Skip to contents

Changelog

Source: NEWS.md

LHWL 0.3.1

  • Fixed Moran and LISA diagnostics with Gaussian or exponential kernels: adaptive kNN radii still set bandwidths, but stored neighbour lists no longer truncate noncompact kernel tails. These diagnostics now use full distance matrices. Added cached/full/explicit-weight parity tests, including permutation inference. Bisquare diagnostics and model estimation are unchanged.

  • Added executable LHWR companions for kernels, bandwidth selection, F1–F3 and Moran/LISA diagnostics, and manuscript-style graphics/export. Documented supported engine combinations, raw versus adjusted classes, plotting transformations, and diagnostic-weight cache behavior. This documentation revision does not change an estimator or historical run.

  • Expanded the introduction and Lasso vignettes into linked, executable tutorials with quick-start examples, interpreted graphics, numerical safeguards and explicit scope limits.

  • Added local pkgdown navigation and a shared citation-maintenance workflow. Public manuscript identifiers and hosting remain pending; no publication or deployment is performed by the documentation tools.

  • Synchronized the release candidate with the corrected Tangent-E prediction distances. Direct, scaled coordinate differences preserve exact zero distances for identical locations without rounding nearby distinct points to zero. Added calibration, duplicate, reordered-row, and batch tests.

  • Made the near-coincident-point test’s absolute error bound explicit; this changes the test assertion only, not a distance or estimation tolerance.

  • Expanded the geometry, reference-center, prediction, conditional-validation, tuning, and inference documentation without changing estimation settings.

  • Matched the C++17 requirement in DESCRIPTION with explicit Unix and Windows Makevars declarations; retained portable R-provided native-library flags.

  • Prepared the public source metadata, package citation, data attribution, and self-contained WDI quick start for GitHub and CRAN review.

  • Recorded the frozen location and scale constants used to construct the packaged WDI Lorentz coordinates and clarified that the representation variables define locality rather than enter the regression formula.

  • Hardened joint tuning, exact selected-model refits, and LHWPR graphical diagnostics.

  • Required every selected tuning row to pass full-fit convergence and KKT gates; cross-validation candidates must also pass the same validation-fit gates, and the minimum, one-standard-error, and sparse-one-standard-error rules now share exactly the same eligibility set.

  • Changed the final selected-model refit to reuse the exact decreasing lambda prefix evaluated for its tuning configuration. The fitted tuning object now records this prefix and a convergence/KKT refit certificate.

  • Added the non-interpolated plot.tune.lhwpr(type = "landscape") display, which distinguishes coarse, refined, ineligible, and selected candidates.

  • Added plot_lhwpr_surface() for coefficient and focal-selection displays. Numerical zeros are hollow and the documentation explicitly distinguishes penalized activation from statistical significance.

LHWL 0.3.0

  • Declared the R-provided LAPACK, BLAS, and Fortran runtime libraries in the Unix and Windows native-link recipes, preventing unresolved Armadillo linear-algebra symbols on clean installations.
  • Restored the local R-reference regression tests for the C++ bandwidth, fitting, and Lorentz-geometry implementations, and added a deterministic Gaussian LHWPR-Lasso test for the p > n setting.
  • Added a non-overwriting, versioned review audit that records the R runtime, compiler/link configuration, user Makevars provenance, package versions, test/check logs, source hashes, and SHA-256 artifact inventory. The audit now stops before compilation when a required compiler or link-library path is unavailable.
  • Redesigned tune.lhwpr() around a cached coarse-to-refined joint search, optional alpha-midpoint refinement, automatic adaptive-kNN limits through n - 1, and explicit boundary diagnostics.
  • Added Gaussian pathwise AIC, AICc, BIC, GCV, and EBIC. Effective degrees of freedom are computed in C++ from the conditional local smoother. Primary EBIC adds the exact coefficient-surface model-space penalty to pathwise BIC; the separately named ebic_refit sensitivity score retains post-selection local refits and excludes saturated refits with at most one residual degree of freedom.
  • Excluded invalid effective-degree-of-freedom rows from Gaussian information-criterion selection and exposed zero EBIC model-space terms for nonempty full-surface models.
  • Added rule = "sparse_1se", richer convergence/active-support selection tables, and a same-fold unpenalized LHWR comparison.
  • Added plot.tune.lhwpr() bandwidth, lambda, joint-surface, and alpha views.
  • Added unpenalized binomial LHWR through lhwr(..., family = "binomial").
  • Added local logistic coefficient, probability, link, convergence, effective-neighborhood, class-variation, Fisher-rank, standard-error, and z-value outputs.
  • Added binomial prediction at calibration and new Lorentz locations.
  • Added binomial bandwidth selection by leave-one-out Bernoulli log loss.
  • Made summaries, reports, diagnostics, and Gaussian-only inference checks response-family aware.
  • Documented LHWL as one package containing the LHWR and LHWPR modules, with later learning modules governed by an explicit staged roadmap.

LHWL 0.2.0

  • Added LHWPR Gaussian and binomial ridge, lasso, elastic-net, MCP, and SCAD paths and joint bandwidth-regularization tuning.