Lorentz Hyperbolic Weighted Learning on the Lorentz hyperboloid.
Details
The LHWR module provides unpenalized Gaussian and binomial local regression. The LHWPR module provides Gaussian and binomial ridge, lasso, elastic net, MCP, and SCAD paths with joint bandwidth-regularization tuning. Geometry, bandwidth selection, local diagnostics, and plotting utilities support both modules where statistically appropriate.
LHWR is the original unpenalized regression foundation. LHWPR-Lasso is a separate penalized extension, not a replacement name for LHWR. Software support for other penalties or response families does not imply that the original Gaussian LHWR theory supplies guarantees for those extensions.
Use make_lhwr_Z_features() or make_lhwr_Z_network() to construct a
representation, bw.lhwr() and lhwr() for unpenalized regression, and
tune.lhwpr() for joint penalized tuning. For a complete introduction use
vignette("lorentz-hgwr-introduction", package = "LHWL"); for Lasso use
vignette("lhwpr-lasso", package = "LHWL").
The supplied representation defines locality; it is not automatically a predictor or a validated latent geometry. Validation of a learned representation requires rebuilding it within the relevant training splits. Local Gaussian diagnostics condition on the fitted representation and smoothing specification. Penalized activation is not a significance test.