MRF unsupervised clustering – Reconstruction method

Usage

mrf3_reconstr(
  recon = NULL,
  rfit = NULL,
  k = NULL,
  weights_cluster = TRUE,
  dat_use_to_cluster = "ALL",
  model_top_v = 10,
  recon_fusion = c("weighted", "uniform"),
  global_fusion = c("average", "pmin"),
  connection_score = NULL,
  response_blocks = NULL,
  score_power = 1,
  score_floor = 0,
  fallback_uniform = TRUE,
  fused_top_v = NULL,
  fused_row_normalize = TRUE,
  fused_keep_ties = TRUE,
  fusion_mode = "average",
  gamma = 0.1,
  alpha_init = NULL,
  ao_max_iter = 20,
  ao_tol = 1e-04,
  knn_q = NULL,
  hollow = TRUE,
  ao_symm = TRUE,
  ao_verbose = FALSE,
  ...
)

get_reconstr_matrix(
  rfit,
  model_top_v = 10,
  recon_fusion = c("weighted", "uniform"),
  global_fusion = c("average", "pmin"),
  connection_score = NULL,
  response_blocks = NULL,
  score_power = 1,
  score_floor = 0,
  fallback_uniform = TRUE,
  fused_top_v = NULL,
  fused_row_normalize = TRUE,
  fused_keep_ties = TRUE
)

Arguments

recon: Optional reconstruction object returned by get_reconstr_matrix(). If NULL, reconstruction is built from rfit.
rfit: A model list of random forest models.
k: Pre-defined number of clusters. The default is selecting the optimal k by tuning method.
weights_cluster: Logical; whether to cluster based on reconstructed weight affinity.
dat_use_to_cluster: If clustering method selected as Reconstr, the user can choose the data to conduct clustering. In weight-clustering mode, shared clustering uses W_all; this argument only applies when weights_cluster = FALSE.
model_top_v: Model-level top-v cutoff applied to each single-model forest weight matrix before fusion.
recon_fusion: Reconstruction fusion mode. "weighted" (default) uses connection_score; "uniform" uses equal weights.
global_fusion: Fusion across per-response matrices. "average" implements the uniform Eq. 8 average; "pmin" is an optional element-wise intersection extension.
connection_score: Optional directional score matrix from find_connection(return_score = TRUE).
response_blocks: Optional character vector containing every data block that must have a response-side forest. The workflow supplies all input block names so an entirely omitted block cannot be silently ignored.
score_power: Exponent applied to raw connection scores before normalization.
score_floor: Non-negative floor applied to raw connection scores before normalization.
fallback_uniform: Logical; whether to fallback to uniform averaging when weighted scores are unavailable.
fused_top_v: Optional integer. If set, apply row-wise top-v truncation to fused weights (W_all) after fusion. If FALSE, no fused top-v truncation is applied.
fused_row_normalize: Logical; whether to row-normalize fused weights after optional truncation.
fused_keep_ties: Logical; whether fused top-v truncation keeps ties at cutoff.
fusion_mode: Clustering-affinity fusion mode. "average" keeps current behavior; "ao_reg" uses alternating optimization with simplex-regularized weights.
gamma: L2 regularization strength for the simplex weights in "ao_reg" mode.
alpha_init: Optional initial fusion weights for "ao_reg" mode.
ao_max_iter: Maximum alternating-optimization iterations.
ao_tol: Convergence tolerance for ||alpha_t - alpha_{t-1}||_1.
knn_q: Optional kNN sparsification per base similarity (rows keep top-q).
hollow: Logical; whether to zero out diagonal in base/fused similarities.
ao_symm: Logical; whether to symmetrize kNN-sparsified similarities by pmax(S, t(S)).
ao_verbose: Logical; whether to print AO iteration diagnostics.
...: Additional arguments passed to clustering backends.

Value

mrf3 clustering object