R/generalize-userfunction.R
mlvsbm_estimate_generalized_network.RdThe inference use a greedy algorithm to navigate between model size. For a given model size, the inference is done via a variational EM algorithm. The returned model is the one with the highest ICL criterion among all visited models.
By default the algorithm fits a single level SBM for each level, before
inferring the generalized multilevel network. This step can be skipped by
specifying an initial clustering with the init_clustering. Also,
a given model size can be force by setting the parameters nb_clusters
to a given value.
mlvsbm_estimate_generalized_network(
gmlv,
nb_clusters = NULL,
init_clustering = NULL,
nb_cores = NULL,
init_method = "hierarchical",
fit_options = list(ve = "joint")
)A GenMLVSBM object, the network to be inferred.
A vector of L integer, the model sizes.
If left to NULL, the algorithm
will navigate freely. Otherwise it will navigate between the specified model
size and its neighbors.
A list of L vectors of integers of the same length as the number of node of each level. If specified, the algorithm will start from this clustering, then navigate freely.
An integer, the number of cores to use. Default to 1
for Windows and detectCores()/2 for Linux and MacOS
One of "hierarchical" (the default) or "spectral", "spectral" might be more efficient but can lead to some numeric errors. Not used when int_clustering is given.
A named list to be passed to the VE-M inference algorithm.
A FitGenMLVSBM object, the best inference of the network
## fit_options
### ve : Using the default ve = "joint" will update all the block
memberships of all levels at each VE step before performing a M step.
Using ve = "sequential" will update the block memberships of one level
at a time before performing a M step only on the concerned parameters. Use
this option if the running time of the algorithm is too long and the number
of levels is large.
my_genmlvsbm <- MLVSBM::mlvsbm_simulate_generalized_network(
n = c(20,20), # Number of nodes for the lower level and the upper level
Q = c(2,2), # Number of blocks for the lower level and the upper level
pi = list(c(.3, .7),NULL), # Block proportion for the upper level, must sum to one
gamma = list(matrix(c(.9, .2, # Block proportion for the lower level,
.1, .8), # each column must sum to one
nrow = 2, ncol = 2, byrow = TRUE)),
alpha = list(matrix(c(.8, .2,
.2, .1),
nrow = 2, ncol = 2, byrow = TRUE), # Connection matrix
matrix(c(.99, .3,
.3, .1),
nrow = 2, ncol = 2, byrow = TRUE)),# between blocks
directed = c(FALSE, FALSE), # Are the upper and lower level directed or not ?
affiliation = "preferential",
distribution = rep("bernoulli", 2)) # How the affiliation matrix is generated
fit <- MLVSBM::mlvsbm_estimate_generalized_network(
gmlv = my_genmlvsbm, nb_cores = 1)
#> 0/1:
#> 0/1:
#> 0/2:
#> 0/2:
#> [1] "Infering level 1:"
#> [1] "# blocks: 2, ICL = -95.8665176037591 !"
#> [1] "Infering level 2:"
#> [1] "# blocks: 2, ICL = -82.8923080740211 !"
#> [1] "======= # Blocks : 2, 2, ICL : -173.150312448671========"
#> [1] "ICL for independent levels : -178.75882567778"
#> [1] "ICL for interdependent levels : -173.150312448671"
#> [1] "=====Interdependence is detected between levels!====="
# A more complex example. A 4 levels network with different structures and
# level direction.
n <- 100
L <- 4
alpha <- list(
diag(.4, 3, 3) + .1, # Undirected assortative community
-diag(.2, 3, 3) + .3, # Undirected disassortative
matrix(c(.8, .2, .1, # Undirected core-periphery
.4, .4, .1,
.2, .1, .1), 3, 3),
matrix(c(.3, .5, .5, # Directed mixt structure
.1, .4, .5,
.1, .3, .1), 3, 3)
)
gamma <- lapply(seq(3),
function(m) matrix(c(.8, .1, .1,
.1, .8, .1,
.1, .1, .8), 3, 3, byrow = TRUE))
pi <- list(rep(1, 3)/3, NULL, c(.1, .3, .6), NULL)
directed = c(FALSE, FALSE, FALSE, TRUE)
gmlv <- mlvsbm_simulate_generalized_network(n = rep(n, 4),
Q = rep(3, 4),
pi = pi,
gamma = gamma,
alpha = alpha,
directed = directed,
distribution = rep("bernoulli", 4))
if (FALSE) {
fit <- mlvsbm_estimate_generalized_network(gmlv,
fit_options = list(ve = "joint"))
plot(fit)
fit2 <- mlvsbm_estimate_generalized_network(gmlv,
fit_options = list(ve = "sequential"))
fitone <- mlvsbm_estimate_generalized_network(gmlv2, nb_clusters = rep(3, 4),
fit_options = list(ve = "sequential"))
fit_from_scratch <- mlvsbm_estimate_generalized_network(
gmlv,
init_clustering = lapply(seq(4), function(x)rep(1, n)),
init_method = "merge_split",
fit_options = list(ve = "joint"))
}