Control of a missSBM fit

Helper to define the list of parameters that controls estimateMissSBM(). All arguments have defaults.

Usage

missSBM_param(
  threshold = 0.01,
  maxIter = 50,
  fixPointIter = 3,
  imputation = c("median", "average", "zero"),
  similarity = l1_similarity,
  useCov = TRUE,
  clusterInit = NULL,
  polish = TRUE,
  iterates = 1,
  maxMergeCandidates = 30,
  stopOnDegenerate = TRUE,
  maxConsecutiveDegenerate = 2,
  warmChain = FALSE,
  trace = TRUE
)

Arguments

threshold

V-EM algorithm stops when an optimization step changes the objective function or the parameters by less than threshold. Default is 1e-2.

maxIter

V-EM algorithm stops when the number of iteration exceeds maxIter. Default is 50.

fixPointIter

number of fix-point iterations in the V-E step. Default is 3.

imputation

character, the imputation strategy used to build the initial clustering (see partlyObservedNetwork's imputation()). Either "median" (default), "average" or "zero".

similarity

an R x R -> R function to compute similarities between node covariates. Default is l1_similarity, that is, -abs(x - y). Only relevant when the covariates are node-centered (i.e. covariates is a list of size-N vectors).

useCov

logical. If covariates is not empty, should they be used for the SBM inference (or just for the sampling)? Default is TRUE.

clusterInit

initial clustering: NULL (default) for a cold spectral clustering, or a list with length(vBlocks) vectors, each with size ncol(adjacencyMatrix), providing a user-defined clustering.

polish

logical, should each model be node-swap-polished (see missSBM_fit's polish()) after the initial VEM fit, fixing individually misclassified nodes at that model's own number of blocks? Cheap relative to exploration (iterates), since it does not search across numbers of blocks. Default is TRUE.

iterates

integer, the number of forward/backward exploration passes searching for a better number of blocks by splitting/merging clusters (unlike polish, which only refines each model at its own number of blocks): more expensive than polish. 0 disables exploration entirely (only the initial VEM fit, plus polish if enabled, is returned). Default is 1.

maxMergeCandidates

integer, caps the number of cluster-pair merge candidates tried during backward exploration (quadratic in the number of blocks otherwise). Beyond this cap, only the pairs with the most similar fitted connectivity profiles are tried, since merging two blocks with very different connectivity is rarely competitive anyway. Default is 30.

stopOnDegenerate

logical. A requested number of blocks can be higher than what the network actually supports: VEM then collapses one or more classes (see missSBM_fit's repair()), and even a fair recovery attempt (repair(), then a full exploration pass letting explore_forward() try to split a healthy neighbor into that range) can still collapse right back down. When this persists over maxConsecutiveDegenerate consecutive (increasing) values of vBlocks after a pass, forward (split) exploration stops growing further into that range on subsequent passes (only relevant when iterates > 1); the affected models stay as fitted, flagged via $degenerate, and a warning is issued. Default is TRUE.

maxConsecutiveDegenerate

integer, the run length that triggers stopOnDegenerate. Default is 2.

warmChain

logical (work in progress). If TRUE, each model is initialized by splitting (see missSBM_fit's split()) the already-converged, smaller neighbor in vBlocks instead of an independent cold spectral clustering (see missSBM_collection's estimate_chain()) – meant to reduce VEM component collapse at higher numbers of blocks. Sequential in the number of blocks, unlike the default (FALSE), which fits every model in parallel: can be slower in wall-clock time when several workers are available. Default is FALSE.

trace

logical for verbosity. Default is TRUE.

Value

a list of parameters configuring the fit, with class missSBM_param.

See also

estimateMissSBM()

Examples

my_control <- missSBM_param(iterates = 2, polish = FALSE)
my_control$iterates
#> [1] 2