Profiles every candidate on the training rows only, decides its fate
and materializes the clean data for train and hold-out with the same
values (training median, "MISSING" level). A sentinel or missing mass
with weight (special_min_share) and signal (special_min_woe) becomes a
categorical flag <column><flag_suffix>, which the engine bins and emits
in SQL natively.
Usage
scr_triage(split, config = scr_config())Arguments
- split
An object from
scr_split().- config
An object from
scr_config().
Value
An scr_triage object with profile (one row per candidate and
derived flag), ledger (the source of truth of the pre-processing the
SQL reproduces), keep, derived, clean (target + survivors + flags,
with no NA) and the originating split.
Details
Failures at this stage: CONSTANT, NEAR_CONSTANT, TOO_MANY_MISSING,
HIGH_CARDINALITY, NO_SIGNAL (coarse IV below min_iv_quick) and
DUPLICATE_OF:<column>. A failed numeric whose special-population flag
survives carries the suffix ;RESCUED_AS_FLAG.
See also
Other stages:
scr_align(),
scr_bin(),
scr_cutoff(),
scr_model(),
scr_reject(),
scr_scorecard(),
scr_select(),
scr_split(),
scr_strategy()
Examples
sp <- scr_split(scr_demo, "default", date_col = "ref_date", drop = "id")
#> OOT: 4 period(s) in train, 2 in hold-out (hold-out starts at 2026-05-01, 33.3% of rows)
tr <- scr_triage(sp, scr_config(verbose = FALSE))
tr
#> <scr_triage> target "default" | 38 candidates -> 37 survivors (7 derived)
#> CONSTANT 2
#> TOO_MANY_MISSING;RESCUED_AS_FLAG 1
#> TOO_MANY_MISSING 1
#> NEAR_CONSTANT 1
#> DUPLICATE_OF:vl_score_01 1
#> HIGH_CARDINALITY 1
#> NO_SIGNAL 1
table(tr$profile$triage_reason)
#>
#> CONSTANT DUPLICATE_OF:vl_score_01
#> 2 1
#> HIGH_CARDINALITY NEAR_CONSTANT
#> 1 1
#> NO_SIGNAL OK
#> 1 37
#> TOO_MANY_MISSING TOO_MANY_MISSING;RESCUED_AS_FLAG
#> 1 1