randomForestSGT 2.0.0
New features
- Added node centering (
node.center = TRUE) before
forming polynomial and interaction terms, with support in fitting,
prediction, filtering, and tuning.
- Added restricted lasso predictions
(
node.restrict = TRUE), which clamp predictors and scores
to each node’s inbag predictor and response ranges. Applies to OOB and
new-data predictions; predict() can override the saved
setting. Older forests without this setting remain unrestricted by
default.
- Added
cv.rfsgt() and make.folds() for
cross-validation with reusable folds and per-fold seeds. Filtering and
optional hcut tuning are repeated within each training
fold, with held-out predictions and error summaries returned.
- Added
plot.rfsgt() for per-tree OOB risk curves and
comparisons of forest OOB error across tree sizes.
- Extended
get.beta() to report centered-model
coefficients in the original coordinates, including induced lower-order
terms. Restricted predictions are reconstructed from term contributions
plus response.adjustment.
- Added relaxed coefficient estimation with
get.beta(..., relaxed = TRUE) using OOB least squares,
global or local active terms, standard errors, and leave-one-out
predictions for training observations.
- Added
partial.beta() for variable-effect plots and
importance.beta() for marginal and conditional importance
summaries based on those effects.
Changes to defaults and
behavior
- Changed the default
hcut from 1 to 2, enabling
quadratic terms and variable filtering by default.
- Centered fits now default to
bsf = "inbag"; uncentered
fits with hcut > 0 retain bsf = "oob".
tune.hcut() now defaults to bsf = "inbag".
Fitting with a tuning result inherits its centering and BSF settings
unless explicitly overridden.
- Reduced the default tree-size limit for
hcut >= 2
from min(100, n/5) to min(25, n/5), and
changed the grow block.size default from 10 to 1.
- Changed the
get.beta() default from
bag = "inbag" to bag = "oob".
- Revised random-seed initialization; results for a fixed seed can
differ from version 1.0.0.
- Printed fits now report node centering and restriction settings and
flag potentially optimistic OOB error when variable filtering was
used.
Fixes and robustness
- Preserved source predictors and term maps needed to reconstruct
retained polynomial terms after filtering, including during tuning and
prediction.
- Corrected saved-forest restoration of terminal membership and inbag
counts, including repeated bootstrap draws, and preserved CART fallback
and split types when saving trees.
- Corrected forest offsets for predictions using selected trees
(
get.tree).
- Improved
get.beta() handling of restored objects and
observations without contributing trees, preserving missing predictions
instead of returning zero.
- Strengthened validation of saved forest dimensions and membership
output, widened native array-length and offset calculations, and
released native allocations used by custom bootstrap samples.
- Improved the SGT-importance fallback when varPro screening fails or
returns no result.
randomForestSGT 1.0.0