Prepare reusable training-fold inputs for SCAD/MCP cross-validation

Prepares full-data and per-training-fold scaling and neighbor structures for reuse by SCAD and MCP. Pass it as cv_cache to cv_gwr_nonconvex, cv_gwr_scad or cv_gwr_mcp.

Usage

gwr_nonconvex_cv_cache(x, coords, k, fold, standardize = TRUE)

Arguments

x

Numeric predictor matrix with observations in rows.

coords

Numeric coordinate matrix or an sf point object.

k

Retained neighbor count, no larger than any training-fold sample.

fold

One non-missing spatial-fold label per observation.

standardize

Whether to center and scale predictors within each training fold.

Value

A read-only gwrs_nonconvex_cv_cache environment with locked bindings.

Details

Predictors, coordinates, neighbor count, fold labels and standardization must match exactly, including row order. Each fold uses its own training-only transformation. Responses, penalties, coefficients, warm-start state and lambda paths are never cached. Automatic lambda anchors retain their existing method-specific full-data definition. Kernel weights are not cached, allowing preparation to be reused across kernels and bandwidths. Holding all prepared folds costs additional memory; cv_cache = NULL retains the one-fold-at-a-time route. No global cache is created.