nonlinear_engine = "gam" adds a penalised GAM
(mgcv) alternative to the CART conditional model, with
optional tensor interactions (nonlinear_gam_interactions)
and local residual resampling (nonlinear_local_k).nonlinear_engine = "auto" selects CART, additive GAM or
GAM with interactions by training-only K-fold cross-validation
(nonlinear_cv_folds, nonlinear_cv_rule).
Candidates are scored by the continuous ranked probability score of
their conditional predictive distribution
(nonlinear_cv_score = "crps", default), which accounts for
the conditional spread as well as the conditional mean;
nonlinear_cv_score = "mse" scores the conditional mean
only. Both scores are stored in
fit$pooled$nonlinear$selection.nonlinear_residual ("leaf",
"global", "none") and
nonlinear_calibration ("quantile",
"none") control the residual generator and the marginal
calibration of the response.fit_synthesizer() now warns and
lists such strata in fit$strata$conditional_fallback;
print() reports them.nominal_target generates one unordered factor,
character or logical response from a multinomial model
(nnet) of its conditional distribution given the
copula-generated predictors. Numeric predictors enter through natural
cubic spline bases (nominal_spline_df, default 3), so
category probabilities can depend nonlinearly on them; the linear model
is fitted too and the one with the smaller AIC is used (0
fits the linear model only). The response is omitted from the predictor
copula; its missingness indicator stays in the joint missingness model.
Mutually exclusive with nonlinear_target. Strata without a
converged model fall back to the copula, with the same warning and
record as above.nominal_decay and nominal_maxit (default
1000 iterations) are passed to nnet::multinom().splines (a base R package) is now imported.t_profile = TRUE records elapsed times of the fitting
stages under fit$pooled$copula$profile.haven_meta attribute added by forthcoming versions of
‘haven’ (contributed by Hadley Wickham).arrow::read_ipc_file() and
arrow::write_ipc_file(), which replace the deprecated
read_feather() and write_feather() (used only
with older ‘arrow’ versions). This removes two deprecation
warnings.nonlinear_engine = "auto" with
training-only K-fold selection among CART, additive GAM and GAM with
numeric tensor interactions. The default engine remains CART.nonlinear_cv_folds (default 5) and
nonlinear_cv_rule ("one_se" or
"min"). Model-selection diagnostics, including per-fold MSE
and candidate failures, are saved in the fitted nonlinear model.nonlinear_engine = "gam" for one
continuous hybrid target. Numeric predictors enter through penalised
cubic regression splines and categorical predictors through factor
effects. Tensor interactions may be enabled explicitly with
nonlinear_gam_interactions = TRUE.nonlinear_residual = "leaf"
option uses nearest fitted-mean residual neighbours rather than tree
leaves. CART, Gaussian copula, and Student-t copula defaults and results
are unaffected.mgcv. This is experimental and no superiority or privacy
guarantee is claimed.nonlinear_residual = "leaf", "global", or
"none", and nonlinear_calibration = "quantile"
or "none". Defaults reproduce the original hybrid method.
The six combinations allow controlled study of stochastic residual
resampling and response marginal calibration without changing the public
default.nonlinear_target argument to
fit_synthesizer() to generate one continuous response from
a CART conditional model while retaining copula-based generation of
predictors. Leaf-grouped residual resampling and empirical quantile
calibration are used to approximate the response conditional
distribution and preserve its marginal. The default copula path is
unchanged.rpart package.levels() order of
unordered factors during marginal fitting and latent-copula synthesis.
Permuting the levels without changing observed labels no longer changes
synthetic observations when the fitting and generation seeds are held
fixed.ordered factors is unchanged. Generated factors retain
their original class and level metadata, including unused levels.Added regression tests verifying that reordering the levels of unordered factors does not change synthetic observations when random seeds and all other settings are identical.
Added checks for preservation of factor classes, original level metadata, and ordered-factor behaviour.
Existing public function signatures remain unchanged.
Fitted models involving unordered factors may produce different results compared with earlier releases because of the corrected internal category ordering.
Users requiring the corrected behaviour should refit models created with previous versions.
write_data()/read_data() were shifted by the
time-zone offset (e.g. 2 or 3 hours for Europe/Bucharest). The instant
is now preserved with every ‘writexl’ version and with ‘openxlsx’.
?write_data documents how each format stores
date-times.inst/CITATION and CITATION.cff give the
Zenodo DOI.?fit_synthesizer.ts
selected rows with EuStockMarkets[1:500, ], which returns a
matrix rather than a ts, so the stationary bootstrap was
never used and a warning was printed; it now uses window().
Passing time-series arguments (method, order,
…) for a non-ts input now gives a warning that explains the
cause.synthesize(file = )
example now writes a file, shows its first lines and reads it back with
read_data().Statistical validation, and the bias it uncovered.
x and the missingness of y
was reproduced at about 60% of its strength (0.30 instead of 0.51). Such
pairs are now estimated by maximum likelihood under the Gaussian copula:
polyserial (discrete-continuous) and polychoric (discrete-discrete)
correlations, with thresholds from the marginal probabilities (Olsson
1979; Olsson, Drasgow and Dorans 1982) and the bivariate normal CDF of
Genz (2004). The estimates agree with polycor’s two-step
estimators to about 1e-5. The t copula uses them for these pairs.y depends on x, the rows where
both are observed have a restricted range; the new estimator reproduces
the relation among the observed rows (e.g. cor(x, y) among
observed rows 0.568 synthetic vs 0.572 real, where the old estimator,
combined with the first fix, would give 0.646). With complete data
nothing changes.iris now keeps the species separation much better (a
petal-based rule classifies 82% of synthetic records correctly vs 66%;
real data 96%).latent:
"normal_scores" or
"polyserial_polychoric").tests/testthat/test-validation.R: data simulated
from known distributions, dependence structures, missingness mechanisms
(MCAR, MAR), strata and VAR processes; synthetic samples are compared
with the truth or the training data within Monte Carlo tolerances
(marginals, latent correlations, rank association with discrete margins,
2 x 2 tables, MAR missingness probabilities, observed-row relations,
stratum means and correlations, t-copula df and tail dependence, VAR
auto- and cross-correlations). Run against 0.2.6, it fails 11
expectations in exactly the areas affected by the bias above. The large
designs are skipped on CRAN; one fast MAR check always runs.fit_synthesizer.ts() warns when an argument does not
apply to the chosen method (block_length outside the
bootstraps, order/burn_in outside
"copula_var", max_lag when order
is given).disclosure_risk() gains abs_tolerance for
numeric targets: a prediction is correct if within
max(tolerance * |y|, abs_tolerance), so targets with
meaningful zeros can be assessed.missing = "drop" leaves columns without
any observed value entirely missing; missingness that depends on the
unobserved value itself (MNAR) is not reproduced; the time-series
methods assume approximately stationary series
(?fit_synthesizer.ts, time-series vignette).airquality; its own output showed otherwise (also in
0.2.6). A single copula only represents monotone relations; the section
now shows a monotone MAR example and that by = "Month"
reproduces the non-monotone pattern.testthat::local_mocked_bindings()
(testthat >= 3.1.7) or data.frame() on an
integer64 column (needs bit64), which failed on R 4.2 with
an older testthat.Bugs fixed.
seed argument in fit_synthesizer()
(data frame, matrix, vector and ts methods). Fitting is
stochastic (random tie-breaking, the randomised distributional
transform, Kendall subsampling for the t copula), so the fitted model
can now be made reproducible when fitting and generation are separate
steps; the global RNG state is restored afterwards.k could be read as a
total.n_pair alone does not flag all of them
(?copula_correlation, ?fit_synthesizer).
summary() labels its count as referring to the pooled
model.Bugs fixed.
n_pair) and the
identification of pairwise correlations were computed on the random
2000-row subsample used for Kendall’s tau (bug introduced in 0.2.4).
With more than 2000 rows,
copula_correlation(fit, what = "n_pair") under-reported the
counts, and a pair with sparse overlap (e.g. 10 joint rows out of
10,000) could be declared unidentified, or estimated from only a few of
its rows, depending on the seed. Counts and identification now always
use the full data, and any pair that the subsample leaves with fewer
than min(n_pair, 500) joint rows is re-estimated from up to
2000 of its own joint rows.p(p - 1)/2 separate O(n^2) calls of
cor(method = "kendall"). Estimates for complete data are
unchanged.completion: "none", "maxdet" or
"zero_fallback") and the largest change made by the
subsequent positive-definiteness repair (pd_adjustment);
summary() reports both. The fit-time warning now states the
method actually used and that the repair may follow, instead of always
claiming maximum-determinant completion.Bugs fixed.
fit_synthesizer() warns when this happens, the fitted
copula stores the pairwise counts (n_pair) and the number
of unidentified pairs,
copula_correlation(fit, what = "n_pair") returns the
counts, and summary() reports them. Fully observed data are
unaffected.r_mvskewnorm() requires a positive-definite
omega, as in the standard definition of the skew-normal
density; singular sigma remains allowed for
r_mvnorm(), r_mvt() and
r_mvlnorm() (degenerate distributions), and this is now
documented..check_prob() (used for closeness,
close_quantile, …) rejects non-finite values explicitly,
consistent with the other validators.Bugs fixed.
r_mvnorm(),
r_mvt(), r_mvlnorm(),
r_mvskewnorm() and r_mvmixture() are validated
by one shared check: non-empty, square, numeric, finite, symmetric and
positive semi-definite, each with a clear error message (previously some
invalid inputs failed with internal errors such as “infinite or missing
values in ‘x’”).mean, meanlog,
xi, skew-normal alpha, mixture means) must be
finite; Inf/NA used to propagate silently into
the output.r_mvskewnorm() requires a strictly positive diagonal of
omega, as its parameterisation standardises by
sqrt(diag(omega)).r_dirichlet() now draws the gamma variates on the log
scale (with the Y * U^(1/a) boost for shapes below 1) and
normalises with log-sum-exp, so very small concentration parameters no
longer give NaN rows. Draws for alpha < 1
therefore differ from 0.2.2 for the same seed.make_corr(d = 1) returns the 1 x 1 identity without
consulting rho.df is not estimated but fixed at 10; this now triggers a
warning, is recorded as df_estimated = FALSE in the fitted
copula and is shown by summary().?fit_synthesizer gains an “Estimation details” section
documenting the pairwise-complete estimation of latent correlations, the
subsequent projection to a valid correlation matrix, and the
grid/profile-likelihood estimate of the t-copula df.Bugs fixed.
r_dirichlet() requires finite
positive alpha; make_corr() requires finite
numeric rho; disclosure_risk() requires a
finite non-negative tolerance; seed arguments
must be finite and in the integer range.r_mvmixture() validates every component (finite means,
finite and positive semi-definite covariance matrices) before sampling,
so an invalid component is reported even when its weight is zero.NaN like NA (both are
missing values).dcr(): exact_match_rate now counts only
records identical to a real record on all variables. New
zero_distance_rate (distance zero on the variables
compared) and compared_share (share of variables on which
the nearest real record was compared) make the meaning explicit under
na = "exclude". real_duplicate_rate is defined
in the same way.compare_synthetic() flags a discriminator AUC well
below 0.5, which indicates synthetic records that copy real ones. The
AUC is deliberately not converted to a symmetric measure such as
max(AUC, 1 - AUC), which would report copies of the real
data as easy to distinguish.Bugs fixed.
r_copula() and r_mvdist() now require
corr to be a correlation matrix (square, symmetric, unit
diagonal, entries in [-1, 1], positive semi-definite). Previously a
covariance matrix was accepted and produced margins that were not
uniform. The t copula and r_mvt() validate
df (positive; Inf gives the Gaussian copula);
Archimedean theta must be finite.disclosure_risk() and
exact-copy detection are now built from integer codes: a missing value
can no longer collide with the string "<NA>", and
values containing the old separator can no longer make different rows
look identical. Strata are listed in order of first appearance.dcr() gains na = c("category", "exclude").
The default treats missingness as a category (as before, now
documented); "exclude" gives the standard Gower
coefficient, ignoring variables missing in either record.discriminator_auc() learns all preprocessing
(imputation, scaling, retained categories) within the training folds, so
the cross-validated AUC has no leakage from the held-out fold. Its
documentation explains that values clearly below 0.5 indicate copies of
real records.r_mvmixture() rejects all-zero or non-finite weights
and components of inconsistent dimension with clear messages;
margin_categorical() rejects duplicated levels and
non-finite probabilities; the quantile functions it and
margin_empirical() return check that probabilities lie in
[0, 1].match_types() reads date-time strings without zone
information as local times in the template’s time zone (strings ending
in Z are UTC, explicit offsets such as +02:00
are honoured), and accepts mixed formats within a column.noise perturbation preserves mean and variance
approximately rather than exactly.Complete rewrite. the API changed and is not backward compatible with 0.1.0.
fit_synthesizer() / generate() /
synthesize() with Gaussian, Student-t and independence
copulas; categorical, ordinal and logical variables take part in the
dependence structure through the distributional transform.by) with
pooled fallback, identifier columns, compressed quantile functions
(knots), dependence strength control and
simulate() for multiple synthetic data sets.closeness argument of generate(),
synthesize() and augment_data();
perturbation() for record-level (noise, swap) and
distribution-level (shift, scale, temperature, dependence up to 2)
settings; and calibrate_closeness() to reach a target
discriminator AUC, DCR ratio, pMSE ratio or marginal distance.discriminator_auc() now uses squares and pairwise
interactions of the numeric variables by default
(model = "quadratic", ridge-penalised), so that differences
in spread and correlation are detected; pmse() gains the
same option.disclosure_risk() for statistical disclosure control:
replicated uniques and correct attribution probability (with no-data
baseline) for chosen key and sensitive variables. dcr()
gains close_share, the share of near-copies of real
records. calibrate_closeness() accepts any custom metric,
e.g. a risk threshold.census_sim and
vignette “Releasing confidential microdata as synthetic data”.augment_data() and ts_windows() for
enlarging small training sets of tabular data and time series for
machine learning.r_mvdist() for arbitrary margins.compare_synthetic() bundles them.read_data() / write_data() for CSV, TSV,
RDS, RData, Parquet, Feather, HDF5, JSON, NDJSON, Excel, SPSS, Stata,
SAS and fst; generate_to_file() for chunked generation of
very large data sets.closeness given to fit_synthesizer()) now
trigger a warning instead of being silently ignored.closeness, perturb and
dependence are not available for time-series synthesizers;
passing them now warns instead of being silently ignored.discrete_threshold values and these are whole
numbers or repeat.noise, shift,
scale) also apply to numeric variables modelled as
discrete, e.g. within strata; whole-number variables are rounded
stochastically, so shifts are not biased by rounding.difftime, keep
their class and units; 64-bit integers (integer64) are
rejected with a clear message.disclosure_risk() gains ignore to leave
out direct identifiers, which otherwise mask exact copies.augment_data() accepts plain vectors.pd_method = "none".rankcor < 1 changed the marginal distributions.make_synthesizer.POSIXct() returned a character
vector.seed = left a global .Random.seed
behind.synthesize() failed for series longer than 1000
observations.