SelectBoost.gamlss 0.2.2
- Code and description fixes requested by CRAN.
SelectBoost.gamlss 0.2.1
- Additional fixes to code and documentation to silence CRAN check
notes for R devel.
- Added the boys7482 dataset.
- Fix knockoff filters to coerce grouped design matrices to numeric
and supply the response to
knockoff::create.fixed(),
preventing failures when smooth proxies carried non-numeric classes or
the design needed augmentation, and reuse any augmented design/response
returned from create.fixed() to avoid downstream dimension
mismatches.
SelectBoost.gamlss 0.2.0
Highlights
- Grouped selection for all parameters (μ, σ, ν, τ)
with
engine = "grpreg" (group lasso/MCP/SCAD) and
engine = "sgl" (sparse group lasso), including
factors, splines (pb()/cs()), and
interactions treated as single groups.
- Per-parameter engines:
engine,
engine_sigma, engine_nu,
engine_tau can be mixed (stepwise / glmnet / grpreg /
sgl).
- Glmnet support (lasso/ridge/elastic-net) extended
beyond μ via working-response proxies for σ/ν/τ.
- Glmnet selectors now accept
glmnet_family (gaussian/binomial/poisson) and handle factor
predictors via model-matrix expansion.
- Tuning framework:
tune_sb_gamlss()
with stability or deviance metrics
(K-fold), progress bars, and a complexity
penalty.
- Fast deviance paths for common families (auto-used
in deviance CV):
NO, PO, LOGNO,
GA, IG, LO, LOGITNO,
GEOM, BE, NBI, NBII,
BI, and native shortcuts via gamlss.dist for
many others (e.g., LOGLOG, DEL,
ZAGA, ZIP/ZIP2, ZAIG,
ZALG, ZIBI/ZIBB, PARETO,
SEP1/SEP2, ZIPF/ZIPFmu, BCT,
BCPE, SICHEL, GLG,
BETA4, RS, WEI,
GIG), with graceful fallbacks.
- Fast deviance dynamically calls
gamlss.dist::d<family>() when available, broadening
zero-inflated/hurdle coverage without manual whitelists.
- Group knockoffs (approximate) for FDR-style
control:
knockoff_filter_mu(),
knockoff_filter_param().
- Robust to rows dropped by
model.matrix() (e.g., missing
predictors) by aligning the response / working response before building
knockoffs.
- Compiled speedups (Rcpp/RcppArmadillo) for
scaling/cor; parallel bootstraps via
future.apply.
New user-facing functions
tune_sb_gamlss(), knockoff_filter_mu(),
knockoff_filter_param()
fast_vs_generic_ll(),
check_fast_vs_generic()
effect_plot() (quick partial effect visualizer for the
final selected model)
New arguments in
sb_gamlss()
engine_sigma, engine_nu,
engine_tau — choose engines per-parameter
grpreg_penalty
(grLasso/grMCP/grSCAD),
sgl_alpha
df_smooth — basis size for grouped-smoother
proxies
progress — progress bar for sequential bootstraps
- (still)
glmnet_alpha — 0=ridge, 1=lasso, (0,1)=EN
glmnet_family — choose gaussian/binomial/poisson for
glmnet selectors
Documentation & vignettes
- Real Data Examples (including
growth/BCT on
gamlss.data::boys)
- Advanced Real Data Examples (ZIP/ZINB on
bioChemists, ZAGA on airquality::Ozone,
longitudinal growth on nlme::Orthodont with random
intercepts)
- Benchmarks (engine timings) and Fast
deviance microbenchmarks
- Fast deviance equality (accuracy checks) +
wide-family sweep with per-family tolerances & skip
reasons
- pkgdown site scaffolding + GitHub Actions
workflow
- README quick start now covers factor effects, BCT four-parameter
example, and deviance-based tuning metrics.
Testing & quality
- Unit tests for fast vs generic deviance (accuracy and presence of
native densities)
- Opt-in long tests via
options(SelectBoost.gamlss.run_long_tests=TRUE) or
RUN_LONG_TESTS=true
Notes
- Some grouped/knockoff features are optional and require packages in
Suggests (
grpreg, SGL,
knockoff, glmnet, etc.).
- Smooths are proxied with
splines::bs(df = df_smooth)
for selection only; the final gamlss fit remains
as specified.
SelectBoost.gamlss 0.1.0
- First draft: bootstrap stability-selection over GAMLSS parameters
(mu/sigma/nu/tau).
- Optional pre-standardization of numeric predictors (stored for
prediction).
- AICc helper.
- Plotting + prediction helpers.