Skip to contents

Estimate a colBiSBM on a collection of networks

estimate_colBiSBM.Rd

Estimate a colBiSBM on a collection of networks

Usage

estimate_colBiSBM(
  netlist,
  colsbm_model,
  net_id = NULL,
  distribution = "bernoulli",
  nb_run = 3L,
  global_opts = list(),
  fit_opts = list(),
  Z_init = NULL,
  sep_BiSBM = NULL
)

Arguments

netlist

A list of matrices.

colsbm_model

Which colBiSBM to use, one of "iid", "pi", "rho", "pirho".

net_id

A vector of string, the name of the networks.

distribution

A string, the emission distribution, either "bernoulli" (the default) or "poisson".

nb_run

An integer, the number of run the algorithm do. Default to 3.

global_opts

Global options for the outer algorithm and the output. See details.

fit_opts

Fit options for the VEM algorithm. See details

Z_init

An optional bi-dimensional list of size Q1_max x Q2_max containing for each value a list of two vectors of clusters memberships. Default to NULL.

sep_BiSBM

A pre-fitted sep_BiSBM. Used to avoid end computations. The best way to obtain one is to extract from a fitted bisbmpop object. Defaults to NULL.

Value

A bisbmpop object listing a collection of models for the collection. of networks

Details

The list of parameters global_opts essentially tunes the exploration process.

  • nb_cores integer for number of cores used for parallelization. Default is 1

  • verbosity integer for verbosity (0, 1, 2, 3, 4). Default is 1. 0 will disable completely the output of the function. Note: you can access the $joint_modelisation_preferred attribute to check which modelisation is preferred

  • Q1_max integer for the max size in row to explore. Default is computed with the following formula: floor(log(sum(sapply(netlist, function(A) nrow(A)))) + 2)

  • Q2_max integer for the max size in columns to explore. Default is computed with the following formula: floor(log(sum(sapply(netlist, function(A) ncol(A)))) + 2)

  • nb_models the number of models to keep for each values of Q1,Q2. Default is 5.

  • depth specifies how large will the moving window be. Default is 1, meaning the window will go from (Q1 - 1, Q2 - 1) to (Q1 + 1, Q2 + 1) and all the values in the square defined.

  • plot_details integer to control the display of the exploration and moving window process. Values are 0 or 1. Default is 1.

  • max_pass the maximum number of moving window passes that will be executed. Default is 10.

The list of parameters fit_opts are used to tune the Variational Expectation Maximization algorithm.

  • algo_ve a string to choose the algorithm to use for the variational estimation. Available: "fp"

  • verbosity an integer to choose the level of verbosity of the fit procedure. Defaults to 0. Available: 0,1

  • max_vem_steps an integer setting the number of Variational Expectation-Maximization steps to perform. Defaults to 1000.

  • minibatch a boolean setting wether to use a "minibatch" like approach. If set to TRUE during the VEM the networks will be optimized in random orders. If set to FALSE they are optimized in the lexicographical order. Default to TRUE.

  • tolerance a numeric, controlling the tolerance for which a criterion is considered converged. Default to 1e-6.

  • greedy_exploration_max_steps the maximum number of iteration of greedy exploration to perform. Defaults to 50.

  • greedy_exploration_max_steps_without_improvement an integer indicating for which number of steps the best model must not change to be end greedy exploration. Defaults to 5.

  • kmeans_nstart an integer indicating the number of random starts to use for kmeans in spectral clustering.

  • kmeans_iter_max an integer indicating the maximum number of iterations to use for kmeans in spectral clustering.

See also

clusterize_bipartite_networks(), bisbmpop, fitBipartiteSBMPop, browseVignettes("colSBM")

Examples

alpha1 <- matrix(c(0.8, 0.1, 0.2, 0.7), byrow = TRUE, nrow = 2)
alpha2 <- matrix(c(0.8, 0.5, 0.5, 0.2), byrow = TRUE, nrow = 2)
first_collection <- generate_bipartite_collection(
  nr = 50, nc = 25,
  pi = c(0.5, 0.5), rho = c(0.5, 0.5), alpha = alpha1, M = 2
)
second_collection <- generate_bipartite_collection(
  nr = 50, nc = 25,
  pi = c(0.5, 0.5), rho = c(0.5, 0.5), alpha = alpha2, M = 2
)

netlist <- append(first_collection, second_collection)

if (FALSE) { # \dontrun{
# A collection where joint modelisation makes sense
cl_joint <- estimate_colBiSBM(
  netlist = first_collection,
  colsbm_model = "iid",
  global_opts = list(nb_cores = parallelly::availableCores(omit = 1L))
)
# A collection where joint modelisation doesn't make sense
cl_separated <- estimate_colBiSBM(
  netlist = netlist,
  colsbm_model = "iid",
  global_opts = list(nb_cores = parallelly::availableCores(omit = 1L))
)
} # }