figs(), fit_figs() and bagging_figs() gained na_method. The default
"omit" is the previous behavior. "mia" (missingness incorporated in
attributes, Twala et al. 2008) keeps every row: at each split the search
also tries sending missing values left or right of the cutpoint, and
splitting on missingness itself, and uses the learned direction at
prediction time. summary() prints these as IS MISSING clauses, plot()
labels the edge missing values take, and figsr_importance() reports a
split on missingness on its own missing(<predictor>) row.
ggfigs() draws a fitted model with 'ggplot2', in five views:
"tree" (node-link diagram of every tree), "rules" (every tree as a
horizontal rule list, with + between trees), "waterfall" (where the
prediction for one observation comes from), "cuts" (the numeric
cutpoints drawn over the data) and "profile" (exact partial dependence).
It returns a ggplot; with interactive = TRUE it returns a 'ggiraph'
widget whose tooltips show the rule behind each node, leaf or cut.
'ggiraph' is only suggested. autoplot() gives the same ggplots.
plot() gained main, an overall title drawn once above every tree
panel, and tree_names, overriding the automatic per-tree panel titles
(a single string applies to every tree; a vector is indexed by each
tree's position in the fit, so it stays correct when tree_idx only
plots some of them; NA in a position keeps that tree's default title).
In the "scientific" and "modern" styles, a leaf's background, border
and text now depend on the sign of its contribution -- green for
positive, red for negative; "classic" stays monochromatic by design.
max_splits = 10), allocating about 30 times less memory. Everything that depends on
the predictors alone (row order, cutpoints, level groups) is computed once
per node instead of once per iteration, every cutpoint of a predictor is
scored from a single cumulative sum of the residuals instead of one pass per
cutpoint, and a leaf whose residuals did not change since the last iteration
keeps its previous best split. The fitted model is unchanged: the same
splits, cutpoints, missing-value directions and sample indices as the
previous search, which a new regression test checks on fourteen data sets.The parsnip engine now fits through the formula interface. The previous
data.frame interface ran model.frame() with na.omit before the engine,
so fit() silently dropped rows with a missing predictor.
plot() no longer draws boxes on top of each other. Nodes used to sit on a
fixed grid, so in deeper trees neighboring leaves and branch conditions
overlapped and their labels were cut off. Each split now places its
subtrees from the width of their labels, every tree gets a panel width
proportional to the room it needs, and all trees share one vertical scale,
so their roots line up. A drawing too large for its panel is scaled down,
text included, instead of clipped.
pima_diabetes: the Pima Indians Diabetes Database (Smith et al. 1988),
768 patients, ships lazily loaded like iris. Its five physiologically
impossible zeros (Glucose, BloodPressure, SkinThickness, Insulin,
BMI) are already recoded to NA, so it is ready to demonstrate
na_method = "mia" without an internet connection or preprocessing.vignette("missing-values"), works through na_method = "mia" on the real Pima Indians Diabetes Database, where missingness is
informative but modest, and shows plot()'s new main, tree_names and
sign-colored leaves on it.figs() now carries the parent leaf's value into the two leaves created by
splitting it. Predictions from any tree deeper than one split were wrong
before this fix.
Classification probabilities are no longer passed through a logistic
function. The engine fits squared error on the 0/1 encoding of the outcome,
so the sum of the leaf values already estimates the probability of the second
level; it is now clamped to the unit interval instead. fitted_values is
clamped the same way, so it agrees with predict().
predict() now rebuilds the model frame from the fitted terms, so a formula
with a transformed term such as y ~ log(x) can be predicted from. It
previously failed with a missing-predictor error.
A matrix-valued term such as poly(x, 2) raised a cryptic
"missing value where TRUE/FALSE needed" during fitting; it is now rejected
with a clear message.
figs() gained subset and na.action arguments, which were previously
swallowed by ... and silently ignored. Passing weights now raises an
error rather than fitting an unweighted model in silence.
A model that accepts no split at all predicts the outcome mean instead of zero. In classification this was a probability of zero for every observation.
Integer predictors with large values no longer overflow when the candidate cutpoints are computed.
Numeric cutpoints above the 30-candidate cap are now quantiles of the sample rather than of its distinct values, so they follow the density of the data.
Factor predictors are judged by the levels present in a node rather than by the declared level set, so a factor carrying many unused levels is no longer skipped as if it had too many levels.
A missing value in a predictor at prediction time raises a clear error, and a
factor level unseen in training is rejected the way other stats model
functions reject it, instead of failing deep inside the tree traversal.
predict() rejects an unsupported type instead of quietly returning
numeric predictions.
bagging_figs() records the mode that was actually fitted. A factor outcome
with the default mode = "regression" produced an ensemble that returned
numeric scores instead of class predictions. Bagged classification also
averages member probabilities rather than hard labels, and supports
type = "prob". A character outcome is converted to a factor once, so a
bootstrap resample that misses the minority class no longer aborts the fit.
fit_figs() no longer drops a predictor genuinely named .outcome.
plot() validates style and tree_idx instead of rendering the wrong
style silently or failing with an internal error on NA. The "classic"
style labels leaves dy =, like the other styles, because a leaf holds one
tree's contribution rather than the prediction.
Registering the parsnip model is now idempotent, so reloading the package in
a live session no longer errors.
figs_tree() gained an engine argument defaulting to "figsr", and an
update() method, so specifications behave like the rest of parsnip.
print() reports which predictors actually carry a split, and summary()
states that classification leaf values are contributions to the probability
of the second class.
Description field no longer quotes acronyms, following the request from CRAN.parsnip integration and bagging, and its example data no
longer draws the outcome independently of the predictors.First release.
figs() fits Fast Interpretable Greedy-Tree Sums for regression and two-class
classification, growing a sum of shallow trees by greedily taking the split
that most reduces residual impurity, whether that split opens a new tree or
deepens an existing one.
figs_tree() registers the model with parsnip, so FIGS can be used inside
tidymodels workflows. max_splits() and max_trees() provide dials
parameter objects for tuning; min_n reuses dials::min_n().
summary() prints the tree sum as IF-THEN decision rules and plot() draws
the trees, in a "scientific", "modern" or "classic" style.
figsr_importance() ranks predictors by the total residual sum-of-squares
reduction attributable to each of them.
bagging_figs() fits a bootstrap ensemble of FIGS models.