ggRandomForests 3.5.5

CRAN release, accepted 2026-10-25. Tagged on ba5c923e, the tree submitted to CRAN (recorded in CRAN-SUBMISSION).

• gg_ivarpro() no longer fails with "replacement has 1 row, data has 0"
  when every local-importance value for a class is nonfinite. The empty
  per-class frame now takes its class column without error: a fit where
  every class is empty returns an empty gg_ivarpro, and one where only some
  are drops those classes. Reported by Udaya Kogalur.
• gg_ivarpro() on a binary classification forest labelled its profile
  with the wrong class. varPro::ivarpro() explains one class there, the
  first factor level (its target attribute says so), and gg_ivarpro()
  filed that profile under the last level, the default. Every binary
  gg_ivarpro() plot therefore showed the first class's slopes, sign
  included, under the second class's name. Because the two class
  probabilities sum to one, the second class's slopes are the first's
  negated, so gg_ivarpro() now derives it exactly and both levels are
  available through which_class. The default stays the last level, now
  with the right values. Under use.abs = TRUE the two profiles are equal.
• gg_ivarpro() mirrors a binary profile into the second class only when it
  explains the forest's own probability of the first level. A y.external
  target named after a class is any numeric response, not a class
  probability, and is kept as a single profile, as before.
• varPro (>= 3.3.0) is now required, up from 3.1.0. ivarpro() records
  the class it explains (its target attribute) only from 3.2.1, and the
  corrected binary gg_ivarpro() depends on it; 3.3.0 is the version tested
  and the one on CRAN.
• gg_partial_varpro() labelled categorical partials in the wrong order.
  Each column of varPro::partialpro()'s output belongs to one value of
  $xvirtual (sorted values, or factor levels), but the labels were taken
  from the order values first appear in the sampled cases, so a 0/1 variable
  could have "0" and "1" swapped and a multi-level variable had its levels
  shifted, whenever the two orders differed. On the PBC data this made stage
  4 and severe edema look protective. Categorical partials are now labelled
  by $xvirtual, in the factor's own level order (and ordered status) when
  part_dta carries factor-valued xorg.
  The precomputed varpro vignette objects are rebuilt, so the Boston rad
  and PBC edema panels now carry the right labels.
• gg_varpro(conditional = TRUE) on a binary classification forest now
  labels its classes with the outcome's own level names. varPro recodes a
  binary outcome to 0/1 internally and names its conditional columns after
  those codes, so the facets read 0 and 1 instead of, say, N and R.
  gg_ivarpro() already mapped them back; gg_varpro() now does the same,
  and $conditional$class is a factor in the outcome's level order.
  Multiclass outcomes, which varPro does not recode, are unchanged, including
  ones whose own labels are "0", "1", ...
• plot.gg_varpro() draws no bar for an undefined class-conditional z
  instead of a zero bar that reads as "tested and unimportant". The z is
  0/0 when too few trees tested a variable for that class, which happens
  with default settings because gg_varpro() calls varPro::importance()
  with local.std = FALSE. The variable keeps its row, including one varPro
  leaves out of every class, every class keeps its facet, and the caption
  says the z is undefined (#344).
• gg_partial_varpro(scale = "surv") no longer returns survival above 1.
  partialpro() smooths each case's S(tau) with an unbounded polynomial, so
  where survival is near 1 (early horizons, before most events) the averaged
  curve ran up to about a percentage point past it. The curve is now clamped
  to [0, 1] after averaging, and the categorical frame per value. Only the
  impossible values change; everything in range is exactly as before.
  Smoothing on the logit scale was tried and rejected: partialpro() replaces
  each case's intercept with the cohort mean, so on that scale the curve's
  level depends on how S = 0 and S = 1 are clamped. A precomputed part_dta
  labelled "surv" passes through unchanged.
  The varpro vignette's precomputed PBC partial is rebuilt: its edema
  panel had 60 values below 0, now 0.
• ?gg_partial_varpro no longer says that scale = "surv" is computed
  through gg_partial_rfsrc(). Only scale = "chf" takes that route;
  "surv" uses the partialpro() S(tau) learner, and its time is the
  horizon tau, defaulting to the median follow-up time.
• Two tests no longer depend on the randomForestSRC random-number stream.
  The live gg_sdependent() test computes its beta fit once and gives it to
  both sides, and the plot.gg_varpro() test for a class with no defined z
  sets that z by hand. Both failed against the randomForestSRC development
  release, whose new seed generator grows a different forest from the same
  seed.