Skip to contents

Hybrid DSBM Estimation with Perturbed SVD Multi-Start

hdsbm_fit.Rd

Estimates a Dynamic Stochastic Block Model (DSBM) by combining the Hybrid CEM algorithm with a multi-start strategy based on the spectral decomposition (SVD) of the time-averaged adjacency matrix and stochastic perturbations.

Usage

hdsbm_fit(
  Y,
  k,
  nrep = 10,
  perturb_range = c(0.05, 0.35),
  maxit = 200,
  tol = 1e-10,
  verbose = TRUE,
  seed = NULL
)

Arguments

Y

Binary 3D array of dimensions \(n \times n \times T\) (the temporal adjacency matrix).

k

Integer, number of latent communities/blocks.

nrep

Integer, number of random restarts perturbed via SVD (default: 10). If 0, only the initial fit is performed.

perturb_range

Numeric vector of length 2 indicating the lower and upper limits of the random perturbation rate (default: c(0.05, 0.35)).

maxit

Integer, maximum number of CEM iterations per restart (default: 200).

tol

Numeric, convergence tolerance for the log-likelihood (default: 1e-10).

verbose

Logical; if TRUE, prints restart progress to the console (default: TRUE).

seed

Optional integer to ensure reproducibility.

Value

An object of class hdsbm containing:

  • best_fit: the estimate with the highest log-likelihood (parameters \(\pi_0, \Pi, \Psi\), classifications cl and clv).

  • lk_trace: vector containing the log-likelihoods obtained from each restart. #'

  • best_restart: the index of the restart that achieved the optimal value (0 = baseline).

  • icl: ICL score calculated on the optimal fit.

  • params: model dimensional parameters (n, k, TT).