Karl Kumbier kkumbier@berkeley.edu

randomForest, readForest

iRF documentation built on May 2, 2019, 11:02 a.m.