Select chain responses by a shadow comparison

Select chain-feature responses whose supplied importance is strictly greater than a same-row shadow importance in more than half of their rows. Use this table-only gate when each row contains the feature and shadow scores to be compared directly.

Usage

activeByShadow(IMP, suffix = ".o")

Arguments

IMP

A data.table with one row per comparison cell and required columns feature, normImportance, and shadowImportance. feature must contain non-missing character names ending in suffix; both importance columns must be numeric on the same normalized-percentage scale. Extra columns are ignored. Values and row uniqueness are not validated.

suffix

A single non-missing, non-empty character suffix to remove from every feature, default ".o". Removal is positional: the function does not verify that names end in this suffix.

Value

A character vector of unique stripped response names in first-appearance order. It is empty when no response has a strict majority.

Gate definition

Each input row is one comparison cell; method, response, and dataset columns are not inspected. A row is a success only when shadowImportance is not missing and normImportance > shadowImportance. Equality is not a success. For each stripped response name, the denominator is the full row count, including duplicate rows and rows with a missing shadow score. The response passes only when successes are strictly greater than half of that count, so a tied half does not pass.

A missing shadowImportance counts as an unsuccessful row and remains in the denominator. A missing normImportance paired with a present shadow makes the group's success count NA, so that group is excluded. The helper assumes that the two score columns are comparable within each row; it does not establish comparability across learners or sources.

This strict-majority rule is ssel package policy. It uses a supplied shadow score as a reference but does not implement the iterative random-probe tests of the Boruta algorithm.

Validation and side effects

NULL, a zero-row table, or an input without shadowImportance returns character(0). Otherwise a non-empty input must be a data.table; missing feature or normImportance columns and incompatible types produce underlying base R or data.table errors or warnings. The function does not modify IMP by reference. It reads and writes no files, uses no random numbers, and emits no message or warning for valid inputs.

References

Kursa, M. B. and Rudnicki, W. R. (2010). Feature Selection with the Boruta Package. Journal of Statistical Software, 36(11), 1–13. doi:10.18637/jss.v036.i11 .

See also

activeByImportance() for the inclusive fixed-threshold gate. extractChainImportance() is the package table producer; its model-file contract is separate from this table operation.

Examples

importance <- data.table::data.table(
  feature = c("alpha.o", "alpha.o", "alpha.o",
              "beta.o", "beta.o", "gamma.o", "gamma.o"),
  normImportance = c(2, 3, 0, 2, 0, 2, 2),
  shadowImportance = c(1, 1, 1, 1, 1, 1, NA)
)

# alpha wins 2/3 cells; beta ties 1/2; gamma's missing shadow counts.
activeByShadow(importance, suffix = ".o")
#> [1] "alpha"