Create controls for greedy concentration-index trees

Usage

ci_tree_control(
  minsplit = 200L,
  minbucket = 100L,
  minprob = 0.01,
  maxdepth = 4L,
  min_gain = 0,
  min_relative_gain = 0,
  mtry = NULL,
  split_engine = c("cpp", "R"),
  factor_split = c("auto", "partition", "order"),
  max_factor_levels_partition = 10L,
  max_factor_partitions = 1e+06
)

Arguments

minsplit: Minimum weighted node size required before attempting a split.
minbucket: Minimum weighted child-node size.
minprob: Minimum child-node weight proportion.
maxdepth: Maximum tree depth, with the root at depth 0.
min_gain: Minimum concentration-index gain required to split a node.
min_relative_gain: Minimum concentration-index gain as a share of the parent-node impurity required to split a node.
mtry: Optional number of variables sampled as candidates at each node.
split_engine: Split-search implementation to use. "cpp" uses the Rcpp-backed split search, while "R" uses the reference R implementation.
factor_split: Strategy for unordered factor split search. "partition" evaluates binary partitions, "order" orders levels by weighted mean outcome and tries cumulative splits, and "auto" uses partition search when it is small enough and ordered search otherwise.
max_factor_levels_partition: Maximum number of observed factor levels eligible for partition search when factor_split = "auto".
max_factor_partitions: Maximum number of binary factor partitions eligible for partition search when factor_split = "auto".

Value

A list of greedy CI tree controls.

Details

Structural controls minsplit, minbucket, minprob, maxdepth, and min_relative_gain must be finite. min_gain = Inf is allowed and prevents splitting.