LCDA-GR: Reactive variant with self-tuning of (alpha_c, alpha_s). Implements Algorithm 4 of the paper.

LCDA-GR: Reactive variant with self-tuning of (alpha_c, alpha_s). Implements Algorithm 4 of the paper.

Usage

lcda_gr(
  g,
  variant = 1,
  B = 150,
  centrality = "eigen",
  similarity = "hpi",
  m = 20,
  y = NULL,
  alpha_c_range = c(0.1, 0.9),
  alpha_s_range = c(0.1, 0.5),
  p_floor = 0.05,
  verbose = FALSE,
  seed = NA_integer_
)

Arguments

g

an igraph object (undirected, simple).

variant

construction variant, 1 or 2.

B

number of GRASP iterations.

centrality, similarity

metric names.

m

pool size (default 20, per paper).

y

refresh period (default 3m).

alpha_c_range, alpha_s_range

bounds for the uniform initial pool.

p_floor

minimum probability mass reserved across the pool at each refresh, spread uniformly so every pair keeps `p_k >= p_floor/m > 0`. This prevents a pair that happened not to be sampled in a block from being permanently excluded (and matches the `p_k >= delta > 0` premise of Proposition 6). Set to 0 to recover the raw proportional rule.

verbose

logical; show a cli progress bar and a final summary.

seed

integer RNG seed, or `NA` to leave the RNG untouched.

Value

an object of class `lcda_gr_result`: best partition, traces, the reactive pool state, the H-decisive iterations, the wall-clock `elapsed` time in seconds, and the (simplified) input `graph`, so that [lcda_metrics()] and [plot()] can be called on the result alone.

Details

Weighted graphs: a numeric `weight` edge attribute is honoured by the modularity objective and the local search, but the construction (similarity and centrality) and the NCE leader score remain *structural* (unweighted).

See also

[lcda_metrics()], [plot.lcda_gr_result()].

Examples

g <- igraph::make_graph("Zachary")
res <- lcda_gr(g, B = 30, seed = 1)
res$best$Q
#> [1] 0.4197896
length(res$lex_decisive_iters)  # how often H broke a Q-tie
#> [1] 0