MRF unsupervised clustering – Proximity method

Usage

mrf3_cl_prox(
  rfit,
  k = NULL,
  enhanced = TRUE,
  size_min = 5,
  use = "X",
  symm = TRUE,
  leaf_embed_dim = 10,
  merge_quantile = 0.9,
  merge_mode = c("soft", "hard"),
  sibling_gamma = 0.5,
  sibling_fun = c("constant", "positive", "shift01", "abs", "identity"),
  sibling_cap = TRUE,
  hard_prox_mode = c("soft_enhanced", "binary"),
  parallel = TRUE,
  sparse = FALSE,
  method_cl = "PAM",
  tune_method = "silhouette",
  gap_w = "uniform",
  cores = NULL,
  ...
)

Arguments

rfit: A model list of random forest models.
k: Pre-defined number of clusters. By default, the optimal value is selected by the requested tuning method.
enhanced: Logical; whether to calculate enhanced proximity.
size_min: Minimum terminal-node size when computing enhanced proximity.
use: Which importance side to use in tree traversal, "X" or "Y".
symm: Logical; whether to symmetrize directional proximities.
leaf_embed_dim: Dimension used for low-dimensional leaf embedding.
merge_quantile: Quantile threshold for sibling-pair merges in enhanced mode. 0.9 keeps the top 10\ iteration. Used only when merge_mode = "hard".
merge_mode: Enhanced-proximity merge mode: "soft" (default) or "hard". "hard" performs iterative sibling merges; "soft" directly builds per-tree similarity with same-leaf = 1 and sibling-leaf = sibling_gamma * f(corr). "soft" is generally more stable.
sibling_gamma: Multiplicative weight gamma used in soft merge mode.
sibling_fun: Correlation transform f(corr) used in soft merge mode. One of "constant", "positive", "shift01", "abs", or "identity".
sibling_cap: Logical; whether to cap soft sibling similarities at 1.
hard_prox_mode: Proximity construction used after hard merge: "soft_enhanced" uses same-enhanced-leaf = 1 and sibling-enhanced-leaf = sibling_gamma * f(corr), while "binary" uses same-enhanced-leaf only.
parallel: Logical; whether to allow forest-level computation in fresh PSOCK worker processes. Parallel execution also requires an explicit cores value.
sparse: Logical; whether to sparsify the enhanced proximity matrix.
method_cl: Clustering backend ("PAM" or "Spectral").
tune_method: Tuning criterion for PAM when k is NULL ("silhouette" or "ratio").
gap_w: Weighting scheme for spectral eigengap when k is NULL ("uniform" or "log").
cores: Number of CPU cores used by parallel steps. Default NULL keeps matrix-heavy forest reductions serial to limit peak memory.
...: Additional arguments passed to downstream clustering helpers.

Value

An mrf3 clustering object.